Greetings, and GO KNICKS.

Since I last wrote, which if memory serves was the day before, my Knicks ended a 53 year drought and brought basketball glory back to the city that never sleeps. Their first title since 1973. Jalen Brunson closed it out with 45 in Game 5 on the road and walked off as Finals MVP, which felt especially sweet given who spent two years doubting him. Becky Hammon, the Aces coach and ESPN analyst, went on air in 2023 to declare Brunson too small to be a “1A dude” and incapable of carrying this team anywhere. After he carried it all the way, the city waited for an apology. What it got was Hammon allowing that her “opinion was wrong” while flatly refusing to actually say sorry, until she admitted “my opinion was wrong” at a press conference, which was about as lukewarm as the crowd that walked out on Trump at the State Fair this week.

I happened to be in the city for a podcast on the day of the parade, and I saw a WAVE of Knicks fans filling the streets. It made me proud of my city. (I also do not care what Jennifer Lopez thinks makes someone a New Yorker. I am one! White Plains, baby!)

In other news, my hot takes on the Fox and Roku deal. Fox needed to get younger, and Roku, well, Roku had no takers. It is hard to keep innovating as a smart app provider once the platform matures. Fox gains access to over 100 million streaming households and becomes the dominant connected TV gateway in the US, though on share of viewing the combined company lands third, behind YouTube and Disney. They are committed to keeping it a “partner friendly platform,” but we will see what the Murdoch triumvirate ultimately does. At the end of the day Fox is still buying a distribution platform, and they are now a bit more entrenched. The advertising piece is the strong part. Every app on Roku already hands over roughly 30 percent of its ad inventory, which Roku sells and keeps for itself, so once the deal closes the company reselling those slots and harvesting that data is no longer a neutral party. It is a direct content rival. Fox gets to skim a share of inventory from its own competitors, which is, at the very least, awkward.

This week I wanted to examine Scrubs, the medical comedy that ran on NBC and then ABC from 2001 to 2010 and has now come back to ABC for a tenth season. I started in on the revival episodes recently, and the experience of watching it land in 2026 turned out to say more about the country than about the show.

At the onset, Scrubs was saddled with the wrong comparisons. It shared a setting with the prestige hospital dramas of its moment, ER and the early run of Grey’s Anatomy, and it lived in the long shadow of the workplace dramedy that invented the form, MAS*H. Like a lot of shows that arrive into a crowded genre, it inherited a legacy as much as it tried to forge one. But the comparison missed what it was actually doing. Where ER treated the hospital as a theater of competence and Grey’s treated it as a theater of romance, Scrubs treated it as a theater of labor: young, indebted, exhausted doctors discovering that the institution they served was indifferent to whether they survived it. Its formal signature, the fantasy cutaways and the slapstick colliding with sudden death inside a single edit, was never just a comic trick. It was a thesis about how people stay intact while standing that close to mortality and bureaucracy at the same time.

The medical comedy often tasks itself with the institution it lives inside, and Scrubs, for all its absurdity, never pretended the institution was benign. Its recurring antagonist was rarely a person. It was the cost structure: the denied claim, the bed that had to be cleared, the patient who could not be treated because the arithmetic did not allow it. Bob Kelso, the smiling administrator who reduced every life to a line item, was the most honest character on television about how American medicine actually runs.

None of this exists in a vacuum. Scrubs debuted in October 2001, into a country that had just been reorganized by September 11 and was settling into the managed care era, when the distance between what medicine could do and what a patient could afford was widening into a permanent feature of the system rather than a temporary failure of it. J.D. and Turk were not asking whether they could become doctors. They were asking what kind of doctors a system loaded with debt and rationed by cost would permit them to become. That was the quiet engine under the goofiness, and it is why the show aged into relevance instead of out of it.

By the time the revival arrives, that question has grown considerably more complicated, and the country has too. The season picks the original timeline back up in the present day. In the years off screen, J.D. became a concierge doctor to the wealthy before returning to Sacred Heart, and that is not an incidental piece of biography. It is the entire economics of contemporary American medicine compressed into one character’s resume: the steady migration of talent toward the patients who can pay out of pocket, and the two tiers that migration leaves behind. He left to serve the people who could afford him. The show brings him back to the people who cannot, and it does not have to underline why that is funny and sad at once.

The other revived tension is labor. Dr. Cox, still Chief of Medicine, is shown trying and failing to reconcile his abrasive teaching style with a new generation of interns who arrive with protections he never had. It plays for laughs, but it is the same structural story told from a different angle. The workplace has changed, the tolerance for cruelty as pedagogy has changed, and an institution built on overwork is being asked, gently and then less gently, to account for what it extracts from the people inside it. The original show diagnosed a profession that ground its young down for the love of the work. The revival watches that profession get asked, for the first time, to pay for the grinding.

The reason the revival works, and it has, drawing strong reviews and an early renewal for an eleventh season, is that its premise needed no updating. A comedy about idealistic doctors inside an indifferent, cost governed institution did not have to revise its thesis for 2026. The country revised it for them. Hospitals consolidated, ownership moved toward balance sheets that had never set foot on a ward, burnout stopped being a personality flaw and became a measured crisis, and concierge medicine quietly normalized the idea that the quality of your care is a function of the size of your account. What Black-ish asked of the Black middle class, what does making it cost you, Scrubs has always asked of medicine: what does staying good cost you, inside a system optimized for everything except that.

That sociological backdrop helps explain both why the show worked in 2001 and why it works again now. The Sacred Heart that J.D. walks back into is the same hospital, lit the same way, scored with the same Lazlo Bane theme. The bill attached to it is just much larger, and far fewer people in the audience need it explained.

What I Measured - and How Well It Worked

The study analyzed nearly every episode of Scrubs by breaking each script into its three acts and measuring dialogue features at each stage, including things like analytical language density, emotional register, and how much the characters said. The goal was to see which of those features, in which act, predicted how many people watched the following week’s episode.

Think of it as asking: if you could read the script before the episode aired, what would tell you whether viewers were going to come back?

36.79% Variance explained (Adjusted R²) — in-sample model fit

29.11% Out-of-sample fit tested on unseen episodes

CV folds 0.31/ 0.65 / 0.25 / 0.39 / -0.15 Mean: 0.2911

180 of 182 Total episodes analyzed — near-complete series corpus

When we tested the model on episodes it had never seen before, it explained about 36.79% of the variation in viewership from one week to the next, from dialogue features alone, with no information about scheduling, marketing, or what was on the other channel. That’s a meaningful signal from something as seemingly mechanical as word choice. While the dependent variable is next-episode viewership in millions, I don’t believe the coefficients that you are about to see impact millions of viewers directly, as there are far too many other forces not captured in the model. But directionally, I believe that my database helps decode why particular TV series work based on character dialogue.

Six features turned out to be statistically significant predictors. Here’s what they were, what they mean in plain English, and which episodes make the pattern obvious.

What Helped: The Features That Brought Viewers Back

1. Power language in the closing act (Power_3)

The strongest positive predictor in the model was the presence of power language in Act 3: words associated with dominance, command, and hierarchical structure. The coefficient was +4.11.

The reading is not that audiences wanted to watch people get bossed around. It is that Scrubs is, structurally, a show about a hospital hierarchy, and its resolutions land hardest when they are delivered with the weight of that hierarchy behind them. Sacred Heart runs on rank: attending over resident, resident over intern, Kelso over everyone, Cox over J.D. in the way that actually matters. When the third act resolves through that structure rather than around it, the lesson arrives with authority instead of sentiment. Dr. Cox does not comfort J.D. He pronounces. The closing truth carries because someone with standing is delivering it, and the show has spent the runtime establishing why that standing was earned.

Power language in Act 3 is the sound of a resolution that means it.

Episode in focus · Season 8, Episode 17 · “My Chief Concern” · Alignment score: 0.7199

J.D. is furious when Turk, newly installed as chief surgeon, overrules him on a medical decision. The question the episode circles is whether Turk pulled rank for the right reason or because J.D. is leaving. The entire third act is a negotiation of authority between two people whose friendship has always been horizontal and is suddenly, painfully, vertical. The Power_3 coefficient of +4.11 captures what the episode stages directly: the resolution lands because it runs through the new hierarchy rather than pretending the hierarchy is not there.

2. Tentative language in the closing act (Tentat_3)

The second positive predictor sits right next to the first and pulls in what looks like the opposite direction: words that indicate uncertainty or possibility, the maybe, the perhaps, the unfinished thought. The coefficient was +3.36.

This is the part the show understood better than almost any of its peers. Scrubs almost never closed clean. The signature move was a final voiceover that gestured at a lesson and then declined to fully land it, leaving J.D., and the audience, holding a question rather than an answer. The tentative register in Act 3 is the texture of that refusal. It is what keeps the resolution from curdling into a moral. The audience comes back to a show that trusts them to sit with doubt, not one that tucks them in.

Read the two positives together and you get the show’s actual fingerprint. Act 3 wants authority and uncertainty in the same breath: Cox delivers the hard truth with the full weight of rank, and then J.D.’s narration quietly admits he is not sure it is the whole truth. Power without doubt is a lecture. Doubt without power is mush. Scrubs at its best held both.

Episode in focus · Season 6, Episode 14 · “My No Good Reason” · Alignment score: 0.6878

Dr. Cox and Nurse Roberts spend the episode arguing about whether things happen for a reason. The episode never settles it, and that is the point. The closing act does not resolve the debate so much as let it breathe, with the most authoritative figures in the building admitting they cannot close it. The Tentat_3 coefficient of +3.36 captures exactly this: an ending built on possibility rather than proof, which is the register audiences returned for.

What Hurt: The Features That Drove Viewers Away

3. Visual perception language in the middle act (See_2)

The first negative predictor: episodes where Act 2 scored higher on visual perception language, words tied to looking, seeing, watching, describing what is in frame. The coefficient was −3.78.

The structural reading matters here. Scrubs had a deep bench of visual machinery, the fantasy sequences and sight gags being the most obvious. When Act 2 leaned on that machinery, on describing and staging spectacle rather than escalating a relationship, the development act stopped developing. Visual world building is cheap attention; it fills the screen without advancing the conflict. The middle act is where the emotional argument is supposed to get harder, and an Act 2 that is busy showing you things is an Act 2 that has stopped making things harder.

If your second act is describing what the audience can already see, ask what the characters are no longer doing to each other.

4. Verbosity in the middle act (WC_2)

Episodes where Act 2 carried a higher raw word count predicted fewer viewers the following week. The coefficient was −3.81.

This is the same problem the CSI and Black-ish analyses surfaced from different angles, and it is becoming a pattern worth naming. A long middle act is usually an act that is explaining rather than escalating. In Scrubs terms, Act 2 is where the medical case and the emotional case are supposed to collide and complicate. When the word count climbs, the act is typically narrating that collision instead of letting it happen. More words in the development act is not more development. It is more throat clearing.

5. Analytical language in the middle act (Analytic_2)

Episodes where Act 2 scored higher on analytical, formal, or logical language predicted fewer viewers. The coefficient was −4.05.

The trap is specific to medical television. Scrubs had a standing license to be smart about medicine, and the temptation was to spend the middle act being smart at the audience: the differential diagnosis delivered as a logic puzzle, the procedure explained as a flowchart, the problem solved on a whiteboard. Analytical language in Act 2 is the sound of the show treating a case as a problem to be reasoned through rather than a situation people are inside. The audience did not tune in for the chart. They tuned in for what the chart did to Cox, or J.D., or Carla. When the middle act over intellectualizes, it trades the second thing for the first.

6. Verbosity in the closing act (WC_3)

The single heaviest penalty in the entire model: episodes where Act 3 carried a higher word count predicted significantly fewer viewers the following week. The coefficient was −6.94, nearly double the weight of anything else in the set.

This is the finding that ties the whole model together, and it should be read against the two positives. The closing act wants power and it wants doubt, but the data is emphatic that it does not want words. A long Act 3 is a resolution that explains itself. It is the voiceover that does not trust the silence, the lesson restated three times in case you missed it, the ending that keeps talking after it has already landed. Everything the show’s best closings did through compression, the authority of a single line from Cox, the open question of a five second narration, a wordy Act 3 undoes by over delivering. The strongest thing the data says about Scrubs is this: when the ending ran long, the audience left.

The recommendation text in the source flags this penalty against Act 2 by copy error. It belongs to Act 3, and given that it is the largest coefficient in the model, the placement is worth correcting. The resolution is where verbosity costs the most.

One More Wrinkle: What the Medical Comedy Genre Says

The findings above read Scrubs from the inside. Two observations from the broader genre data are worth holding alongside them.

The verbosity penalties travel. The word count findings here, WC_2 at −3.81 and WC_3 at −6.94, are not idiosyncratic to Scrubs. The same penalty showed up in the Black-ish sentence length and dictionary density results and in the CSI word count work before that. Across single camera comedies and procedurals alike, acts that run long tend to be acts that explain, and explanation is consistently punished. This is not a scene level note you fix by trimming one monologue. It is a room level understanding of what each act is for.

The Act 3 combination may be the show’s own. The pairing of power and tentativeness in the closing act, authority and doubt delivered together, looks more genre and structure conditional than the verbosity findings. It is the specific product of a show built on a mentor hierarchy and an anxious narrator: Cox supplies the power, J.D.’s voiceover supplies the doubt, and the architecture needs both poles to be present. A show without that vertical mentor structure, or without a narrating interior, would not necessarily reward the same combination. Replication across other medical and workplace comedies is the way to find out, and that work is ongoing.

So What Would You Actually Change? The Writer’s Takeaway

The model is not a style guide. It is a set of structural pressures, and they point somewhere usefully specific.

Positive relationships, consider emphasizing

+4.11 · Power language in Act 3. Let your resolution run through the hierarchy, not around it. The lesson lands harder when it comes from someone with standing to deliver it.

If your ending feels like comfort, ask who in the building has the authority to make it cost something.

+3.36 · Tentative language in Act 3. Resist the clean close. The endings audiences returned for left a question open rather than stamping a moral shut.

If your final beat resolves everything, you have probably taken away the reason to come back and find out what happens next.

Negative relationships, use with caution

−3.78 · Visual perception in Act 2. Spectacle in the middle act is attention without escalation. Use the fantasy machinery to deepen conflict, not to fill the runtime.

If Act 2 is describing what we can see, the characters have stopped doing things to each other.

−3.81 · Word count in Act 2. A long development act is usually a narrated one. Let the collision happen instead of explaining that it is happening.

Cut the connective tissue. The middle act should escalate, not summarize.

−4.05 · Analytical language in Act 2. The case is not the point. The differential delivered as a logic puzzle is the show being smart at the audience instead of putting people inside a situation.

If the whiteboard is doing the work, ask what the whiteboard is doing to Cox.

−6.94 · Word count in Act 3. This is the one to internalize. The closing act wants weight and ambiguity, not volume. A wordy resolution explains itself out of its own power.

If the ending is still talking, it has already landed. Trust the silence and the single line.

Technical Notes

Each feature was extracted with LIWC, applied at the act level, with the suffix _1, _2, or _3 denoting the act in which it was measured. Coefficients were estimated with HC3 robust standard errors. Given the sample size, the responsible reading is directional: positive versus negative association, with magnitude as relative weight rather than a unit effect on viewership in millions.

Appendix A · Glossary of All Six Measures

Significant positive predictors, more of this, more viewers next week

Power_3 · Power language, Act 3. LIWC Power category in the closing act: words associated with dominance, command, and hierarchical structure. Captures a resolution delivered through rank and standing rather than sentiment. Sig. positive · coef +4.11.

Tentat_3 · Tentative language, Act 3. LIWC Tentative category in the closing act: words indicating uncertainty or possibility, such as maybe and perhaps. Captures an ending that leaves a question open rather than stamping a conclusion shut. Sig. positive · coef +3.36.

Significant negative predictors, more of this, fewer viewers next week

See_2 · Visual perception language, Act 2. LIWC See category in the development act: words tied to looking, watching, and describing the visual field. High scores indicate a middle act leaning on spectacle and world building rather than escalating conflict. Sig. negative · coef −3.78.

WC_2 · Word count, Act 2. Total words in the development act. High scores indicate a middle act that narrates its collision rather than letting it happen. Sig. negative · coef −3.81.

Analytic_2 · Analytical language, Act 2. LIWC Analytic measure in the development act: formal, logical, problem solving register. High scores indicate an act treating its case as a puzzle to be reasoned through rather than a situation people are inside. Sig. negative · coef −4.05.

WC_3 · Word count, Act 3. Total words in the closing act, and the heaviest penalty in the model. High scores indicate a resolution that over explains itself, undoing through volume the authority and ambiguity its best endings achieved through compression. Sig. negative · coef −6.94.

Appendix B · Most Aligned Episodes

What the data’s best matching cases actually show

The alignment score measures one specific thing, and it is worth being precise about what. It is how closely an individual episode’s writing matched the full feature pattern the show level Random Forest identified as predictive of retention. A high alignment score does not mean an episode performed well. It means the writing was operating inside the show’s structural fingerprint, exhibiting the features the model found important in proportions consistent with what it learned across the full nine season run of 182 episodes.

With that definition in place, the most aligned episodes in the dataset are these.

The first thing the table shows is where the most aligned episodes came from. Three of the top five sit in Seasons 8 and 9, the ABC years, after the move from NBC and into the soft reboot that Scrubs: Med School became. The writing was more precisely calibrated to the retention pattern during the show’s contraction than during much of its peak, with one conspicuous exception near the top.

The second thing the table shows is what that precision was worth. “My Rabbit,” the most aligned episode in the series by this measure, drew 5.28 million. “My Lucky Charm,” ranked second in alignment and aired in Season 4 at the show’s commercial peak, drew 7.02 million, the highest figure in the table. The Season 8 and 9 episodes clustered behind them drew between 3.06 and 4.56 million. Alignment is not performance. It is a measure of internal consistency with the writing pattern, and the gap between “My Lucky Charm” and the late season episodes here illustrates that gap precisely.

The third thing the table invites is the harder question: why do late season episodes align so highly in the first place. The intuitive answer is that the room had refined the formula over time. The data suggests two less flattering explanations.

The first is a training artifact. The show level Random Forest was fit on episodes across the entire run, which means Seasons 8 and 9 are a meaningful share of the training data. The learned retention pattern is therefore not purely a portrait of peak Scrubs. It is partly a portrait of the contracting and rebooted show, because that is part of what it was trained on. When a Season 9 episode scores high on alignment, some of what is being measured is that the model recognizes it as structurally typical of the era it partly learned from. That is closeness to the training distribution, not convergence on the show’s best self.

The second explanation is variance. Peak Scrubs was doing more things. The room took bigger swings, the fantasy structure ran wilder, and the show had the commercial security to experiment. That variability spread peak episodes further from the model’s central tendency, producing lower alignment scores not because the writing was weaker but because it was more dispersed. By the ABC years, and especially in the Med School reboot, the creative range had narrowed. Less experimentation means more episodes clustering near the same feature values, which the model reads as high alignment regardless of whether those values point anywhere productive. Consistency is not the same as correctness.

Generalizability. Findings are specific to Scrubs and should not be applied directly to other series without replication. Cross genre and cross network analysis is in progress. Preliminary results suggest the word count penalties generalize across single camera comedies and procedurals; the Act 3 combination of authority and doubt appears more conditional on a show’s mentor hierarchy and narrating structure.

Last time out I put the Stranger Things pilot through my script analysis tool and walked through what its structure quietly revealed beneath the surface. A few of you wrote back asking what other pilots would look like under the same lens, and one title kept surfacing. So here we are. Today I want to talk about 24.

A quick reminder for anyone new here. For the better part of a year I have been building a tool that reads movie and TV scripts and breaks them down in an organized, repeatable way. It runs on large language models, but I have programmed it strictly to read and analyze rather than to learn or generate. The goal is simple. Give writers, producers, and business development people a shared analytical foundation, a common language they can actually agree on when they sit down to argue about what a script is doing and why. I am aiming to make it commercially available in August. Until then I am sharing what I find here, in public, while it is still being built.

And I will be honest about why I reached for 24.

When I wrote about Stranger Things, I spent a fair amount of time on the retention problem facing streaming services. The core issue is cadence, or rather the loss of it. When everything is available all at once and the slate never stops expanding, audiences have no structural reason to commit to any one show over time. They chase whatever is hot this week and then move on to the next thing.

24 is fascinating to me precisely because it comes from the opposite world and solved the opposite problem in the most literal way imaginable. It premiered on Fox on November 6, 2001, deep in the age of appointment television, when a network needed you to show up at a specific hour on a specific night or you simply missed the episode. And 24 took that constraint and made it the entire premise of the show. Each season covers a single day. Each episode covers a single hour of that day, told in something very close to real time, with a digital clock ticking on screen and a split screen pushing several storylines forward at once. The form itself was a retention mechanism. You could not skip ahead without breaking the pact, because the pact was the clock.

I find that genuinely remarkable as a piece of media engineering. Most shows try to earn your return through character and mystery, the softer debts I wrote about last time. 24 did that too, and did it well. But it also wired your return directly into the structure of the thing. The clock is a promise and a threat at the same time. Tune in next week, same hour, or the hour happens without you.

There is also the matter of timing, and I do not mean the one on screen. 24 arrived less than two months after the September 11th attacks, and a show about a counterterrorism agent racing the clock to stop a plot landed in a country that was suddenly, painfully primed to receive it. I am going to set aside the long and legitimately fraught conversation about how the series handled torture and politics across its full run, because that deserves its own piece and is not really what the pilot is doing. But it would be dishonest to analyze this episode as a pure structural object and pretend the cultural moment had nothing to do with why it worked.

So that is the setup. Below I want to walk through what the first hour of 24 actually does, what the tool surfaced when I ran the script through it, and why I think this pilot holds up as one of the more confident opening hours network television has produced. For what it is worth, it won the Emmy for Outstanding Writing for a Drama Series, so the people who hand out trophies agreed.

As always, I am eager for feedback. I am still figuring out how to make these breakdowns clearer and more useful for the people who actually make this stuff. Find me on LinkedIn at https://www.linkedin.com/in/aphd/. Happy to chat.

Moving on.

What the First Hour Actually Did

The pilot was working with a structural problem that is, in its own way, as constrained as the Stranger Things premise. You have sixty minutes of story that has to occupy sixty minutes of clock. You cannot cut to three weeks later. You cannot montage your way past the boring parts. Every minute you spend is a minute the audience spends with you, and the meter is running the whole time. That sounds like a cage. The Duffer Brothers earned their tension through restraint and silence. Robert Cochran and Joel Surnow earned theirs through a different bet entirely, and a few things became clear once I ran it through the tool.

First, the pilot establishes simultaneity before it establishes plot. The episode opens in Kuala Lumpur, in the shadow of the Petronas Towers, with a spy named Victor Rovner sending a message halfway around the world. Within minutes we are in Los Angeles with Jack Bauer and his family, then inside the Counter Terrorist Unit, then on the move with a presidential campaign. The show is not easing you into a single storyline. It is teaching you, immediately and without apology, that you are going to be holding several places at once and that they are all happening right now, in parallel. The split screen is not a stylistic flourish. It is the grammar of the entire show, introduced in the first hour so the audience learns the rules early.

Second, the pilot fuses the personal and the professional and refuses to let you separate them. Jack is called in because there is a credible threat on the life of Senator David Palmer, a leading presidential candidate, on the day of the California primary. At the very same moment his teenage daughter Kim has climbed out her bedroom window, and his wife Teri is trying to find her, in a marriage that the episode tells us is already strained and recently fractured. A lesser pilot would treat the domestic thread as relief from the political one. 24 does the opposite. It builds toward the realization that the two worlds are the same world, that the people coming for Palmer understand exactly who Jack loves and exactly how to use them. The stakes at CTU and the stakes in Jack’s own home are not running on separate tracks. They are the same track, and the pilot lays the rail for that collision in its opening hour.

Third, and this is where I think the pilot earns its reputation, it makes a deliberate bet on the audience’s tolerance for not knowing. There may be a mole inside CTU. Palmer receives a phone call that visibly unsettles him, and we are given almost nothing about who is on the other end. The full shape of the plot against him stays offstage. Like the best pilots, 24 treats those gaps as promises rather than as problems to be patched over with exposition. It trusts that a viewer who is denied an answer at midnight will come back at the same hour next week to get one.

The study analyzed nearly every episode of 24 by breaking each script into its three acts and measuring dialogue features at each stage, including things like analytical language density, emotional register, and how much the characters said. The goal was to see which of those features, in which act, predicted how many people watched the following week’s episode.

Think of it as asking: if you could read the script before the episode aired, what would tell you whether viewers were going to come back?

40.45% Variance explained (Adjusted R²) — in-sample model fit

32.78% Out-of-sample fit tested on unseen episodes

CV folds 0.33/ 0.39 / 0.29 / 0.17 / 0.47

192 of 195 episodes analyzed — near-complete series corpus (98.5%)

When we tested the model on episodes it had never seen before, it explained about 40.45% of the variation in viewership from one week to the next, from dialogue features alone, with no information about scheduling, marketing, or what was on the other channel. That’s a meaningful signal from something as seemingly mechanical as word choice. While the dependent variable is next-episode viewership in millions, I don’t believe the coefficients that you are about to see impact millions of viewers directly, as there are far too many other forces not captured in the model. But directionally, I believe that my database helps decode why particular TV series work based on character dialogue.

Seven features turned out to be statistically significant predictors. Here’s what they were, what they mean in plain English, and which episodes make the pattern obvious.

What Helped: The Features That Brought Viewers Back

1. Sophisticated vocabulary in the opening act (Sixltr_1)

The strongest positive predictor in the entire model was the share of long words, words over six letters, in Act 1. The coefficient: +2.20.

This one runs directly against the conventional wisdom about cold opens, which says keep it simple, keep it visceral, do not make the audience work before you have earned their attention. 24 says the opposite, and the reason is structural rather than stylistic. Long words in the opening act of this show are not ornamental. They are the sound of expertise. Protocols, jurisdictions, intelligence assessments, the names of foreign capitals and government agencies and the people who run them. When the first act of a 24 hour is lexically dense, it is telling the audience that the world they are entering is real, intricate, and consequential. The density is the stakes.

A thriller that opens in plain, common language is a thriller that sounds like it could happen anywhere, to anyone, over nothing in particular. A thriller that opens in the specialized vocabulary of counterterrorism sounds like it could only happen here, and only now, and only to people who know what they are doing. The audience leans in not because the words are hard but because the difficulty signals that something important is underway.

Episode in focus · Season 8, Episode 10 · “Day 8: 1:00 a.m.-2:00 a.m.” · Alignment score: 0.6216

This is the most aligned episode in the data I am looking at, and its opening register is a clean demonstration of why. President Taylor is leaning on President Hassan, Kayla has gone missing, a New Jersey swamp gives up two bodies, and Jack is constructing a ruse to draw out the Kamistani rebels. Read that sentence again and notice how much of it is proper nouns and procedural machinery. The hour opens inside a dense lattice of diplomatic names, fictional geography, and operational planning, and the lexical weight of all that is exactly what the Sixltr_1 coefficient of +2.20 is rewarding. The episode sounds expert from its first minutes, and the model says that is part of why audiences stayed.

2. Anger in the opening act (Anger_1)

The second positive predictor is anger in Act 1. The coefficient: +1.59.

If sophisticated vocabulary is what makes the opening act feel real, anger is what makes it feel urgent. 24 is a show whose entire premise is a day going wrong, and the first act is where the wrongness announces itself. Hostility, confrontation, the threat made explicit, the alliance fracturing. Anger in the opening act is the ignition. It is the moment the clock starts to matter.

The instinct to read this as a warning, to file anger under things that alienate an audience, gets the genre backward. In a sitcom, opening in anger might cost you the warmth the form depends on. In a real time thriller, anger is the genre contract. The audience tuned in precisely because they expect the day to detonate, and a first act that delivers that conflict is a first act that confirms the bargain. The danger is not too much anger early. The danger, as we are about to see, is the wrong emotion late.

Episode in focus · Season 7, Episode 8 · “Day 7: 3:00 p.m.-4:00 p.m.” · Alignment score: 0.5526

I will not pretend to a tidy logline for this specific hour, but it sits in the heart of the Season 7 arc, the stretch where an exiled Jack is pulled back in to chase Iké Dubaku and his conspiracy while the betrayal of a former ally tightens in the background. That is a season built on antagonism and broken trust, and this hour carries the highest next episode viewership in the entire dataset at 11.22 million. An episode that opens in that register of confrontation, and then hands the audience an 11 million viewer return, is the Anger_1 coefficient made visible.

What Hurt: The Features That Drove Viewers Away

3. Generic vocabulary in the opening act (Dic_1)

The mirror image of finding number one. Episodes whose Act 1 dialogue scored higher on dictionary density, a larger share of words drawn from common vocabulary, predicted fewer viewers the following week. The coefficient: −1.59.

Sixltr_1 and Dic_1 are two readings of the same dial. When the opening act reaches for sophisticated, specific language, retention rises. When it settles into ordinary, common words, retention falls. Generic vocabulary is the register of generic stakes. If the first act of a 24 hour could have been spoken on any procedural on television, then it is not yet the opening of this show. It is the opening of a category. The audience came for CTU, and an opening act that sounds like any control room is an opening act that has not yet earned the clock.

4. Disgust in the opening act (Disgust_1)

Episodes whose Act 1 dialogue scored higher on disgust, the language of revulsion, moral violation, and the transgression of norms, predicted fewer viewers the following week. The coefficient: −1.72.

24 promises competence under pressure, the satisfaction of watching capable people impose order on chaos against a ticking clock. Revulsion does not serve that promise. It stalls it. An opening act soaked in moral nausea asks the audience to recoil at the exact moment the show needs them to lean forward. The same lexical signal, two genres, opposite signs. That is not noise. That is the genre contract showing up in the data.

5. Bonding language in the closing act (Affiliation_3)

Episodes whose Act 3 dialogue scored higher on affiliation, the language of social bonds, alliance, and friendship, predicted fewer viewers the following week. The coefficient: −1.34.

Here the model starts describing the back of the episode, and a clear rule emerges: the closing act of 24 is for propulsion, not for reunion. Affiliation language in Act 3 tends to mark an hour that is winding down emotionally, settling its characters into connection and resolution. That is the wrong motion for this show. A 24 hour does not end. It hands off. The closing act is a launchpad for the next one, and the bargain that keeps audiences returning is the promise that the danger is not over, that someone is compromised, that the clock has merely reset toward the next crisis. When the final act turns toward the warmth of bonds rather than the tension of what comes next, it spends the very momentum the cliffhanger is supposed to carry forward.

6. Fear in the closing act (Fear_3)

Episodes whose Act 3 dialogue scored higher on fear, the language of dread, threat, and apprehension, predicted fewer viewers the following week. The coefficient: −2.01.

This is the counterintuitive heart of the whole analysis. 24 is a show that traffics in dread. It is, on its surface, an hour by hour machine for generating fear. And yet fear in the closing act is one of the largest penalties in the model. How?

The answer is in the difference between fear and propulsion. A closing act that tilts into fear is a closing act that leaves the audience in apprehension, in the felt sense of threat without agency, in dread that has nowhere to go. That is not the engine of this show. The engine is Jack acting. The cliffhanger that brings an audience back is not “something terrible is going to happen and we are helpless.” It is “something terrible is happening and watch what he does about it.” Fear that ends in paralysis breaks the contract. The clock demands forward motion, and a final act that ends in pure dread is a final act that has stopped the clock at the worst possible moment.

7. Disgust in the closing act (Disgust_3)

And the largest penalty in the entire model: disgust in Act 3, the language of revulsion and moral violation arriving at the climax. The coefficient: −2.58.

Everything that made Disgust_1 corrosive is amplified here, because the closing act is where the audience decides whether to come back. Disgust in the opening act stalls the engine. Disgust in the closing act poisons the handoff. An hour that leaves its audience revolted rather than propelled is an hour that has converted its final, most important real estate into a reason to look away. For a show whose closing act is supposed to be a coiled spring pointing at next week, ending in moral nausea is the surest single way to lose the audience the clock was built to keep.

Episode Deep-Dives

Episode in focus · Season 7, Episode 23 · “Day 7: 6:00 a.m.-7:00 a.m.” · Alignment score: 0.5852

Jack is abducted for unknown purposes, Kim sits in danger at the airport, and Olivia matches wits with Ethan Kanin. This is the penultimate hour of the season, and you can feel the apprehension loading into its close. It is a deeply aligned episode, 0.5852, and it still drew a strong 9.64 million the following week, so I am not arguing it failed. I am arguing it paid a cost the model can see. The Fear_3 penalty of −2.01 is the tax on an hour whose closing register leans toward peril and helplessness rather than the propulsive, what does Jack do next motion that defines the show at its most retentive.

Episode in focus · Season 8, Episode 15 · “Day 8: 6:00 a.m.-7:00 a.m.” · Alignment score: 0.5825

Season 8 is where the show darkens, where Jack moves toward the increasingly grim and personal war that defines its final stretch. That darkening register is exactly the territory where the Disgust_3 penalty bites. This hour is highly aligned at 0.5825, structurally typical of late 24, and it carries a real cost in the model precisely because of how far the season’s tone has drifted toward the bleak. The next episode viewership of 7.9 million is the lowest among the most aligned hours in the data, which is the kind of quiet corroboration the −2.58 coefficient would predict.

One More Wrinkle: What the Thriller Genre Says

Step back from the individual coefficients and a single shape emerges. The model is describing where each kind of energy belongs in a 24 hour, and the rule is almost architectural.

The front of the hour wants density and heat. Sophisticated, specific vocabulary and open conflict are the two strongest positive signals in the opening act, and they do complementary work. The density establishes that the world is real and the stakes are serious. The anger establishes that the day has already gone wrong. Together they are the ignition, and the model rewards an opening that fires on both.

The back of the hour wants propulsion, and almost nothing else. Every Act 3 signal in the model is a penalty. Fear, disgust, and the warmth of bonding all drag the closing act in the wrong direction, because the closing act of a 24 hour has exactly one job: to point at next week with enough force that the audience cannot look away. Dread that ends in helplessness, revulsion that ends in nausea, and sentiment that ends in reunion all spend the momentum the cliffhanger is built to carry. The clock does not stop at the end of the hour. It hands off, and the handoff is the whole game.

Everything the model rewards or penalizes comes down to a single structural truth, which is that a real time thriller lives or dies on whether it launches the next hour rather than on whether it explains the current one. The opening of a 24 hour earns the hour you are watching, while the closing act earns the one you will return for, and that is why the show is rewarded for the density and conflict that ignite an episode and penalized for the fear, disgust, and sentiment that would let it coast quietly to a stop. Read in that light, the model is less a scorecard than a map of how 24 keeps its clock running, hour after hour and week after week.

So What Would You Actually Change? The Writer’s Takeaway

The point is not to write a more frightening show, which would be both obvious and, per the data, counterproductive. The implications are more specific, and more useful.

Positive relationships, consider emphasizing

+2.20 · Sophisticated vocabulary in Act 1. Let the opening sound expert. The procedural and geopolitical density that makes the world feel real is doing retention work, not getting in its way. → If your cold open has been simplified for accessibility, ask whether you have sanded off the authority that made it feel high stakes.

+1.59 · Anger in Act 1. Open in conflict. A first act with antagonism, threat, and confrontation starts the clock the audience came to watch. → If your opening act is still arranging furniture by minute ten, the day has not gone wrong yet, and the day going wrong is the show.

Negative relationships, use with caution

−1.59 · Generic vocabulary in Act 1. Common language signals common stakes. → If your opening act could be spoken on any control room set on television, it is not yet the opening of this show.

−1.72 · Disgust in Act 1. Do not open in revulsion. Moral nausea stalls the engine before it starts. → A thriller can feel like the natural home for the transgressive, but in the opening act it reads as friction. Save the recoil.

−1.34 · Affiliation in Act 3. Hold the bonding. The closing act is a launchpad, not a reunion. → If your final act turns toward warmth and connection, you have spent the momentum the cliffhanger needed.

−2.01 · Fear in Act 3. End in motion, not dread. The climax should propel, not paralyze. → Fear that ends in helplessness is not the same as a threat that ends in “watch what he does next.” Only the second one brings them back.

−2.58 · Disgust in Act 3. The largest penalty in the model. Do not end in revulsion. → A final act that leaves the audience nauseated rather than propelled is the surest single way to lose next week.

Technical Notes & Full Output

Appendix A · Glossary of All Seven Measures

Each feature was extracted using LIWC (Linguistic Inquiry and Word Count) and related computational text analysis tools, applied at the act level. The suffix _1, _2, or _3 denotes the act in which the feature was measured.

Significant positive predictors, more of this, more viewers next week

Sixltr_1 · Words over six letters, Act 1. Proportion of opening act words longer than six letters. Captures lexical density and a register of expertise and specificity, the sound of a world worth taking seriously. sig. positive · coef +2.20

Anger_1 · Anger language, Act 1. LIWC Anger category in the opening act: proportion of words associated with hostility, aggression, and confrontation. Captures the conflict that ignites the episode’s clock. sig. positive · coef +1.59

Significant negative predictors, more of this, fewer viewers next week

Dic_1 · Dictionary density, Act 1. Percentage of opening act words drawn from common vocabulary. High scores indicate a generic opening register, language that could belong to any procedural. sig. negative · coef −1.59

Disgust_1 · Disgust language, Act 1. LIWC Disgust category in the opening act: words of revulsion, moral violation, and transgression. Corrosive in a show whose promise is competence, not discomfort. sig. negative · coef −1.72

Affiliation_3 · Affiliation language, Act 3. LIWC Affiliation category in the closing act: words associated with social bonds, alliance, and friendship. Captures a climax that turns toward connection rather than propulsion. sig. negative · coef −1.34

Fear_3 · Fear language, Act 3. LIWC Fear category in the closing act: words of dread, threat, and apprehension. Captures a climax that ends in helplessness rather than forward motion. sig. negative · coef −2.01

Disgust_3 · Disgust language, Act 3. LIWC Disgust category in the closing act: words of revulsion and moral violation arriving at the climax. The largest penalty in the model. sig. negative · coef −2.58

Appendix B · Most Aligned Episodes

The alignment score measures something specific, and it is worth being precise about what that is: how closely an individual episode’s writing matched the full feature pattern the show level Random Forest identified as predictive of retention. A high alignment score does not mean an episode performed well. It means the writing was operating inside the show’s structural fingerprint, exhibiting the features the model found important, in proportions consistent with what it learned from the corpus.

With that definition in place, here are the most aligned episodes in the data, alongside the viewership of the episode that followed each one.

The first thing the table shows is where the most aligned episodes came from. They cluster in Seasons 7 and 8, the show’s late run. The single highest alignment score belongs to a Season 8 hour, and the ranked group is entirely a late seasons affair. The writing was more precisely calibrated to the retention pattern during the show’s later years than at most points earlier in its life.

The second thing the table shows is what that precision was actually worth, and the caution here is one worth stating plainly. Alignment is not performance. The most aligned hour in the data, Season 8 Episode 10 at 0.6216, was followed by 8.91 million viewers. The least aligned hour in the ranked group, Season 7 Episode 8 at 0.5526, was followed by 11.22 million, the strongest return in the entire dataset. The episode that best matched the learned pattern was not the episode that retained the most people. The ceiling moved independently of the fingerprint, which is precisely the gap the alignment score is built to reveal rather than hide.

The third thing the table invites is the question of why late season episodes dominate the very top, and the answer is a two part explanation. The first part is a training artifact. The Random Forest was fit on the full run, so the later seasons make up a meaningful share of the data it learned from. When late season hours score high on alignment, part of what is being measured is that the model recognizes them as structurally typical of the era it was substantially trained on. The second part is variance. Peak era 24 was taking more risks, running wilder plot mechanics, and spreading its episodes further from any central tendency, which produces a wider spread of alignment scores without implying weaker writing. That dispersion is exactly why a Season 1 hour can land at a respectable 0.5663 while other peak era episodes fall well below the group. By the late seasons the show had contracted into a narrower creative range, and less experimentation means more episodes clustering near the model's center of mass, which it reads as high alignment regardless of direction. Consistency is not correctness. The Season 1 hour is a useful reminder of the same point from the other direction: it aligns respectably and was followed by a strong 9.25 million, yet sits below two late season hours that drew fewer viewers the next week. Structural typicality and performance are different axes, and the table refuses to let them collapse into one.

Generalizability. Findings are specific to 24 and should not be applied directly to other series without replication. The emotional predictors in particular appear to be genre conditional rather than universal. What a register like disgust or fear means depends entirely on the promise a show has made to its audience, and 24's promise is a specific one. Cross genre and cross network analysis is ongoing. The vocabulary penalties appear to travel across procedurals. The emotional registers, so far, look more local to the show.

I want to take a brief respite from my typical empirical work. I’ll get back to this next week!

Buon pomeriggio! As I write this, I am sitting in my chic, modern hotel lobby, The Market Urban Hotel, or, given its sustainable bent, perhaps this is more like a vestibule. Soon, I will take a train to Lucca, Italy, where I’ll be for the rest of the week at Scuola IMT Alti Studi Lucca | IMT. I’m here for an AI, data science, and digital twins workshop with other scholars, with the primary aim of building new research collaborations.

Yesterday, I went to the Uffizi Gallery, and this morning I attended the Bargello and Accademia.

Though this Substack is typically anchored in empirical work, I felt an obligation to write this. At the end, I’ll connect these reflections back to my TV script dashboard analyses, and to the forthcoming dataset surrounding 1,400+ movie scripts, 10-K parent financial data, and earnings call transcripts (look out for articles this fall). But as a cultural economist, drawing out what compels consumers toward art in the first place feels like a remarkably important conversation, both for grounded truth and for whatever value my research can contribute to academia and industry.

Speed (or lack thereof)

It quickly became apparent how critical it was for these works to take time. When was the last time you truly took your time with anything? Had the luxury of unlimited time, or even a modicum of flexibility? Our hurried, unremitting lifestyles, particularly in urban centers, have dulled our capacity to sense anything beyond the immediate scope of what needs to get done.

Sculptures like this one, Donatello’s David, didn’t take a half day, a morning or even a few days. It is estimated that he worked on this for 2-4 years! 2-4 years! Most of us can’t be bothered to wait 2 years for another season of a beloved TV series, or fall off after waiting endlessly for a movie sequel (though it took 20 years for a sequel to The Devil Wears Prada, and the second movie so far has grossed $557m against a $100m production budget, rendering it a MASSIVE hit).

There is something more important here: how fast we move through life. The more we consume, do, touch, and feel, the more we are forced to compartmentalize, reduce, and compress these experiences. Our minds, after all, have only so much bandwidth for memory. Work like Donatello’s could only have been born in an era when things were slow, thoughtful, and pensive, when craftsmanship commanded a premium, expectations were lower, and self-actualization with analog tools was not only possible but expected. Perhaps we miss being slower, working slower, engaging slower, and moving through life slower. Art forces us to sync our intellectual curiosities and emotional heart beats with the object itself, slowing us down.

Inspiration

With the ability to day dream, plan, think, challenge, and carefully craft, inspiration could fully unfurl. I welled up with tears and felt shivers seeing Michaelangelo’s David (see below).

The ability to craft something so tall, so big, and so graceful (how does it NOT tip over?!), speaks to the quiet brilliance and ingenuity that Michaelangelo possessed. He also had sculptures that purposely were incomplete, to illustrate the finite and infinite, and to have us reflect too, on the journey. Rarely are we satisfied at the end of a journey; it is the journey itself that is often the reward. The end is just ceremony.

Viewing this reminded me not only what my Italian ancestors are capable of (along with other art, music, food, wine and many other pursuits), but of what any of us can become through sustained effort. I am still deeply moved by Christopher Nolan’s cinematography, sound, and direction in the movie Dunkirk (highly recommended). Art reminds us what we are capable of: the skills that we can cultivate, the grinding that we can endure, and the rewards that can be reaped by any of us. While we are all limited in what we can do, I don’t think any of us are limited in our ability to master something, be it a job or skill.

Luxury

Art patrons carry luxuries that others don’t. The wealth, time, and investment required to travel and experience art in person are an unambiguous privilege. And not everyone who makes the trip does so out of pure love or intellectual curiosity. For some, there is the constant pressure of generating content for social media, carefully scaffolding a profile designed to attract mates, or simply attention. There is a status transference at work in the selfie next to a masterpiece: brand equity borrowed from a work viewed by millions, transferred to someone viewed by perhaps tens of thousands. During the Renaissance, commissioning a painted likeness of yourself beside a work of art would have taken days and a small fortune, so the impulse was never indulged. Now the economics are frictionless, and we are all, in a sense, out on display as our own self-aware art pieces, attempting to imbue ourselves with the same sentiment. Social media has spurred humans to serve as their own mobile, dexterous paintbrushes, painting the reality around them in hues, tones, and schemes that fit their desired projected lives.

Seeing Leonardo da Vinci’s Mona Lisa two summers ago was spectacular. Given its fame, I was genuinely stunned to find myself six feet from the painting. With the rise of a much larger millionaire class in the United States and the glacial erosion of the middle class, experiences that mark distinction have become harder to come by, or, they have simply been tiered upward for high-net-worth individuals, as Disney has done with its theme parks. The underlying hunger is the same: to feel that you have come a long way intellectually, spiritually, or monetarily.

Consumption

You could walk right past Raphael’s cupid painting (see below). Technically. But the scope, the density, and the sheer range of what surrounds you here means that the time required to consume any given piece varies enormously, and is ultimately determined by what the viewer brings to it. The art doesn’t change. You do. That’s the whole point.

Does it move you? Why are you here in the first place? What were you thinking about before the painting? What about during it? And after it? Did it connect to an event, an opinion, or perception? As I mentioned earlier, when I stood before Michelangelo’s David, I welled up with tears and felt genuine shivers. It represents the pinnacle of what a single human being could accomplish by hand, the delicate precision and vision required to chisel away at something hard, impenetrable, and unyielding in its raw form, and to completely transform it into a preconceived design. Empowering isn’t quite the right word, but it’s close. Around me, others stood in similar awe, while others took a quick photo and moved on. Neither response is wrong, but they are not the same experience.

Full consumption requires slowing down. While TikTok has wreaked genuine havoc on our neural pathways and our capacity for sustained attention, the experience of standing before great art, reflecting, having an internal conversation, perhaps an external one, simply takes time. It demands thought. It may even require you to challenge your own preconceptions, sit with ambiguity, and remain open to being wrong.

There is a useful comparison in the food. If you eat a bag of chips in four minutes, is that more satisfying than spending ten minutes with a rich, layered tiramisu? For some, perhaps. However, people come to Florence for the tiramisu. Each moment more deliberate than the last, irreproducible, specific to this place and this menu. You cannot order it somewhere else. You cannot get it delivered. TikTok, by contrast, is engineered to be the opposite, a service that gets better the more you use it, learning your preferences, expanding its offering, ensuring there is always more.

Age

Permanence is a preoccupation of our moment. The ability to live forever, biohack our way to longevity, and engineer some form of continuity beyond death. We are obsessed with outlasting ourselves.

And yet, figuratively speaking, unless you’ve seen the Night at the Museum movies, these paintings and sculptures have been quietly laughing at us for centuries. They were here before we arrived and will be here long after we’re gone. Think about the sheer volume of human experience that has washed past them: depressions and periods of excess, world wars and long peacetimes, plagues and recoveries, revolutions and restorations. Through all of it, people still made the journey to stand in front of these works, still craning their necks at the same ceiling, where they are still welled up in the same rooms.

Why are we fascinated with permanence? For one, it is in great contrast toward the fleeting nature of life, its impermanence, and the pressure it thrusts upon us toward finding and making meaning. We are attracted to things that give us meaning, particularly artifacts from an older generation. Perhaps we are amazed with how much could be accomplished before with little to no technology, where people were tightly married to their own work, tools, and other boundary constraints in their lives.

Transformation

We do not leave museums the same as we entered them. Something shifts, not always dramatically, not always consciously, but something shifts. The encounter with great art plants a question that takes days or weeks to fully surface. What am I doing with my time? What am I building that will outlast me? What would I make if I had two years and no distraction?

There is something primal in the transformation of raw earth, the assertion of human will over raw material, the impulse to take something unformed and press our essence into it. To take marble, which has existed for millennia without being touched or shaped, and impose upon it a new form. To leave a mark. To see ourselves reflected back in what we have made. We are deeply, perhaps irreversibly, obsessed with that mirror. And we gain something profound from the transformative process itself, watching human hands turn stone into giants, gods, and athletes. Into versions of ourselves that are larger, stronger, and more permanent than we will ever be.

I don’t know that I have clean answers to any of those questions. But I think the asking of them is the point.

Ideal self

Michelangelo didn’t carve what men look like, but instead he carved what we imagine we could be. David is not a portrait, as he is an aspiration rendered in marble. Proportionally impossible, serenely confident, caught in the moment before courage rather than after it. The Renaissance ideal of the human form was never documentary; it was prescriptive. This is what the body could be, what the spirit should reach for.

Religion supplied the original architecture for this impulse. Religion as authoritative in nature and scope, and really oppressive. It dominated most facets of life. Renaissance patrons didn’t fund altarpieces and chapel ceilings purely out of devotion, they were, at least in part, purchasing proximity to the divine. Having yourself painted into a biblical scene, donating a nave, commissioning a pietà in your family’s name: these were bids for a better afterlife, attempts to launder the earthly self into something more transcendent. The ideal self extended beyond death, and art was the vehicle. Sound familiar? Consider the billionaire class today, funding longevity research, cryonics, and mind-uploading projects, earnestly attempting to buy real estate toward their own version of the divine. The afterlife has been rebranded as a server farm, but the impulse is identical: transcend the mortal self, extend the legacy, refuse the finitude that everyone else accepts. The Medicis had Botticelli. Bezos and Musk have biohackers and large language models. The cathedral has merely changed shape, and we all bend the knee accordingly, ChatGPT, Claude, and Gemini serving as rival gods within the same restless religion.

What strikes me now is how little the structure has changed. The selfie next to the Mona Lisa is, in its own way, an altarpiece. Where a Medici banker once paid Botticelli to render him in the company of saints, we now press our faces next to masterworks and upload the result to an audience of thousands. The aspiration is identical, proximity to greatness as a transfer of meaning, a quiet argument that we too belong in the frame.

Our ideal selves, though, have morphed considerably, from cherubic, ebullient angels and chiseled embodiments of divine proportion to hyper-sexualized, rage-optimized influencers whose currency is provocation. The algorithm rewards the extreme. Anger travels faster than beauty. I'm not sure this is the ideal self worth aspiring to, and I find myself wondering whether staring up at a marble David, calm, proportioned, caught in a moment of quiet resolve, might simply be a healthier place to look.

Mourning our loss of connection to our work process

I took a step back and looked at David towering over me and everyone else in the room. Why was this such a big deal, beyond the self-evident qualities? Part of the answer, I think, is simply the act of looking up. With the decline of organized religion and the rise of the phone, we don’t look up the way previous generations did. We look down. We have traded the cathedral ceiling for the screen in our palm, and we have done it so gradually that we barely noticed the exchange. To look up at the sky, a piece of art, or anything for that matter, is something that we do less of as human beings.

There is a grief in what we have given up. Most of what I produce these days, papers, platforms, data pipelines, slide decks, passes through a keyboard. It is mediated, abstracted, and increasingly automated. I copy and paste images into software; I don’t make the image, color it, or draw it. I type rather than write. I cannot hold my regression output, and frankly I probably couldn’t calculate it by hand anymore either. All modeling techniques are now largely automated. I cannot run my hand along the edge of a coefficient estimate and feel its texture. The distance between my hands and what I make grows wider with each passing year.

The Florentine craftsmen were attached to their work in a physical, sensory way that is increasingly difficult to replicate. They could touch it, feel it, fine-tune it. The feedback loop between intention and output was immediate and tactile. What I do instead is click “Run” and wait. There is real skill in that, I’ve spent years in graduate coursework grinding toward something like expertise, but there is a different kind of satisfaction that I think we have quietly, gradually mourned without ever quite naming it.

The satisfaction of taking raw materials from the world and reshaping them into something wondrous is an effervescent feeling, and an increasingly rare one. Think about how significant cooking classes have become precisely as fewer and fewer people actually cook. There is something irreplaceable about having made something yourself, a sourdough loaf, a chord progression on the guitar, a painting, a sandcastle built with your daughter on a beach somewhere. That simple act is lost on those of us conditioned to fill every second with an activity, a screen, an image, a song. There is no daydreaming anymore, as we seek to fill each moment with something that appeases us, rather than allowing boredom to creep in.

Art transports us back to a simpler time, sure, but what else are we craving when we seek it out? What are these primitive, ineffable nudges that compel some of us to stand in front of a painting, seek out older media, or wander through antique shops? What, precisely, has technology removed from us?

My cousin works at SNL and can now 3D print masks, props, and makeup pieces. It obviously accelerates the process and likely achieves a precision that hand-sculpting couldn’t always guarantee. But the trade-off is that he didn’t physically make the object himself. I feel this too, as I used to code line by line, which I hated, but now I automate much of it through tools like Claude Code. I can still read code fine. But I am, without question, a worse coder for it. When we optimize, we make trade-offs: we gain efficiency while relinquishing agency and direct contact with the work. The question worth sitting with is which tasks, when removed from our hands, also remove something from us. Which ones made us feel capable? Powerful? Self-actualized?

What does it mean to be increasingly removed from the manual process of making things? Do we take less ownership of the output? Find less meaning in it? Coupled with the loneliness crisis quietly metastasizing in this country, we have fewer and fewer meaningful relationships to fill the space left by that loss, to push back against the dreariness of automated tasks, or to name the particular emptiness that optimization leaves behind.

Relation to TV scripts automation

While I believe my work analyzing TV scripts through natural language processing, machine learning, and econometric modeling genuinely serves artists and the industry, the irony isn’t lost on me. I am applying automated quantitative analysis to something that, whatever mathematical properties may underlie it, is ultimately a qualitative project. Art resists the ledger, as it always has, and as it probably should. That’s the job of art financiers.

I can establish guardrails, identify patterns, and surface what has historically resonated with audiences, but prescribing to an artist what to do with any of that is a different matter entirely, and probably a foolish one. Data narrates the past with some ability to illuminate the future. It has never been very good at telling anyone what to make next. Over time, this may change, as there is a cyclical nature to what is in vogue, what is not, and where there are gaps in entertainment, and media. However, there is also a serendipity that should be left alone to flourish on its own accord. We cannot control every aspect of creation, nor should we aspire to do so.

Scioglimento & Epilogo

As my bloodline pumps through my arteries and veins through my heart, composed of a complex but enviable blend of red wine and marinara sauce (with notes of basil, naturally), I couldn’t help but think about what the human condition is capable of when unencumbered by distraction. What an uninterrupted period might look like. Without noise. Without the ding of social media notifications. With quiet, contemplative inner monologues ruling the day, long lunches, and slow, deliberate conversations along the way to polished work. Craftsmen attached to their objects. Able to touch, feel, and refine, not through a keyboard or semi-automated processes, but through something organic, earthly, and even romantic in the way their work was conducted.

I wonder: if we still had to write and sign letters, would we leave such awful comments on each other? Has the friction that technology removed cheapened our expressions, and perhaps ourselves? There is even a consumer behavior concept that fits here, effort justification, the well-documented phenomenon where we value things more when we have worked harder for them. The IKEA effect, the endowment effect, sunk cost reasoning, all variations on the same theme. Art may be the purest expression of that principle at civilizational scale. We treasure it because we know, at some level, what it cost. And increasingly, we make things that cost nothing, and feel exactly that way. Of course, our tech overlords benefit from the work that consumers do, though we are often not compensated for the digital footprints we leave behind, and are simply told it’ll benefit us if we enable tech firms to measure us.

The irony is not lost on me that this workshop is happening in a country celebrated for its ancient mystique and, in many ways, still rooted in a more relaxed rhythm of life. While we seek to accelerate, optimize, and automate our way to doing more, Italians, and many Europeans, have opted for something closer to the opposite. And yet the people here look healthier and happier. The food tastes better because it is better. A two-bedroom condo here averages just under $300,000, and a three-bedroom runs about $400,000. For this New Yorker, that is a bargain for living in a major city. There’s something about living among beautiful things, art, food, people and culture, but not in excess. There’s a simple intentionality that goes with living as a European that, increasingly, speaks to me.

Maybe the more we rush to optimize everything, the more we risk optimizing away the very things that make life worth measuring.

PS - Yes, this means I’ve seen all four artists the Teenage Mutant Ninja Turtles are based on. I quickly realized that today, too =).

I recently spoke with Michael Daley, former writer on CSI: Crime Scene Investigation and now a professor at Loyola Marymount University, about the craft of television writing, the evolution of procedural storytelling, and what it means to teach narrative in an era increasingly shaped by data, platforms, and AI-assisted creativity.

Daley’s time on CSI: Crime Scene Investigation—one of the defining procedural dramas of its era—offered a front-row seat to a very particular kind of storytelling machine: tightly structured episodes, forensic detail as narrative engine, and a writers’ room calibrated for both repetition and reinvention. He described the experience as “writing within constraints that paradoxically force creativity,” where every episode had to satisfy the procedural logic of the crime while still delivering emotional stakes for characters who, over time, became as familiar as the cases themselves.

Disclaimer: Zoom was a bit spotty so I apologize for some of the screen-in-screen going on during portions of the video recording. I’m not sure how it happened (as I didn’t see it on my screen), but I’ll correct it for next time. Nonetheless, this was an absolute delight to create!

Enjoy,

Anthony

In my fervent and frenetic need to prove myself as a media economist/machine learning whatever, I pushed myself (and, frankly, pushed AI pretty hard too) to code and develop a tool that can break down movie and TV scripts in an organized and thoughtful manner. I hope to make this commercially available in August, but for now I am eager to share it while it is still being built.

The purpose of this tool is to aid writers, producers, and business development professionals in breaking down scripts more holistically. Imagine a shared dashboard that, at the very least, gives every stakeholder a common analytical foundation they can actually agree on. I know what you’re thinking: isn’t this just technical lubricant pushing our beloved content toward a collective regression to the mean? That is truly not my intention. I would be the first to reject the notion that data should force the artist’s hand in some I-know-better-than-you way. Nonetheless, given the rate at which TV series are cancelled, and as a former Disney executive reminded me just yesterday, audience retention has become an existential problem.

This is a genuinely new challenge for streaming services. It is difficult to build an audience from season to season when streaming economics reward audiences for chasing whatever is hot at any given moment, and that slate is ever expanding. The predictable cadence that once governed TV schedules at least gave audiences a reason to commit to a handful of shows and follow them over time. Now there is a visible run toward wrapping series in three to four seasons, shortening the dedication required and cultivating loyal but small audiences in the process.

With all of that as backdrop, I envision this tool, which relies on large language models, but is programmed strictly to read and analyze rather than learn or generate, as a touchstone for every stakeholder involved in a project. To put it through its paces, I tracked down season one, episode one of Stranger Things entitled “The Vanishing of Will Byers”. Personally, this episode just grabbed me. Full stop. If memory serves right, I watched the next two episodes before I had to go to bed. I was absolutely hooked and, frankly, mesmerized by the attention to detail, the plot, and knowing full well, as each scene unspooled in our dark living room, how many micro-decisions were made to figure out what to tell the audience, and what not to tell them. To this day, this particular narrative device/tool/framework amazes me, and the very best know exactly what to tell the audience, at the exact moment, in the exact manner.

Moving on! To showcase the tools prowess, I wanted to walk through how the tool works, and what I think it can offer.

Most importantly, I’m eager for feedback. I’m thinking of ways to program this to provide even better and clearer analyses for all stakeholders involved. Feel free to find me on LinkedIn - https://www.linkedin.com/in/aphd/. Happy to chat.

What the First Episode Actually Did

The Duffer Brothers were working with a remarkably constrained premise in the pilot. A kid disappears. A grieving mother starts losing her mind. A group of twelve-year-olds ride bikes and argue about Dungeons and Dragons. That is, on paper, not a lot.

But the structure of that first episode is doing enormous work beneath the surface. When I ran it through the tool, a few things became immediately clear.

First, the emotional register is set within the first ten minutes and never wavers. The opening sequence drops you into pure dread. Will Byers sees something in the road, crashes his bike, hides in a shed, and disappears. We do not see what takes him. That restraint is doing a tremendous amount of heavy lifting. The fear is ambient, not explicit, and that ambient quality is what allows the show to sustain tension across an entire season rather than blowing its budget in episode one. Moreover, bicycling riding, particularly for those of us who grew up in the 80s, 90s, and maybe 00s, was indicative of childhood, and pre-adolescent freedom just before getting a driver’s license to drive a car.

Second, every major relationship is introduced with a conflict already baked in. Mike, Dustin, and Lucas are not just friends. They are a small social ecosystem with a hierarchy and a set of loyalties the audience has to learn. Joyce Byers is not just a worried mom. She is a woman the town already does not quite believe, which means her arc has a structural antagonist before the supernatural element even enters the frame. Jim Hopper is not just a small-town cop. He is a man with a private grief he has not processed, and the show is betting that grief will make us root for him even when he is doing something we probably should not endorse.

Third, and this is where I think the pilot earns its reputation, the show makes a conscious bet on the audience. It assumes you will tolerate not knowing things. The Upside Down is not explained. The creature is not named. The government subplot is introduced with zero context. A lesser pilot would have papered over those gaps with exposition. The Duffer Brothers left them open, and in doing so, they made the audience a participant in the mystery rather than a passenger.

A Framework You Can Actually Use

So what can writers, producers, and development executives take from this? After running a number of pilots through my tool, I have started to see a pattern in the ones that hold audiences across multiple seasons. I think of it as the Three Debts framework.

The first debt is emotional. Who do I care about, and why do I feel like something bad is going to happen to them? This has to be established in the first fifteen minutes. Not over the course of the pilot. Not by the end. In the first fifteen minutes. Stranger Things does this with Will and with Joyce, separately and quickly.

The second debt is intellectual. What do I not understand yet, and why does that feel like a promise rather than a frustration? This is the hardest one to calibrate. Too little mystery and the show feels flat. Too much and the audience disengages because the gap between what they know and what they need to know becomes insurmountable. Stranger Things threads this by giving you one clear mystery (where is Will?) while seeding two or three secondary mysteries (what is the government doing? what did Will see in the road?) that will pay off later. It is up to the writer to figure out whether to satiate or challenge audience members, or perhaps thought of along a continuum, what the proper blend is that is true to the story and involved characters.

The third debt is social. Who is the ensemble, and what are the fault lines between them that will generate conflict without requiring an external antagonist? The Duffer Brothers understood that Mike and Lucas arguing about whether to trust Eleven is not a subplot. It is the engine. Internal friction in an ensemble is what keeps a show alive in the quieter episodes when the mythology is not advancing.

If a pilot satisfies all three, it is in strong shape. If it satisfies two, it can probably survive on execution. If it only satisfies one, it will likely get one season and a very respectful cancellation announcement. There has to be a reason to follow the characters, just as in movies, there has to be a reason to pay attention to the plot (though sometimes movies can be character driven, and sometimes TV series can be plot driven). Of course, this is a working framework, and based in musings and attempts I’ve made in analyzing scripts.

This is the Plot A/B/C overview tab from your tool analyzing the Stranger Things pilot. Here is what it shows:

The episode contains 94 scenes across 33 characters, with 6,946 words of dialogue and exactly 3 plotlines.

The scene distribution breaks down as follows: Plot A carries 43 scenes (roughly 46% of the episode), Plot B carries 38 scenes (about 40%), and Plot C carries 12 scenes (about 13%).

What is notable here is how unusually balanced the A and B plots are. Most pilots front-load the A plot heavily and use B and C as thin connective tissue. The fact that Plot B is nearly as dense as Plot A suggests the Duffer Brothers were running two near-equal narrative engines simultaneously from episode one, which maps directly onto the Three Debts argument in the article: the ensemble friction between the kids storyline and the Joyce/Hopper adult storyline was never a subplot. It was co-equal weight from the jump. Plot C, likely the government/lab thread, is deliberately thin, which is exactly right for a mystery seed that is not meant to pay off yet.

This is the Character Analysis tab, and it is genuinely revealing. Here is the breakdown:

The numbers: Joyce leads all characters at 16.5% of dialogue, followed by Benny at 11.8% and Hopper at 9.6%. The top three together account for 37.9% of all spoken words. After that, the distribution flattens quickly across Dustin, Mike, Nancy, Steve, and Will in the 4 to 8% range, with everyone else below 4%.

What is interesting analytically:

Benny at 11.8% is a storytelling sleight of hand. He is essentially a one-episode character who exists to introduce Eleven to the audience and then disappear. The fact that he holds the second-highest dialogue share in the entire pilot tells you how much narrative scaffolding the Duffer Brothers needed to make that introduction feel earned rather than convenient. He is doing functional work disguised as character work.

Will Byers, the character the entire episode is nominally about, speaks only 4.2% of the dialogue. He vanishes before he can accumulate words. That is a deliberate structural choice: the show is not about Will so much as it is about what his absence does to everyone around him. The pilot encodes that asymmetry right into the dialogue distribution.

The key insight your tool flags, highly distributed with a steep drop-off after the top tier, maps directly onto the ensemble engine argument in the article. No single character owns the show in episode one. Joyce is the emotional center but not an overwhelming one, which leaves room for the audience to attach to multiple characters simultaneously.

This is another strong data point to weave into the piece, specifically into the section on ensemble friction and the third debt.

This is the Character Presence Map, and it adds a spatial dimension to everything the dialogue share chart already told us. Here is the read:

The quadrant logic: Characters in the upper right (high scenes, high words) are true leads. Characters with many scenes but fewer words are supporting infrastructure. Characters with few scenes and few words are cameos or functional roles.

What stands out:

Joyce is a complete outlier, nearly 12 scenes and over 1,100 dialogue words. She is not just the emotional lead, she is carrying more narrative weight than anyone else by a significant margin on both axes simultaneously.

Benny is the most structurally anomalous point on the entire chart. He appears in roughly 8 scenes with over 800 words, which puts him in lead territory on dialogue but mid-range on presence. That confirms what the dialogue share showed: he is a high-intensity, short-duration character, a narrative burst rather than a throughline.

Will is the quietly devastating data point. He sits at around 7 scenes but only 280 words, noticeably below where a character with that scene count should land. The show keeps him physically present just long enough to register his absence when it comes, but it never lets him accumulate enough voice to feel safe.

One of the features I am most proud of in this tool is the Character Emotion Fingerprint, a radar chart that maps each character’s emotional distribution across seven registers: joy, anger, sadness, fear, surprise, disgust, and neutral. Run it on Dustin and you get something that initially looks underwhelming. His fingerprint is anchored heavily at Neutral, with a secondary spike toward Joy and modest traces of Surprise and Fear. Sadness, Anger, and Disgust are essentially flat.

The tool classifies him as a Neutral archetype at 30.5% of his emotional range, balanced and varied, mixed tone overall. That sounds like a polite way of saying he is not doing much. It is actually the opposite.

Dustin is the ensemble’s pressure release valve. When a scene needs levity he provides it. When it needs a grounded, human reaction he provides that too. He never tips into grief or rage because the show is consciously reserving those registers for Joyce and Hopper. His neutrality is not a character flaw. It is tonal architecture. He is what keeps the kids’ plotline from collapsing into pure dread while the B plot burns around Joyce.

The fact that the tool reads him as emotionally complex and varied despite that neutral anchor is the tell. A flat character would cluster tightly in one emotion. Dustin spreads across the chart in a low, wide shape. That spread is what makes him feel real rather than functional.

Now imagine that fingerprint sitting next to Joyce’s. You would see almost a mirror image: fear and sadness dominant, joy nearly absent, the shape pulling hard toward the upper left of the chart. Two characters, two completely different emotional grammars, both necessary, neither redundant. That visual contrast, I think, is one of the cleaner arguments the tool makes for why this pilot worked as well as it did.

The Character Co-Presence Network maps which characters actually share scenes together, and the shape of that network tells you something a script breakdown alone cannot: where the social gravity of the episode actually lives.

The three network hubs, the most connected nodes in the entire graph, are Mike, Dustin, and Lucas. Joyce and Hopper do, of course, come close. Still, the three twelve-year-olds are the structural center of this pilot, which is either a tremendous creative risk or a very confident bet on an audience that still remembers what it felt like to be that age. Probably both.

What the network also reveals is that the A and B plots are nearly separate social galaxies in episode one. The kids cluster on the right side of the graph with dense connections running between them. Joyce and Hopper occupy their own gravitational zone, connected to each other with three shared scenes and to the wider network through Hopper’s bond with Officer Callahan, the strongest single pairing in the episode at four shared scenes. The two worlds are not isolated from each other entirely, but they are not yet intertwined either. That separation is intentional. The show is building two distinct emotional worlds that the audience has to hold simultaneously, and it is not going to collapse them together until it has earned the right to do so.

Then there is Barbara. Peripheral, limited interactions, one thin edge connecting her to the rest of the cast (RIP!).

The Character Introduction Timing table answers a question that sounds simple but is actually one of the harder structural problems in pilot writing: when do you let someone speak?

The kids come in almost immediately. Dustin, Mike, Will, Lucas, and Karen all enter at scene 6, which is 6.4% of the way through the script. That is essentially the cold open. The Duffer Brothers wanted you bonded to that group before anything else happened, and the timing confirms it was not an accident.

Then there is a deliberate pause before the adult world arrives. Hopper does not speak until scene 34, nearly 36% into the script. Joyce, despite being the emotional anchor of the entire episode and the character who ends up with the most total words at 1,148, does not appear until scene 24, about a quarter of the way in. The show makes you wait for her. It builds the kids’ world first, establishes what normal looks like, and only then introduces the adult infrastructure that will have to carry the grief.

Benny is the most structurally interesting entry in the table. He arrives at the exact midpoint of the script, scene 47, 50% in, with 819 total words. He enters late, speaks enormously, and exits permanently. The show essentially inserts a guest lead at halftime to bridge the gap between Will’s disappearance and Eleven’s emergence as a character. When you see it laid out this way, the precision of that timing is striking.

The checkmarks next to every entry in the Notes column indicate that all major characters are introduced within acceptable development windows. No one arrives so late that their presence feels unearned. That is a cleaner structural record than most pilots manage.

This table is where the tool gets genuinely granular. Rather than summarizing a character’s emotional profile across the whole episode, it breaks down emotion scores by character, plotline, and act, so you can see how a character’s emotional register shifts depending on where they are in the story and which plot thread they are operating inside.

A few things jump out immediately.

Barbara’s dominant emotion in her only act is Surprise at 0.6365, which is nearly double any other score in her row. That is the show telling you, at a keyword frequency level, that Barbara exists in the pilot primarily to react. She is a witness and a sounding board for Nancy, not an agent of her own story. The network chart already showed her on the periphery. The emotion table shows her frozen in a posture of reaction right up until she disappears.

Benny’s arc across his three acts is a clean emotional escalation. He moves from moderate Surprise in act one to Fear dominating at 0.7039 in act three. That is the shape of a character who starts out curious and ends up cornered. The tool is essentially tracing his death in emotional frequency data before it happens on screen.

Dustin’s C plotline entries are worth pausing on. In the scene where he mentions the Demogorgon, Fear spikes to 0.5269. That is the highest fear score in the visible rows for any of the kids, and it is attached to what the audience initially reads as a game reference. The show is seeding genuine dread inside what looks like playful genre banter. The kids are scared before they know they should be scared, and the emotion data catches it.

Dr. Brenner in act two registers Fear at 0.1981 in his own dialogue, which is quietly unsettling. He is the one who should be in control. The fact that fear is present in his language at all suggests the show is already encoding instability into the villain before the audience has any reason to distrust him.

Here is the Substack-ready version:

The Three-Act Structure Analysis is probably the tab I find myself returning to most, because it shows the episode’s skeleton in the most unambiguous terms.

The script divides as follows. Act one covers 25% of the script, 3,026 words across 23 scenes with only 7 speaking characters. Act two covers 50%, 6,052 words across 47 scenes with 27 characters. Act three returns to 25%, 3,027 words across 24 scenes with 17 characters. The word counts for acts one and three are nearly identical down to the single digit, which is either an extraordinary coincidence or evidence of writers who understood symmetry at a structural level.

The character explosion in act two is the most striking number on the page. The show goes from 7 characters to 27 in a single act. That is not escalation. That is detonation. The world cracks open in the middle of the episode and everyone comes flooding in simultaneously, which is exactly what it feels like to watch it.

The words per character metric tells its own story. Act one gives each character 432 words on average. By act two that drops to 224, and by act three it falls further to 178. As the cast expands, each individual voice gets compressed. The pilot is deliberately narrowing the bandwidth per character as the stakes rise, which creates a mounting sense of urgency without the show ever having to announce it.

Then there is the emotion tagging. Act one is Fear. Act two is Surprise. Act three returns to Fear. The show opens in dread, pivots to discovery, and lands back in dread. It does not resolve. It does not offer comfort. The pilot ends where it began emotionally, which is precisely why you cannot stop watching.

The Word Count Distribution and Scene-by-Scene Pacing Map together are the closest thing the tool produces to an EKG of the episode. If the three-act table showed you the skeleton, this shows you the heartbeat.

The act distribution is visually stark. Acts one and three are gray, nearly identical in height at 3,026 and 3,027 words respectively. Act two is a deep blue tower at 6,052 words, exactly double either bookend. The color coding is not decorative. The tool is signaling that act two is where the gravitational mass of the episode lives, and the symmetry of the outer acts is what makes that central weight feel controlled rather than chaotic.

The scene-by-scene pacing map is where it gets interesting. Each bar is one scene, height represents dialogue word count, and color represents plotline: blue for A, yellow for C, and light blue for B. What you see is not a smooth escalating curve. It is a series of spikes and valleys, which is exactly right for a pilot that is managing three simultaneous narrative threads across 94 scenes.

A few specific moments worth noting. The tallest spike in the entire chart lands around scene 52, deep in act two, in the B plotline. That is almost certainly the Benny and Eleven sequence, the longest sustained dialogue exchange in the episode, which tracks with everything the character presence map already told us about Benny’s outsized word count arriving at the midpoint.

The yellow C plotline bars, representing the government thread, are sparse and low throughout. Brief intrusions, never dominant, always receding. The Duffer Brothers are keeping that story on a very short leash in episode one.

And the valleys matter as much as the spikes. The clusters of near-zero bars scattered across the map are action and visual scenes, moments where the script stops talking and trusts the image. A pilot that never goes quiet is a pilot that is afraid of itself. This one is not.

Here is the Substack-ready version covering all three plots:

The Plot Summaries tab is where the tool steps back from individual data points and tries to read the episode as a whole. What it produces for each plotline is part structural summary, part thematic diagnosis, and it is worth walking through all three because together they explain something important about why this pilot holds together the way it does.

Plot A is the main engine: 67 scenes, 5,098 words, 28 characters. Will disappears, Joyce unravels, Hopper investigates, Hawkins Labs lurks in the background. The tool identifies the theme as the terror of losing a child and the desperate, helpless love that drives a parent and a community to confront the unknown. That is accurate, but what the act breakdown adds to it is instructive. Act one carries 31 scenes and only 1,370 words. Act three compresses to 12 scenes but 1,807 words. The episode is doing the opposite of what most pilots do: it starts wide and visual, letting the world breathe, and then tightens into language as the emotional stakes arrive. Joyce is carrying 37% of the A plot dialogue on her own. She is not sharing the lead. She is the lead.

Plot B is the most economically constructed thread in the episode: just 6 scenes, 1,042 words, 7 characters. Nancy navigating the early pull of Steve Harrington while Barb watches with quiet alarm, Jonathan connecting briefly with Joyce. The dominant emotion is Surprise at 35%, which is the right register for a plotline about teenagers discovering that the social world they thought they understood has different rules than they expected. What the tool’s summary captures cleanly is the theme: the tension between social belonging and personal integrity, and the way romantic distraction can blind you to the dangers around you. Barbara’s danger, in other words, is seeded in Plot B’s emotional logic before the Upside Down ever touches her.

Plot C is the sharpest piece of structural writing in the pilot, and the tool’s summary makes the argument plainly. Benny finds a shaved-head girl stealing food, feeds her ice cream, calls what he believes are social services, and is killed by the agents who arrive instead. Twenty-one scenes, 806 words, 7 characters. Benny carries 71% of the dialogue in this thread, which confirms everything the character presence map already showed: this plotline exists almost entirely to give Eleven a human face and a moment of genuine warmth before the show takes both away. The dominant emotion is Surprise at 33%, and the act breakdown shows the thread backloading heavily into act three at 48% of its words, which is when Benny dies and Eleven disappears into the night. The tool identifies the theme as misplaced trust and the danger of innocence, a moment of ordinary human kindness exploited and punished by institutional malevolence. That is probably the most precise single sentence summary of what Plot C is doing, and it took a language model reading keyword frequencies to surface it this cleanly.

Three plots, three distinct emotional grammars, all converging on the same structural truth: this episode is about what institutions, whether family, government, or social hierarchy, do to the people who fall outside their protection.

The last chart is in some ways the most clarifying one, because it answers a question that sits underneath everything else we have looked at: how much of this pilot is the Duffer Brothers trusting the camera over the page?

The answer is: most of it. 67 of 94 scenes, 71%, are classified as Action/Visual, meaning fewer than 100 dialogue words. Only 22 scenes are Mixed, and just 5 are Dialogue-Heavy. The tool flags this with a note that is worth quoting directly: action/visual-driven structure, ensure key character moments are not being underwritten.

That is a reasonable caution for most scripts. For this one it is almost the wrong question. The Stranger Things pilot is not a show that is avoiding dialogue because it does not know what its characters would say. It is avoiding dialogue because it understands that dread is a visual and physical experience before it is a verbal one. Will does not explain what he sees in the road. Joyce does not articulate her grief in a monologue. Eleven says almost nothing. The show is making a deliberate argument that silence and image carry more emotional freight than words in the registers it is working in.

The five Dialogue-Heavy scenes are the load-bearing moments: Benny and Eleven at the diner, Joyce on the phone trying to reach Will, Hopper at the station piecing together what happened. These are the scenes where language is allowed to do real work because the emotional ground has already been prepared by everything the camera showed you first.

71% action and visual is not a weakness in this script. It is the method. The pilot earns its dialogue by making you wait for it.

What Comes Next

My tool is still in development, but the analysis above is exactly what I want it to surface for every script that moves through it. Not in a prescriptive way, not as a checklist that replaces creative instinct, but as a shared language that writers, producers, and executives can use to talk about what a script is actually doing at a structural level. I am sorry if this felt kind of check-listy!

August is the target. In the meantime, I plan to keep running scripts through it and writing up what I find here. Next up: I am curious what the pilot of The Bear looks like through this lens, because I have a theory about why that show retained its audience in a way that almost nothing else on streaming has managed to do.

More soon.

As of writing this, I’ve just closed four sections of data storytelling with 171 MBA students…3,000 slides+, and 1,500+ charts later, I’m wiped out! I apologize for the delay here, and will be making this up in the next month with a few more articles before I head out for a data science/AI summit in Italy! Wahoo!

On media front, it’s been interesting to see that Reed Hastings’ departure is driving down Netflix stock. Of course, this comes at a time when subscriber growth is decelerating (Netflix Stock Is Down | The Motley Fool), but I think a valuation concern as the streaming space is no longer a growth market but, perhaps, a mature value one. It didn’t help that they didn’t acquire WBD, but perhaps this ill-fated engagement raised more questions than answers among investors. Yes, a $2.8 billion break up fee isn’t a bad deal (Paramount Paid Netflix $2.8 Billion Breakup Fee for Warner Bros.) (I wish I got that after my last break up), but more importantly I think it illuminates potential insecurities Netflix may have with which direction to grow. But don’t worry, Netflix is creating Netflix clips, reconfiguring its mobile app to help users discover new content. Like a best friend late to a party, some of us may be saying it’s about time…while others might be thinking “why bother at this point?”

This week, I wanted to examine the TV series Black-ish that ran on ABC from 2014-2022. At the onset, it was saddled with unfair comparisons to Cosby, Sanford and Son, and other African-American driven TV series that promulgated equality, fairness, and success, the TV series inherited a legacy as much as it attempted to forge and further one. It should be stated that The Cosby Show, for all of its cultural weight and impact, was repeatedly criticized for not addressing systemic racial inequality. Black-ish never side-stepped this issue, and instead elected to repeatedly step in it. Specifically, how the family unit should trade off particular aspects of being a middle-class Black family for external acceptance in groups.

Minority-driven TV series often tasked themselves with challenging institutions in modern society that, though endowed with inherited layered racism and inequalities, are often too timid to thoughtfully confront these gaps and discrepancies. Still, this TV series still managed to cultivate a loyal following, sustain genuine humor, and, most importantly, address racism and assimilation in a way that felt accessible rather than accusatory.

None of the following data exists in a vacuum. Black-ish launched in 2014, the second year of Barack Obama's second term, who is the first Black presidency in American history, and a cultural moment that made the Johnson family's premise feel both timely and slightly fraught. The show debuted into a country still metabolizing the Great Recession, which had been particularly brutal for Black homeowners: Black household net worth, concentrated disproportionately in home equity, had been gutted by the housing collapse. And yet the period from 2008 to 2017 also saw a measurable expansion of the Black middle class as a share of the overall population, as the percentage of the Black population in the middle class increased across the country, driven by gains in education, white-collar employment, and suburban migration toward prosperous metros like Atlanta, Dallas, and Houston. The Black family that Black-ish depicted — aspirational, upwardly mobile, culturally conflicted — was not a fantasy. It was, for a growing segment of the audience, a mirror. That sociological backdrop helps explain both the show's early commercial traction and the specific texture of its assimilation anxiety: the Johnsons were not asking whether they could make it. They were asking what making it cost them. By the time the show reached its decline years, that question had grown considerably more complicated — the country had, too.

What I Measured — and How Well It Worked

The study analyzed nearly every episode of Black-ish by breaking each script into its three acts and measuring dialogue features at each stage, including things like analytical language density, emotional register, and how much the characters said. The goal was to see which of those features, in which act, predicted how many people watched the following week’s episode.

Think of it as asking: if you could read the script before the episode aired, what would tell you whether viewers were going to come back?

36.15% Variance explained (Adjusted R²) — in-sample model fit

32.04% Out-of-sample fit tested on unseen episodes

CV folds 0.33/ 0.22 / 0.33 / 0.38 / 0.33

166 of 176 Episodes analyzed — near-complete series corpus

When we tested the model on episodes it had never seen before, it explained about 36.15% of the variation in viewership from one week to the next, from dialogue features alone, with no information about scheduling, marketing, or what was on the other channel. That’s a meaningful signal from something as seemingly mechanical as word choice. While the dependent variable is next-episode viewership in millions, I don’t believe the coefficients that you are about to see impact millions of viewers directly, as there are far too many other forces not captured in the model. But directionally, I believe that my database helps decode why particular TV series work based on character dialogue.

Six features turned out to be statistically significant predictors. Here’s what they were, what they mean in plain English, and which episodes make the pattern obvious.

What Helped: The Features That Brought Viewers Back

1. Auditory language in the opening act (Hear_1)

The strongest positive predictor in the model was the presence of auditory perception language in Act 1 — words associated with hearing, listening, sound. The coefficient: +2.24. Remember, though this is in millions for the next episode, these coefficients should be treated as directional (e.g. positive or negative).

The finding is less about literal acoustics than about a specific mode of attention the opening act asks of its audience. Dialogue built around auditory perception tends to be scene-setting in a particular way: it situates characters in relation to each other, cues the audience to listen closely, and signals that something worth paying attention to is about to happen. In a show whose central dramatic device is Dre processing the world around him — narrating, reacting, translating — this register has an obvious structural function. It tells the audience, before the argument has been made, that there is an argument incoming.

The implication is not cosmetic. Act 1 that opens in an auditory register is an Act 1 that has already begun pulling focus.

Episode in focus · Season 4, Episode 7 · “North Star” · Alignment score: 0.5325

The episode opens in the aftermath of a dispute — Dre and Bow negotiating holiday custody between families, each side arriving with its own demands and its own cuisine. The setup is inherently conversational: multiple voices, competing claims, the comedy of two extended families discovering each other across a dining room table. The auditory dimension of the Act 1 dialogue is structural here, not ornamental. Before the ideological confrontation between soul food and beet dishes, before Grandma Mabel’s closing argument about what it means to make a plate for your husband, the episode’s opening act is doing the work of listening — staging the characters so that what they are about to say to each other carries weight. The Hear_1 coefficient of +2.24 captures something the episode’s architecture demonstrates: audiences come back to episodes that open by making them lean in.

2. Disgust language in the opening act (Disgust_1)

The second positive predictor is the one most likely to be misread. Episodes where Act 1 dialogue scored higher on disgust — the presence of language associated with moral violation, revulsion, or the transgression of social norms — predicted stronger viewership retention. The coefficient: +1.56.

The instinct is to read this as a penalty on comfort. The more precise reading is that it rewards provocation with an argument. Black-ish is a show built around moments of social discomfort that the Johnson household then has to navigate, metabolize, and resolve. When Act 1 introduces that discomfort through language that signals something has been violated — a norm, an expectation, a family boundary — it is not alienating the audience. It is presenting the audience with the very thing the show’s genre contract promises to address. Disgust in the opening act is not a flaw. It is the hook.

The condition is that the discomfort has to land as meaningful, not merely provocative. The difference is whether the episode has somewhere to take it.

Episode in focus · Season 2, Episode 1 · “The Word” · Alignment score: 0.5106

Jack says the n-word at his school talent show, quoting Kanye West’s “Gold Digger,” and the episode immediately distributes the disgust response across every character in the room — Bow finding the word hateful, Dre finding the situation complicated, Pops landing somewhere in between. What the opening act does precisely is establish that no position in the room is simple. The disgust register of Act 1 is not designed to resolve the discomfort. It is designed to make the audience feel why the question matters before the episode begins arguing through it. The Disgust_1 coefficient of +1.56 captures what “The Word” demonstrates in its first act: that moral discomfort, deployed as an opening move rather than a resolution, is the condition for the kind of episode audiences return for.

What Hurt: The Features That Drove Viewers Away

3. Reward language — in the act that should be escalating (Reward_1)

The most quietly damaging predictor in the model: episodes where Act 1 dialogue scored higher on reward language — words associated with receiving benefits, positive reinforcement, or resolution — predicted fewer viewers the following week. The coefficient: −1.26.

The logic is structural. An opening act that satisfies before the problem has been established is an opening act that has spent the audience’s attention before the episode has earned it. Black-ish operates on dramatic irony — Dre’s narration frames every episode as a conflict before the family knows it is in one. When Act 1 resolves into warmth and reward before the tension has been introduced, the show abandons the mechanism that makes its premise work. The audience feels the floor disappear. They don’t always articulate it. They just don’t come back.

4. Verbose sentence structure — in the investigative act (WPS_2)

Episodes where Act 2 featured longer average sentence length — measured as words per sentence — predicted significantly fewer viewers the following week. The coefficient: −2.05.

The structural reading matters more than the surface one. Act 2 in black-ish is where the argument is developed — Dre explaining, debating, teaching, usually through a combination of direct address and family confrontation. Longer sentences in this act tend to indicate an episode that is walking the audience through its reasoning rather than letting the reasoning emerge through conflict. When the show tells you what it’s about rather than demonstrating what it’s about, the argument becomes a lecture. The audience for a lecture is smaller than the audience for a fight.

Episode in focus · Season 7, Episode 8 · “Compton Around the Christmas Tree” · Alignment score: 0.5112

Dre’s identity crisis after Junior calls him a “valley dad” sends the family to Compton for the holidays, and the episode’s second act carries the argumentative load of two competing arcs: Dre trying to reassert his street credentials and Bow teaching Jack and Diane the actual meaning of charity. The dual structure is ambitious, but it means Act 2 has more ground to explain than a single-thread episode can cover in the same runtime. Longer sentences accumulate as the episode tries to move both arguments forward simultaneously. The WPS_2 penalty of −2.05 registers what the episode’s architecture produces: a middle act that is working harder to explain itself than it is letting its characters discover themselves through action. An alignment score of 0.5112 places the episode in the high-alignment tier, but the sentence-length cost is visible in the register.

5. Overuse of common vocabulary — in Act 2 (Dic_2)

Episodes where Act 2 dialogue scored higher on dictionary density — a higher percentage of words drawn from common vocabulary — predicted fewer viewers the following week. The coefficient: −1.73.

Ordinary language is the register of ordinary stakes. When the middle act of a black-ish episode settles into plain, familiar vocabulary, it tends to signal that the argument has become routine — that what is happening on screen is the kind of conflict the show has done before, in the kind of language any show would use to do it. The episodes that retained audiences were the ones whose Act 2 dialogue felt specific: to the Johnson family, to the particular conflict, to the cultural argument being made. Generic language in the argumentative core is a symptom of generic argument.

6. Psychological engagement language — in the argumentative core

Episodes where Act 2 dialogue scored higher on psychological engagement — dense use of personal pronouns, emotional vocabulary, and cognitive processing language — predicted fewer viewers the following week. The coefficient: −1.49.

The finding runs against the instinct of ambitious television writing. High engagement scores in Act 2 tend to mark episodes that are doing serious emotional work: scenes that ask the audience to track multiple characters’ interiority simultaneously, to hold personal stakes alongside the central argument, to be both moved and oriented at the same time. The problem is sequencing. Black-ish is a show whose Act 2 is designed to develop an argument through confrontation — not to process the emotional fallout of one. When the middle act tilts toward interior reflection, personal pronoun density, and the language of feeling rather than the language of conflict, the show slips out of its own genre register. The audience came for the debate. What they’re getting instead is the characters’ feelings about the debate. That’s a different show, and audiences notice the substitution even when they can’t name it.

One More Wrinkle: What the Sitcom Genre Says

The analysis above looks at black-ish from the inside. But there is a second layer worth considering: how does the show sit within the broader sitcom genre?

Two observations from the genre-level data are worth holding alongside the episode-level findings.

Provocation requires payoff. The Disgust_1 finding — that moral discomfort in Act 1 predicts stronger retention — is not a license for provocation without resolution. Cross-genre data from sitcoms broadly suggests that discomfort in the opening act predicts retention only when the episode’s argument structure provides somewhere for that discomfort to land. Black-ish at its best did this consistently: “The Word,” “North Star,” and similar episodes introduced the violation in Act 1 and spent Acts 2 and 3 metabolizing it. When the show introduced discomfort without the argumentative architecture to resolve it, the audience felt the imbalance.

Concision is a register, not a scene-level fix. The WPS_2 and Dic_2 penalties both point toward the same structural problem: Act 2 that over-explains. Like the word count findings in CSI, these are not problems you fix by tightening one scene. They reflect a room-level understanding of what the show’s argumentative middle act is supposed to do. Black-ish at its most economical let its characters argue without narrating the argument. When the show began explaining what its characters were feeling instead of letting them feel it in front of the audience, the Act 2 register softened. The ratings reflected the softening.

So What Would You Actually Change? The Writer’s Takeaway

The model’s implications are not a mandate to write more provocative television. They are more specific — and more useful.

Positive relationships — consider emphasizing

+2.24 · Auditory perception in the opening act Make the audience lean in before the argument begins. Dialogue that situates characters in relation to each other through sound and attention cues the audience that something worth hearing is coming. → If your Act 1 feels like setup, ask whether it’s asking the audience to listen yet.

+1.56 · Moral discomfort in the opening act The genre contract of black-ish is built on introducing friction the family has to navigate. Act 1 discomfort is not a risk — it is the premise. Deploy it with the confidence that the rest of the episode will earn it. → If your cold open resolves into warmth before the problem has arrived, you’ve spent the audience’s attention too early.

Negative relationships — use with caution

−1.26 · Reward language in Act 1 Don’t satisfy before you’ve provoked. An opening act that signals resolution before conflict removes the reason to keep watching. → If Act 1 feels warm and resolved, the audience has no reason to stay for Act 2.

−2.05 · Sentence length in Act 2 The argumentative core of the show should escalate, not explain. Longer sentences in Act 2 tend to mean the episode is walking the audience through its reasoning instead of letting them experience the argument. → Cut the exposition. Let the fight do the work.

−1.73 · Dictionary density in Act 2 Generic vocabulary signals generic stakes. If the middle act could belong to any show, it doesn’t belong to this one. → The Johnson household has a specific voice. Act 2 should sound like them, not like television.

Technical Notes & Full Output

Appendix A · Glossary of All Six Measures

Each feature was extracted using LIWC (Linguistic Inquiry and Word Count) and related computational text analysis tools, applied at the act level. The suffix _1, _2, or _3 denotes the act in which the feature was measured.

Significant positive predictors — more of this, more viewers next week

Hear_1 · Auditory perception language, Act 1 LIWC Hear category in the opening act: proportion of words associated with listening, sound, and auditory experience. Captures a dialogue register that orients characters — and the audience — toward active attention before the episode’s central argument has been made. sig. positive · coef +2.24

Disgust_1 · Disgust language, Act 1 LIWC Disgust category in the opening act: proportion of words associated with moral violation, revulsion, or transgression of social norms. Captures the introduction of friction that the episode’s argument structure is then asked to metabolize. sig. positive · coef +1.56

Significant negative predictors — more of this, fewer viewers next week

Reward_1 · Reward language, Act 1 LIWC Reward category in the opening act: proportion of words associated with receiving benefits, positive reinforcement, or resolution. High scores indicate an opening act that signals satisfaction before conflict has been established. sig. negative · coef −1.26

WPS_2 · Words per sentence, Act 2 Average sentence length in the argumentative core. Higher scores indicate dialogue that walks the audience through reasoning rather than letting conflict do the work. sig. negative · coef −2.05

Dic_2 · Dictionary density, Act 2 Percentage of words in Act 2 drawn from common vocabulary. High scores indicate a generic register in the middle act — language that could belong to any show rather than this one. sig. negative · coef −1.73

Engaged_2 · Psychological engagement, Act 2 Composite LIWC measure in the argumentative core: elevated use of personal pronouns, emotional vocabulary, and cognitive processing language. Captures an Act 2 oriented toward interior reflection rather than outward confrontation. sig. negative · coef −1.49

Appendix B · Most Aligned Episodes

The 10 Most Aligned Episodes: What the Data’s Best-Matching Cases Actually Show

The alignment score measures something specific, and it is worth being precise about what that is: how closely any individual episode’s writing matched the full feature pattern that the show-level Random Forest identified as predictive of retention. A high alignment score does not mean an episode performed well. It means the writing was operating inside the show’s structural fingerprint — exhibiting the features the RF found important, in proportions consistent with what the model learned from the complete corpus.

With that definition in place, the ten most aligned episodes in the 334-episode dataset are identifiable from the data:

The first thing the table shows is what seasons the most aligned episodes came from: six of the top ten are from Seasons 6 and 7 — the decline era. The three highest alignment scores in the series belong to a Season 4 mid-season episode, a Season 7 episode, and a Season 2 episode, in that order. That distribution is worth pausing on. The writing was more precisely calibrated to the retention pattern during the show’s audience contraction than at most points during its peak years — with one notable exception at the top.

The second thing the table shows is what that precision was worth. “North Star,” the most aligned episode in the series by this measure, drew 8.58 million viewers — and it aired in Season 4, not the decline era. The Season 6 and 7 episodes clustered in the top ten drew between 2.39 and 3.17 million. Alignment is not performance. It is a measure of internal consistency with the writing pattern. The ceiling it describes moved independently of the quality it captured, and the viewership gap between “North Star” and the late-season episodes in this table illustrates that gap precisely.

The third thing the table invites — and the most important interpretive caution — is the question of why late-season episodes align so highly in the first place. The intuitive answer is that the writers had, over time, refined the formula. The data suggests a different explanation, or rather two of them.

The first is a training artifact. The show-level Random Forest was fit on all episodes across the full run, which means the later seasons constitute a meaningful share of the training data. The RF’s learned retention pattern is not solely a portrait of what peak-era Black-ish looked like — it is partly a portrait of what the contracting show looked like, because that is what it was substantially trained on. When late-season episodes score high on alignment, part of what is being measured is that the RF recognizes them as structurally typical of the era it largely learned from. It is not that the writing converged on the peak pattern. It is that the peak pattern and the decline pattern were both inside the RF’s training set, and the alignment score reflects closeness to the RF’s full learned distribution — including its late-season component.

The second explanation is variance. Peak-era Black-ish was doing more things. The writing room was taking risks, the ensemble cast was generating less predictable dynamics, and the show had the commercial security to experiment. That variability spreads peak-era episodes further from the RF’s central tendency — producing lower alignment scores, not because the writing was weaker, but because it was more dispersed. By Seasons 6 and 7, the show had contracted into a narrower creative range. Less experimentation means more episodes clustering around similar feature values, which the RF reads as high alignment regardless of whether those values are pointing in a productive direction. Consistency is not the same as correctness.

Generalizability: Findings are specific to Black-ish and should not be applied directly to other series without replication. Cross-genre and cross-network analysis is in progress. Preliminary results suggest certain dialogue-density penalties generalize across ensemble comedies; the cultural specificity signal appears genre- and network-conditional.

There’s a particular kind of television criticism that mistakes economy for emptiness. CSI: Crime Scene Investigation attracted plenty of it during its fifteen-season run on CBS (2000–2015). Reviewers catalogued the show’s procedural formula, its forensic pseudoscience, its characters defined more by their equipment than by inner lives. The Atlantic once dismissed the entire franchise as “a machine for producing closure.”

And yet at its peak the show drew 26 million viewers. For years it ranked as the most-watched drama on American television — and, by some international syndication measures, the most-watched series on the planet. It ran for 337 episodes. People kept coming back. So the question worth asking isn’t whether the critics were right about the formula. It’s: what was the show doing, at the level of the actual words, that kept the audience showing up?

To find out, I built a statistical model using dialogue features extracted from every episode’s script — measuring things like how analytically precise the characters sounded, how much they said in each act, and how emotionally charged the language got — then asked which of those features predicted whether viewers returned the following week. What came back was counterintuitive. And it reframes the show’s legacy entirely.

“The critics weren’t wrong that CSI was cold. But they were watching the coldness. The audience was watching the thinking.”

What I Measured — and How Well It Worked

The study analyzed every episode of CSI: Crime Scene Investigation by breaking each script into its three acts and measuring dialogue features at each stage — things like analytical language density, emotional register, and how much the characters said. The goal was to see which of those features, in which act, predicted how many people watched the following week’s episode.

Think of it as asking: if you could read the script before the episode aired, what would tell you whether viewers were going to come back?

38.65% Variance explained — in-sample model fit

34.57% Out-of-sample fit tested on unseen episodes

CV folds 0.29 / 0.39 / 0.45 / 0.14 / 0.46

334 of 337 Episodes analyzed — near-complete series corpus

When we tested the model on episodes it had never seen before, it explained about 34.57% of the variation in viewership from one week to the next — from dialogue features alone, with no information about scheduling, marketing, or what was on the other channel. That’s a meaningful signal from something as seemingly mechanical as word choice. While the dependent variable is next-episode viewership in millions, I don’t believe the coefficients that you are about to see impact millions of viewers directly — there are far too many other forces not captured in the model. But directionally, I believe that my database helps decode why particular TV series work based on character dialogue.

Seven features turned out to be statistically significant predictors. Here’s what they were, what they mean in plain English, and which episodes make the pattern obvious.

What Helped: The Features That Brought Viewers Back

1. The emotional range of the middle act

The strongest positive predictor in the model was not a specific emotion. It was the variation between emotions. Episodes where the Act 2 emotional register shifted across scenes — not holding a single affective pitch for twenty minutes, which predicted significantly higher viewership the following week. The coefficient: +12.03.

To understand why this matters, you have to think about what Act 2 actually does in a procedural. It’s the investigative core. Interviews are conducted, evidence catalogued, hypotheses assembled and discarded. The mechanical risk of the procedural form is that this act becomes a flat, professional sequence, as competent people doing competent things in a competent register. What the data suggests is that the episodes audiences returned for were the ones that treated Act 2 as emotionally dynamic rather than emotionally inert. Not a second act defined by grief, or tension, or levity, but one that moved between them.

The implication is structural. A moment of dry humor landing inside an otherwise tense interview. A beat of unexpected stillness before evidence breaks open a case. A small human exchange dropped into an otherwise affectless evidence review. These aren’t flourishes. The model suggests they’re load-bearing and the variation itself is the hook that keeps an audience present.

Episode in focus · Season 13, Episode 10 · "Risky Business Class"

Alignment score: 0.806

A private plane crashes just off the Vegas strip with no manifest and no immediate identification of victims. The investigation's second act is structurally unusual for the series: evidence gathering happens across multiple survivor interviews in a hospital setting, intercut with lab work on wreckage, intercut with Sara navigating the emotional weight of a former relationship tied to the crash. The tonal range of the act is accordingly wide, clinical observation alongside interpersonal tension alongside the dark procedural logic of reconstructing a crash. An alignment score of 0.806 places this episode in the top five most pattern-consistent in the series. The SD_scaled signal feeling transported into the TV episode in Act 2, is visible in the episode's architecture. What the coefficient of +12.03 captures is what the episode demonstrates: not a second act holding a single emotional pitch, but one that modulates it across the full investigative sequence.

2. Neutral tone as structural breathing room

The second positive predictor runs against the instinct of most television writers. Episodes where Act 2 dialogue scored higher on neutral tone — measured as the relative absence of emotionally loaded language — predicted stronger viewership retention. The coefficient: +8.59.

The immediate reading is that this rewards flatness. The more accurate reading is that it rewards architecture. Neutral language in the investigative act is not a failure of dramatic ambition — it is the structural equivalent of negative space. It gives an audience time to process what they’ve been shown before the third act begins collapsing the possibilities into a resolution. A second act running at sustained emotional intensity doesn’t build to the finale; it pre-spends it. By the time the answer arrives, the audience has already been holding its breath for twenty minutes. The payoff lands softer, not harder.

The neutral passages aren’t dead air. They are the silences that make the resolution land.

Episode in focus · Season 12, Episode 20 · “Altered Stakes” · Alignment score: 0.8019

A convicted murderer's lawyer files for a new trial, arguing that a detective coerced the confession. The CSI team has a narrow window to find independent evidence to keep him imprisoned, working through archived physical materials with one hand effectively tied by evidentiary rules. The second act is procedural in the most classical sense: Nick and the team moving through old evidence, running tests processed years earlier, building a case through incremental inference rather than confrontation. There is no emotional climax in the investigative act. The affect is clean, measured, and deliberate. An alignment score of 0.8019 places this among the three highest in the series. The episode demonstrates what the Neutral_2 coefficient of +8.59 rewards: not a second act that withholds drama, but one that uses the absence of emotional urgency to make the third act's resolution feel earned by accumulation rather than forced by pressure.

3. Analytical language — in the final act especially

This was the finding that most directly challenges the critical consensus about CSI. Episodes where Act 3 featured denser analytical language — formal, logical, structured reasoning — predicted significantly more viewers the following week. The coefficient: +8.88.

The familiar complaint about CSI was that its forensic science was stylized to the point of fantasy. That critique, directed at the accuracy of the analytical sequences, is not wrong. But audiences were not returning for accuracy. They were returning for the experience of watching someone reason through a problem and arrive, with apparent inevitability, at a solution. In a procedural, thinking is not the path to the emotional payoff — it is the emotional payoff. The analytical third act is where the show makes good on its implicit promise: you stayed with us through the evidence, now here is the answer, and here is exactly why it is correct.

When CSI replaced that register with music swells and anguished close-ups, it was substituting affect for the genre’s actual currency.

Episode in focus · Season 14, Episode 12 · “Keep Calm and Carry-On” · Alignment score: 0.8069 — the highest in the series

A petty theft on an inbound flight becomes a murder investigation when one of the passengers is found dead after landing. The case is structurally clean: a closed environment, a finite suspect pool, a sequence of physical evidence pointing logically toward a single resolution. The third act is a textbook example of what the Analytic_3 coefficient rewards — Russell and the team working through the evidence chain step by step, each inference following from the last, the resolution arrived at through structured deduction rather than dramatic confrontation. No emotional epiphany. No music cue signaling that the moment has arrived. The answer is delivered because the logic is complete.

“Keep Calm and Carry-On” holds the highest alignment score of any episode in the 334-episode corpus: 0.8069. It was written in Season 14, when the show was drawing 10.3 million viewers — less than half its peak. The writing was, by the model’s measure, more precisely calibrated to the retention pattern than anything Grissom’s team produced in the show’s prime years. The ceiling had simply moved.

What Hurt: The Features That Drove Viewers Away

4. Fear language in the wrong act

The most counterintuitive finding in the model: for a crime drama built around threat, violence, and death, fear language in Act 2 was negatively associated with next-week viewership. The coefficient: −4.74.

On reflection, the logic is precise. CSI is not a horror show. Its genre contract with the audience is not sustained dread but the resolution of dread. The pleasure of the procedural form is forensic — intellectual, not visceral. When the investigative act sits in fear rather than moving through it toward analysis, the show defaults to a different genre promise than the one its audience signed up for. It amplifies the threat without delivering the satisfaction of dismantling it. The audience that came for the answer has been given more of the problem instead.

Fear in Act 2 is not inherently wrong. Fear in Act 2 that Act 3 fails to analytically resolve is a broken contract.

The observation holds at the scene level too: when Grissom delivers a verdict in three cold sentences, it lands harder than any scene designed to make you feel how frightening the crime was.

Episode in focus · Season 13, Episode 9 · “Strip Maul” · Alignment score: 0.7641

The Texture of Dread

The investigative case involves a murder with connections to Vegas's underground sex trade. But the episode's structural weight sits on a separate axis: Brass's troubled daughter Ellie resurfaces, and the second act carries a persistent undertone of dread and personal anxiety around her storyline that runs alongside the procedural investigation. The fear language is not primarily about the crime. It is about Brass and what Ellie represents, danger, failure, loss of control. Act 2 oscillates between the neutral register of the forensic investigation and the emotionally loaded register of the Brass subplot, with fear-adjacent vocabulary accumulating in the latter. An alignment score of 0.7641 places the episode well within the high-alignment tier, and the procedural backbone is correctly calibrated. But the Fear_2 signal carries a cost that the OLS model registers even inside a well-structured episode. The genre contract, dread introduced, dread analytically resolved, is partially broken by a subplot the show cannot resolve in the forensic register.

5. Too many words — across every act

This finding is less dramatic than the word count penalty but operates on related logic. Episodes where Act 1 dialogue scored higher on psychological engagement — dense use of personal pronouns, emotional vocabulary, and cognitive processing language — predicted fewer viewers the following week. The coefficient: −8.56.

The Engaged score is not a measure of bad writing. It often marks ambitious writing: scenes that ask the audience to track multiple characters’ interiority simultaneously, to hold emotional information alongside investigative setup, to be both moved and oriented before the case has given them anything to anchor to. The problem is attentional sequencing. The procedural form has a natural order — establish the puzzle, then invest the audience in the people solving it — and Act 1 scenes that front-load psychological engagement reverse that sequence. Emotional stakes requested before procedural stakes are established tend to feel effortful rather than involving. The audience notices. They don’t always come back.

The practical implication is specific: in Act 1, let the case arrive first. The personal dimensions of the characters are more effective once the audience has been given a reason to care about the outcome. Personal pronouns and reflective interiority belong downstream in the narrative — in Act 3, when the resolution can absorb and give meaning to what the characters are feeling.

Episode in focus · Season 14, Episode 8 · “Helpless” · Alignment score: 0.7364

The episode runs two cases simultaneously: the death of a man found inside a human-sized hamster ball, and a serial rape case that Morgan is working alongside a swing-shift CSI who subsequently goes missing. The dual structure means the opening act carries significant psychological weight before either case is fully established, the serial assault investigation introduces victim trauma and personal stakes into Act 1 before the procedural hook has been set, and Morgan's relationship with the swing-shift CSI adds an interpersonal layer that asks for emotional investment ahead of narrative context. The Engaged_1↓ signal that appears in Season 14's RF importance rankings is visible here: the episode's opening act is psychologically dense in ways that ask more of the audience than the case, at that point, has earned. An alignment score of 0.7364 indicates the episode was otherwise well-calibrated to the show's retention pattern. The Act 1 engagement penalty is a local cost inside a correctly structured episode, but it is a cost the model registers regardless.

One More Wrinkle: What the Procedural Genre Says

The analysis above looks at CSI from the inside — what made one episode outperform another within the same show. But there’s a second layer worth considering: how does CSI sit within the procedural drama genre as a whole?

Two findings from the genre-level data are worth highlighting for a general audience:

Procedurals reward resolution, not exposure. The fear penalty in CSI is not an isolated finding. Preliminary cross-genre analysis suggests that in procedural formats broadly, emotional language in the middle act is negatively associated with retention when it amplifies threat rather than resolving it. Audiences come to procedurals for the satisfaction of a closed case. Fear that is not analytically dismantled is a genre promise broken.

Economy of language is a voice, not a choice. Just as with 2 Broke Girls — where sentence length was a register set at the room level, not a scene-by-scene decision — CSI’s word count patterns are consistent within seasons. The wordy seasons were wordy throughout. You cannot fix a verbose third act by tightening one scene. The brevity has to be structural, built into the writers’ room’s baseline understanding of what the show’s dialogue is supposed to do.

CSI, airing 2000–2015 and built on spare, logical exchanges and economical forensic language, was in several respects the procedural that best honored those norms during its peak years. The 20 to 26 million viewers it held through its first eight seasons responded to that register — not despite its coldness, but because of it.

This was the largest and most consistent finding in the model. It was not close, and it was not localized. Episodes where characters spoke more words in Act 1 predicted significantly fewer viewers the following week: coefficient −11.75. Act 2 followed at −6.44. Act 3 at −6.89. The penalty ran across all three acts without exception.

This is not a pacing problem in individual scenes. It is a register problem for the entire episode. And read against what CSI actually is — a show whose signature moments involve a character standing over a piece of evidence in silence before delivering a single, precise observation — the finding is almost obvious in retrospect. The show’s voice is economical. Its characters speak in conclusions, not in process. They don’t narrate their reasoning; they present its output. The episodes that forgot this — that explained what was about to happen, showed it, then explained what had just happened — were the ones the audience drifted from.

Viewers can feel when dialogue is filling time rather than advancing thought. The later seasons of CSI were simply wordier than the early ones, consistently and across all three acts. The ratings reflected it.

Episode in focus · Season 12, Episode 22 · “Homecoming” · Alignment score: 0.7353

The Season 12 finale investigates a triple homicide that exposes corruption running through the LVPD, a structurally ambitious case that requires the episode to carry significant expository load across all three acts. The scope of the conspiracy means characters spend extended sequences establishing connections, laying out chains of implication, and walking through institutional context that the audience needs to track the resolution. An alignment score of 0.7353 places the episode in the high-alignment tier. But the word count penalty is structural: a finale-scale episode resolving multiple threads across a corrupt-institution storyline has more ground to cover than a single-case procedural, and the dialogue reflects it. Word counts rise across all three acts. The model registers this as a retention cost even inside an otherwise correctly structured episode. And across the later seasons, that local cost was consistent enough to register as a systemic one.

So What Would You Actually Change? The Writer’s Takeaway

The model’s implications aren’t a mandate to write colder television. They’re more precise than that — and more useful.

Positive relationships — consider emphasizing

+6.88 · Emotional range in the middle act

Vary the register. Don’t let the investigative core hold a single emotional pitch for twenty minutes. A moment of dark humor, a beat of unexpected stillness — the contrast is the hook.

→ If your Act 2 sounds the same from scene to scene, the audience has already checked their phone.

+6.66 · Neutral tone as structural pacing

Build in breathing room. Neutral passages in the middle act are not dead air — they’re the silences that make the resolution land harder. Pre-spend the audience’s emotional attention and the third act has nothing to pay off.

→ The pause before the answer is part of the answer.

+6.62 · Analytical precision in the final act

Let the climax be intellectual. In a procedural, the thinking is the feeling. The answer delivered in precise, logical language is the emotional payoff — not a substitute for it.

→ If your third act is reaching for music and close-ups, consider whether a quieter, more analytical close would honor the genre better.

Negative relationships — use with caution

−3.89 · Fear language in the investigative act

Move through threat, don’t sit in it. The middle act should be resolving dread, not amplifying it. Audiences who came for the answer do not stay for the dread.

→ Fear in Act 2 is a genre promise. If Act 3 doesn’t redeem it analytically, the promise is broken.

−11.75 / −6.44 / −6.89 · Word count, all three acts

The penalty is consistent across every act. This is not a scene-level fix. It is a room-level instruction: CSI’s voice is economical, and when the writing forgets that, the audience notices. Cut the setup. Cut the explanation. Cut the monologue and leave the observation.

→ If a character is saying what they’re about to do instead of doing it, cut the first sentence.

Technical Notes & Full Output

Appendix A · Glossary of All Seven Measures

Each feature was extracted using LIWC (Linguistic Inquiry and Word Count) and related computational text analysis tools, applied at the act level. The suffix _1, _2, or _3 denotes the act in which the feature was measured.

Significant positive predictors — more of this, more viewers next week

SD_scaled_2 · Emotional range, Act 2 Standard deviation of scaled emotion scores across Act 2. Captures tonal variation in the investigative core — how much the emotional register shifts across scenes rather than holding a single pitch. sig. positive

Neutral_2 · Neutral tone, Act 2 LIWC neutrality score in the middle act: relative absence of positively or negatively valenced emotional language. Captures pacing and structural breathing room between Act 1 setup and Act 3 resolution. sig. positive

Analytic_3 · Analytical language, Act 3 LIWC Analytic score in the final act: measures logical, formal, categorical thinking. High scores indicate characters reasoning in structured, evidence-based language rather than affective or personal terms. Range 0–100. sig. positive

Significant negative predictors — more of this, fewer viewers next week

Fear_2 · Fear language, Act 2 LIWC Fear category in the middle act: proportion of words associated with anxiety, dread, threat, and danger. Captures sustained affective exposure rather than investigative resolution. sig. negative

WC_1 · Word count, Act 1 Total words spoken in the opening act. The largest single coefficient in the model — negative — measuring dialogue density in the setup and cold open. sig. negative · coef −11.75

WC_2 · Word count, Act 2 Total words spoken in the investigative act. sig. negative · coef −6.44

WC_3 · Word count, Act 3 Total words spoken in the climactic act. sig. negative · coef −6.89

Appendix B · Most Aligned Episodes

The 10 Most Aligned Episodes: What the Data’s Best-Matching Cases Actually Show

The alignment score measures something specific, and it is worth being precise about what that is: how closely any individual episode’s writing matched the full feature pattern that the show-level Random Forest identified as predictive of retention. A high alignment score does not mean an episode performed well. It means the writing was operating inside the show’s structural fingerprint — exhibiting the features the RF found important, in proportions consistent with what the model learned from the complete corpus.

With that definition in place, the ten most aligned episodes in the 334-episode dataset are identifiable from the data:

The first thing the table shows is what season the most aligned episodes came from: seven of the top ten are from Seasons 12 through 15 — the decline era. The three highest alignment scores in the series belong to a Season 14 mid-season episode, a Season 13 episode, and a Season 12 episode, in that order. The writing was more precisely calibrated to the retention pattern during the show’s audience collapse than at any point during its peak years.

The second thing the table shows is what that precision was worth. “Keep Calm and Carry-On,” the most aligned episode in the series by this measure, drew 10.3 million viewers. Unaligned episodes from Season 5 drew nearly three times that. Alignment is not performance. It is a measure of internal consistency with the writing pattern. The ceiling it describes moved independently of the quality it captured.

The third thing the table invites — and the most important interpretive caution — is the question of why late-season episodes align so highly in the first place. The intuitive answer is that the writers had, over time, figured out the formula. The data suggests a different explanation, or rather two of them.

The first is a training artifact. The show-level Random Forest was fit on all 334 episodes, which means Seasons 10–15 constitute roughly 40% of the training data. The RF’s learned retention pattern is not solely a portrait of what peak-era CSI looked like — it is partly a portrait of what the declining show looked like, because that is what it was predominantly trained on. When late-season episodes score high on alignment, part of what is being measured is that the RF recognizes them as structurally typical of the era it largely learned from. It is not that the writing converged on the peak pattern. It is that the peak pattern and the decline pattern were both inside the RF’s training set, and the alignment score reflects closeness to the RF’s full learned distribution — including its late-season component.

The second explanation is variance. Peak-era CSI was doing more things. The writing room was experimenting, the cast was generating unpredictable dynamics, and the show had the commercial security to take creative risks. That variability spreads peak-era episodes further from the RF’s central tendency — which produces lower alignment scores, not because the writing was worse, but because it was more dispersed. By Seasons 12–15, the show had contracted into a narrower creative range. Less experimentation means more episodes clustering around similar feature values, which the RF reads as high alignment regardless of whether those values are pointing in the right direction. Consistency is not the same as correctness. The ↓ flags are the evidence: if the late seasons had genuinely cracked the formula, the negative predictors would have disappeared from the season-level RF rankings. Instead, Engaged_1↓ appears in Season 14 and Fear_2↓ in Season 9. The convergence was real. It was just convergence toward a consistent pattern that included the costly features alongside the beneficial ones.

Generalizability: Findings are specific to CSI: Crime Scene Investigation and should not be applied directly to other shows without replication. Cross-genre analysis is in progress. Preliminary results suggest the word count penalty generalizes across procedurals; the analytical language signal appears genre-conditional.

There’s a particular kind of television criticism that mistakes simplicity for stupidity. 2 Broke Girls attracted plenty of it during its six-season run on CBS (2011–2017). Reviewers catalogued the show’s reliance on crude jokes, its thin plotting, its dialogue that moved at the speed of a vaudeville act. The AV Club eventually stopped reviewing it altogether.

And yet the series premiere drew 19.4 million viewers — the highest-rated fall comedy premiere since 2001. The show survived, got renewed, and ran for 138 episodes. People kept coming back. So the question worth asking isn’t whether the critics were right. It’s: what was the show doing, at the level of the actual words, that kept the audience showing up?

To find out, I built a statistical model using dialogue features extracted from every episode’s script — measuring things like how emotionally engaged the characters sounded, how long their sentences ran, and how complicated their vocabulary got — then asked which of those features predicted whether viewers returned the following week. What came back was surprising. And it reframes the show’s legacy entirely.

“The critics weren’t wrong that the jokes were often cheap. But they were watching the jokes. The audience was watching the friendship.”

What We Measured — and How Well It Worked

The study analyzed every episode of 2 Broke Girls by breaking each script into its three acts and measuring ten distinct dialogue features in each act — things like emotional engagement, vocabulary complexity, and sentence length. The goal was to see which of those features, in which act, predicted how many people watched the following week’s episode.

Think of it as asking: if you could read the script before the episode aired, what would tell you whether viewers were going to come back?

58%Variance explained in-sample model fit

42%Out-of-sample fit tested on unseen episodes

1.14MAvg. prediction error viewers, typical episode

When we tested the model on episodes it had never seen before, it explained about 42% of the variation in viewership from one week to the next — from dialogue features alone, with no information about scheduling, marketing, or what was on the other channel. That’s a meaningful signal from something as seemingly intangible as word choice. While the DV is viewership in millions for the next episode, I don’t believe the coefficients that you are about to see impact millions of viewers (there are SO many other things not measured in model). However, directionally, I believe that my database helps decode why particular TV series work based on character dialogue.

Five features turned out to be statistically reliable predictors. Here’s what they were, what they mean in plain English, and which episodes make the pattern obvious.

What Helped: The Features That Brought Viewers Back

1. “Should have, could have, would have” — the language of regret

Episodes where characters used more words like should, could, would, ought — the grammar of imagining things differently — predicted higher viewership the following week. In the second act specifically, this kind of language was associated with roughly 1.4 million additional viewers returning the next episode.

It makes intuitive sense when you think about what those words actually do in a scene. They’re not describing what happened. They’re describing what almost happened, what should have happened, what the character wishes were true. They signal an inner life. They signal stakes.

Episode in focus · Season 4 · “And the Childhood Not Included”

The Teddy Bear Scene

Caroline accidentally loses Han’s expensive fish. Max quietly offers to sell the only toy from her childhood — a teddy bear, a gift from her mother — to cover the cost. Caroline finds out and refuses. Neither character makes a speech. The whole scene runs on unspoken logic: what kind of person would I be if I let her do that, what would it say about us, what if I’d just been more careful. Nobody says “should” out loud. But the emotional grammar of the scene is entirely counterfactual. That’s exactly what the model is picking up.

2. Emotional honesty — in the first act and especially the last

This was the strongest finding in the entire study. Episodes where characters sounded emotionally present — using first-person language, expressing feelings directly, reflecting on what things meant to them — predicted significantly more viewers the following week. And the effect was strongest in the third act: the closing scenes of an episode, where the show either earns its emotional resolution or lets it go.

Episodes that scored high on emotional engagement in the final act were associated with nearly 1.8 million more viewers returning the following week — the biggest positive effect in the model.

Importantly, episodes that opened emotionally and closed emotionally did better than episodes that only did one. The data suggests the show’s best episodes weren’t just funny — they established something personal early and paid it off at the end. The warmth between Max and Caroline wasn’t decoration. It was the structure.

Episodes in focus · Season 2, Episodes 6–7 · The Candy Andy Arc

“And the Candy Manwich” / “And the Three Boys With Wood”

Caroline can’t face the cute candy store owner after an embarrassing encounter at his shop — she ate too much candy and got sick in front of him. The premise is pure sitcom. But both episodes are built around genuine vulnerability: Caroline’s humiliation is specific and personal, her courage to try again is earned rather than assumed, and the resolution in the follow-up episode — when Andy discovers who she really is and has to decide whether it matters — closes on something that feels true. These two episodes consistently rank among the highest-rated in the series. The jokes are there. The emotional honesty is what the audience came back for.

What Hurt: The Features That Drove Viewers Away

3. Big words in the wrong place

This was the most powerful finding — and the most counterintuitive. Episodes where the dialogue in the final act used more long, complex words were associated with significantly fewer viewers coming back. Not slightly fewer. The effect was large enough that it was the single strongest predictor in the entire model, positive or negative.

A one-unit increase in vocabulary complexity in the third act was associated with 3.35 million fewer viewers the following week — more than twice the effect of any positive predictor in the model.

Read against 2 Broke Girls specifically, this makes a kind of instinctive sense. The show’s voice is Max Black’s voice: blunt, direct, working-class specific. When that voice gets dressed up — when the climax reaches for elevated language instead of plain truth — something breaks. The audience can feel the effort. It doesn’t read as smarter. It reads as false.

“When Max says she’ll sell her only childhood toy for Caroline in three plain words, that lands harder than any speech with polysyllabic sincerity ever could.”

4. Long sentences at the start

When the opening act of an episode had longer, more complex sentence structures, the model predicted lower next-week viewership. The effect was consistent and statistically reliable.

The intuition here is pacing. A sitcom earns its audience’s attention in the first few minutes or it doesn’t earn it at all. Long sentences are a friction tax — they slow the rhythm before the episode has found its feet. The top-rated episodes of 2 Broke Girls, whether emotionally driven or purely comedic, tend to open at a sprint. Short sentences. Fast exchanges. The premise established before you’ve had time to reach for your phone.

Episode in focus · Season 3 · “And the Kitty Kitty Spank Spank”

Han’s Fake Girlfriend

Han has been emailing a stripper’s photo to his mother in Korea, claiming she’s his girlfriend. When his mother visits, he begs Max and Caroline to find the woman and convince her to play the part. The cold open establishes all of this in under two minutes, with exchanges running four words or fewer. It’s a clinic in how to set a premise at speed. The episode ranks among the top ten on IMDb — not because it’s the show’s most emotionally complex half-hour, but because it never gives you a reason to stop watching.

One More Wrinkle: What the Sitcom Genre Says

The analysis above looks at 2 Broke Girls from the inside — what made one episode outperform another within the same show. But there’s a second layer worth considering: how does 2 Broke Girls sit within the sitcom genre as a whole? A separate genre-level analysis, run across sitcoms broadly, adds important context.

Figure 3. Pairwise correlations among structural and dialogue features across the sitcom genre as a whole. See Appendix B for full technical commentary on these relationships.

Two findings from the genre-level data are worth highlighting for a general audience:

Newer sitcoms are wordier in their opening acts — and getting more so. Across the genre, more recently produced shows pack significantly more dialogue into Act 1 than shows from a decade ago. This is the prestige-television influence: as dramas raised the expectation for dense, information-rich scripts, even network comedies started front-loading more. Read against the 2 Broke Girls finding that wordier, longer-sentence openings hurt viewership, this creates a real industry tension. The genre has been drifting toward something the data suggests doesn’t work.

Sentence length is a voice, not a choice. Shows that write in short sentences in Act 1 write that way throughout. The correlation between sentence length in Acts 1 and 2 across sitcoms is strong enough to suggest this is a register set at the show level — it’s in the DNA of the writers’ room, not a scene-by-scene decision. You can’t fix a slow-opening episode by tightening the ending. The brevity has to be structural.

2 Broke Girls, airing 2011–2017 and built on vaudeville-speed exchanges and plain working-class language, was in several respects swimming against where the genre was heading. The 10 to 15 million viewers it held through most of its run responded to that register, not despite its simplicity, but because of it.

So What Would You Actually Change? The Writer’s Takeaway

The model’s implications aren’t a mandate to write simpler television. They’re more precise than that — and more useful.

Positive relationships — consider emphasizing

+1.79

Emotional honesty in the final act

Let your characters drop the armor in the closing scenes. Personal pronouns, direct feeling, specific stakes — this is what the audience takes home. The joke is the vehicle. The moment of genuine connection is the destination.

→ If your third act ends on a quip instead of a truth, you’re leaving viewers behind.

+1.50

Establish emotional stakes early

Audiences who feel something in Act 1 are primed to feel something in Act 3. The opening doesn’t have to be heavy — it just has to be personal. Give us a reason to care before you give us the premise.

→ Open with stakes that are about the relationship, not just the situation.

+1.38

Give characters something to regret in the middle act

Counterfactual language — “should,” “could,” “would” — signals characters with real inner lives who are processing the gap between what is and what might have been. That gap is where audience identification lives.

→ Unresolved longing is a more durable hook than unresolved plot.

Negative relationships — use with caution

−2.07

Long sentences in the cold open

If a line can be cut in half, cut it. Sitcom audiences calibrate their attention to the show’s rhythm in the first two minutes. Long sentences bleed pacing before the episode has found its feet.

→ If your cold open reads like a speech, it should read like an exchange.

−3.35

Elevated vocabulary in the climax

The most emotionally resonant moments in television are written in plain language. Fancy words in the third act don’t signal sophistication — they signal effort. Audiences feel the effort, and it creates distance.

→ Write your climax like you mean it, not like you went to school for it.

Coming next in this series

Does the pattern hold across other shows?

The 2 Broke Girls model is one data point in a much larger study. The same framework has been applied across dozens of series spanning network procedurals, prestige cable, and streaming originals. Preliminary results suggest the vocabulary complexity penalty is robust across genres — but the emotional engagement signal varies considerably. Drama series reward engagement earlier; comedies more often front-load plot and close with emotion. A full cross-series comparison is in progress.

Technical Notes & Full Output

Appendix A · Glossary of All Ten Measures

Each feature was extracted using LIWC (Linguistic Inquiry and Word Count) and related computational text analysis tools, applied at the act level. The suffix _1, _2, or _3 denotes the act in which the feature was measured.

Significant positive predictors — more of this, more viewers next week

Engaged_3

Emotional depth, Act 3

Composite score measuring first- and second-person pronouns, emotional vocabulary, and reflective thinking in the final act. Captures how personally invested characters sound in the closing scenes.

sig. positivep = 0.035

Engaged_1

Emotional depth, Act 1

Same composite engagement construct as Engaged_3, measured in the opening act. Elevated personal pronouns (I, we, you), affective words, and cognitive processing terms.

sig. positivep = 0.081

Discrep_2

Regret / “what if” language, Act 2

LIWC discrepancy category: proportion of words indicating counterfactual thinking — should, could, would, ought, must. Captures characters processing alternatives to their current situation.

sig. positivep = 0.098

Significant negative predictors — more of this, fewer viewers next week

Sixltr_3

Long-word density, Act 3

Percentage of words with six or more letters in the third act. A vocabulary complexity index — captures register elevation regardless of whether words are standard or non-standard.

sig. negativep < 0.001

WPS_1

Sentence length, Act 1

Average words per sentence in the opening act. Higher values indicate more complex, slower sentence structures in the setup and cold open.

sig. negativep = 0.020

Included in model — not statistically significant

Clout_3

Commanding voice, Act 3

LIWC Clout score: linguistic confidence and social dominance. High Clout speakers use fewer first-person singular pronouns and more certainty words. Range 0–100.

not significantp = 0.168

Dic_3

Standard vocabulary, Act 3

LIWC dictionary word percentage in the third act: proportion of words appearing in standard-English LIWC dictionary. Lower values indicate more slang or invented language.

not significantp = 0.246

Feel_1

Sensory / feeling language, Act 1

LIWC Feel category: references to sensory experience, bodily states, and emotion-adjacent perception (”felt,” “sense,” “touch”). Distinct from cognitive emotion processing.

not significantp = 0.419

WC_1

Word count, Act 1

Total words spoken in the opening act. A measure of dialogue density — how much the episode front-loads in terms of raw speech volume.

not significantp = 0.193

WPS_3

Sentence length, Act 3

Average words per sentence in the third act. Same construct as WPS_1, applied to climactic scenes. Notably, this act’s sentence length does not predict viewership — only Act 1’s does.

not significant

Appendix B · Full OLS Regression Output

Dependent variable: next-episode viewership in millions. OLS with HC3 heteroskedasticity-robust standard errors. N = number of episodes with complete script and viewership data.

Table 1. Full OLS regression output. Intercept = 6.53M baseline viewership. Rows highlighted in the analysis: WPS_1 (x7, p = 0.020), Sixltr_3 (x8, p < 0.001), Engaged_3 (x10, p = 0.035). Discrep_2 and Engaged_1 approach significance at the 10% threshold.

MetricValueInterpretationR-squared0.5806Model explains 58% of variance in next-episode viewership in-sampleAdjusted R-squared0.5369Penalized for 10 predictors — fit holds after complexity adjustmentF-statistic13.291Model is statistically significant overall (p < 0.001)Mean Absolute Error1.142MAverage prediction error per episode, in viewersRoot Mean Square Error1.422MPenalizes larger errors more heavily than MAECV Mean R² (5-fold)0.4239Out-of-sample performance — model generalizes beyond training episodesCV fold scores0.447 / 0.583 / 0.322 / 0.060 / 0.708Variation across folds suggests some seasons more structurally consistent than others

Note on the low CV fold (0.060): one fold covers a portion of the series where viewership was unusually volatile — likely driven by scheduling changes in Seasons 4–5 when the show shifted timeslots twice. This is a contextual issue, not a model failure.

Appendix C · Correlation Matrices — Show-Level and Genre-Level

Pearson correlation coefficients among all features in the model. Values above ±0.30 are generally worth discussing; values above ±0.50 indicate a relationship strong enough to affect interpretation of the regression coefficients.

Figure 2. Show-level correlation matrix, 2 Broke Girls. Key relationships discussed in the main article: Sixltr_3 ↔ Dic_3 (r = −0.51); Engaged_1 ↔ WC_1 (r = +0.49); Engaged_1 ↔ Engaged_3 (r = +0.33).

Three relationships in the show-level matrix merit technical attention:

r = −0.51Sixltr_3 ↔ Dic_3. The strongest off-diagonal relationship in the matrix. Climactic acts with high long-word density are simultaneously those with low dictionary-word coverage — a double signal of register elevation. This collinearity does not create a multicollinearity problem in the regression (VIF remains acceptable) but it confirms the two variables are tapping the same underlying construct: vocabulary complexity in the climax.

r = +0.49Engaged_1 ↔ WC_1. Emotionally engaged first acts are also wordier first acts. This creates a structural tension: Engaged_1 has a positive coefficient (+1.50) while WPS_1 has a negative one (−2.07). Episodes that open with emotional depth tend to be longer in the first act, which partially offsets the engagement benefit. The implication is emotional efficiency — get to the feeling in the fewest words possible.

r = +0.33Engaged_1 ↔ Engaged_3. Moderate correlation suggests emotional engagement is partly a whole-episode property and partly act-specific. Episodes that open with psychological investment tend to close that way too — but not always, and the independent coefficients confirm each act’s engagement contributes distinct predictive value.

Figure 3. Genre-level correlation matrix across sitcoms. Key relationships: No. of seasons ↔ No. of episodes (r = 0.84); Year ↔ WC_1 (r = +0.49); Year ↔ No. of episodes (r = −0.32); WPS_1 ↔ WPS_2 (r = +0.50); Network_Fox ↔ Dic_1 (r = −0.26).

Five relationships in the genre-level matrix merit technical attention:

r = 0.84No. of seasons ↔ No. of episodes. Near-perfect collinearity — these two variables are measuring the same latent construct (longevity) and should not be entered together in any cross-genre regression without dimension reduction or exclusion of one. While one could argument including both in a model doesn’t make sense, there is another argument that seasons captures longevity, while episodes captures output. It’s like thinking about how long you know a friend versus how often you see them annually.

r = +0.49Year ↔ WC_1. Significant positive drift: more recently produced sitcoms front-load more dialogue in Act 1. Likely reflects prestige-TV influence on network comedy norms. Relevant to the WPS_1 finding: if newer shows are both wordier and have longer sentences in Act 1, the viewership penalty identified in 2 Broke Girls may be more pronounced in recent-era sitcom data.

r = −0.32Year ↔ No. of episodes. Streaming-era truncation effect: more recent sitcoms produce fewer episodes per season. This compresses the within-season variance available for show-level regression models and may attenuate the act-level signals in more recent series.

r = +0.50WPS_1 ↔ WPS_2. Sentence length is a stable within-show feature across acts — a voice signature, not a structural decision. This validates the interpretation of WPS_1 as a show-level register property rather than an episode-level pacing choice. Corrective action at the episode level is likely insufficient; the register must be established at the room level.

r = −0.26Network_Fox ↔ Dic_1. Fox sitcoms use markedly less conventional dictionary vocabulary in Act 1 — consistent with the network’s historically edgier comedy brand. This suggests network identity is a dialogue-level variable, not merely a scheduling one, and should be included as a control in any cross-network genre-level analysis.

Appendix D · Model Specification and Limitations

Dependent variable: Next-episode viewership in millions (live + same-day). Using next-episode rather than same-episode viewership is deliberate: it measures retention rather than sampling, which is a more theoretically appropriate target for dialogue-level predictors.

Standard errors: HC3 heteroskedasticity-robust standard errors throughout. Viewership variance in early seasons (when the show had 15–19M viewers) is substantially higher than in later seasons (10–13M), making robust SEs necessary for valid inference.

Feature selection: The ten variables in the final model were selected from a larger candidate set using a combination of domain theory and regularization. Act-level segmentation was chosen over episode-level aggregation to preserve the temporal structure of narrative.

What the model cannot capture: Scheduling changes (the show moved timeslots three times), lead-in programming effects (the premiere benefited from the first post-Charlie Sheen episode of Two and a Half Men), competitive programming, marketing spend, and cultural moment effects. These forces account for the majority of viewership variance not explained by the model. The 42% out-of-sample R² represents dialogue’s independent contribution after those factors are roughly held constant within-series.

Generalizability: Findings are specific to 2 Broke Girls and should not be applied directly to other shows without replication. Cross-genre analysis is in progress. Preliminary results suggest the Sixltr penalty generalizes; the Engaged signal is more genre-specific.

I was surprised they didn’t pull through on this one. Paramount’s aggressive posture suggests something deeper — not confidence, but pressure. Consolidation is accelerating. DOJ hurdles remain. Nothing is final. But the direction of travel is clear.

And that direction should make creatives and consumers pause.

When fewer dominant buyers control distribution, the marketplace begins to resemble something monopsony-adjacent — many sellers (writers, producers, actors) negotiating with fewer powerful buyers. Even if we never reach a pure monopsony, movement in that direction matters. Competitive pressure disciplines pricing. It protects experimentation. It sustains diversity of voices.

Consolidation shifts leverage.

And when leverage shifts, the stage at which value is measured becomes critically important.

The Real Question Beneath the Negotiations

As the next round of SAG-AFTRA negotiations approaches, the entertainment industry once again finds itself circling the same underlying tension that has defined the streaming era: who creates value, how that value is measured, and where, precisely, economic leverage actually lives in the production process. The headlines will focus on residual structures, AI protections, streaming transparency, and the allocation of risk between labor and capital, and those are critically important debates. Yet beneath those negotiations lies a more fundamental economic question, one that receives far less attention than it deserves: at what stage of the creative process do we truly understand value?

What remains largely unmeasured is the development stage, where scripts are evaluated, pilots are greenlit, budgets are negotiated, and portfolio bets are placed. This is the moment when risk is concentrated and when future revenue trajectories are quietly determined. Historically, that stage has relied on experience, taste, reputation, and instinct. Executives read scripts and assess narrative potential; producers place trust in seasoned showrunners; studios depend on pattern recognition shaped by decades in the business. These processes are not arbitrary or unsophisticated. On the contrary, they are often deeply informed by cultural literacy, market awareness, and institutional memory.

Yet despite their sophistication, they are not instrumented. There is no systematic measurement of structural narrative risk at the moment when the largest financial commitments are made. There is no upstream retention probability model embedded in script evaluation meetings. The gap between instinct and measurement remains wide precisely where the financial stakes are highest. And it is in that gap, between creative intuition and economic instrumentation, that the next phase of cultural economics may emerge.

That stage has historically relied on experience, taste, reputation, and instinct.

Can Retention Be Detected in Dialogue?

In my recent study analyzing 25,000+ television scripts across 228 series, I asked a different question:

Can audience retention signals be detected at the script stage, before production, before casting, before marketing, using dialogue alone?

To test this, I trained machine learning models to predict next-episode viewership using only linguistic and emotional features extracted directly from scripts. Well, technically Open Subtitles, a website that houses closed caption data. Despite not having character content analytics available (which will change in the future – securing better data), the results, particularly in Table 3, which ranks the top 50 series by predictive robustness, suggest something economically meaningful:

Dialogue structure contains measurable information about whether audiences return.

This reframes content analytics as an upstream strategic capability rather than a downstream performance dashboard.

In other words, dialogue is not merely expressive. It functions as an upstream lever in the economics of serialized storytelling. When linguistic structure, emotional calibration, and act-level pacing systematically correlate with audience return, they become economically relevant variables rather than abstract artistic choices. And if measurable retention signals exist before a dollar is spent on production, then the development stage is not only a creative space but a form of strategic infrastructure, where narrative architecture quietly shapes long-term economic outcomes.

First, please see below the table 3 from my Journal of Marketing Analytics publication. There is a NLP definition list at the end of the article, but, you may be able to deduce that Fear_1, for instance, means the among of fear language in act 1 (-0.43, so in this case negative for A Million Little Things) is the way to read through this.

The Surprise: Endings Matter More Than You Think

Across dozens of series, including comedies, procedurals, prestige dramas, one pattern appears repeatedly:

Act 3 dialogue features show up as strong predictors of next-episode viewership.

Examples:

• Law & Order → WPS_3 (Words Per Sentence in Act 3) strongly predictive

• Dexter → WPS_3 highly significant

• Blindspot → Word Count_3 and Achieve_3 predictive

• Grey’s Anatomy → Engaged_3 strongly predictive

• JAG → Clout_3 significant

In plain English: how you end the episode matters more than how you start it. This aligns closely with Peak-End Theory from behavioral economics, which suggests that people evaluate experiences largely based on their emotional high point and how they conclude. What’s new here is that we can now measure that effect directly in scripts, before the cameras even roll.

Tight Writing Wins (Usually)

One of the most consistent predictors across series?

WPS — Words Per Sentence.

In many shows:

• Higher WPS in early acts → negative effect on retention

• Tighter sentences → stronger audience return

For example:

• 2 Broke Girls → WPS_1 negative and significant

• 30 Rock → WPS_1 and WPS_2 negative

• Adventure Time → WPS_1 negative

This suggests something counterintuitive: dense, overly complicated dialogue early in an episode may actually reduce the likelihood that viewers return. Audiences don’t reward verbal clutter. They reward clarity, coherence, and narrative momentum.

Emotion Isn’t Optional. It’s Structural.

Fear, joy, disgust, sadness, and anger are not abstract themes. They are statistically significant signals embedded in dialogue.

Examples from Table 3:

• Cold Case → Fear_3 positive predictor

• Criminal Minds → Joy_3 strongly negative (emotional tone matters by genre)

• CSI: NY → Joy_1 negative, Neutral_3 positive

• Black-ish → Disgust_1 and Engaged_2 significant

• Breaking Bad → Feel_3 significant

• Bull → Engaged_2 negative

What does this mean?

Emotional intensity, especially in the final act of an episode, predicts viewer retention. However, the direction and magnitude of that effect are moderated by genre. Procedurals respond differently than sitcoms, and serialized dramas behave differently than episodic comedies. In other words, emotion is not one-size-fits-all; it operates as genre-conditioned leverage.

Engagement Is the Quiet Power Variable

One of the most recurring features across high-performing series:

Engaged

This composite measure reflects psychological involvement, immersion, and emotional investment in dialogue.

It shows up repeatedly in:

• NCIS

• Criminal Minds

• Black-ish

• 2 Broke Girls

• Empty Nest

When engagement in dialogue drops, retention often drops.

When engagement intensifies near the end, retention improves.

In other words:

Audience loyalty follows narrative immersion.

Procedurals Are Predictable. Spectacle Isn’t.

When you compare predictive performance across series, structured procedurals like CSI, NCIS, and Cold Case exhibit stronger dialogue-based predictability. The reason is structural: their storytelling rhythm is consistent, with clear act breaks, deliberate conflict escalation, and a reliable resolution cadence. In these formats, dialogue scaffolds the plot.

By contrast, spectacle-heavy or visually driven series tend to show lower predictability from dialogue alone. In those shows, much of the action does not live in the words. And the model can only measure what is written.

Language Complexity Is a Double-Edged Sword

Another recurring feature: Dic (language complexity).

In some series, higher linguistic complexity predicts retention.

In others, it hurts.

For example:

• Adventure Time → Dic_1 positive

• Empty Nest → Dic_1 negative

• Seinfeld → Dic_3 negative

This reinforces a core insight from cultural economics: audience response is not driven by content being “more complex” or simply “more emotional.” Instead, it depends on alignment, between narrative structure, genre expectations, and the emotional arc delivered to viewers.

What This Means for Creators

Let’s translate Table 3 into creative strategy:

1. End strong. Dialogue in Act 3 disproportionately matters.

2. Keep early dialogue tight. Verbosity is not sophistication.

3. Build emotional peaks intentionally. Fear and tension often outperform neutrality.

4. Genre discipline matters. What works in a crime procedural won’t work in a sitcom.

5. Engagement is measurable. Immersion is not mystical.

This doesn’t replace artistry.

It enhances it.

A Practical Preview

To illustrate what this looks like in practice, I recently ran the pilot episode of Stranger Things (“The Vanishing of Will Byers”) through a new analytics system I’ve been building.

A few observations:

• Joyce’s dialogue share was 16.5% of all dialogue.

• Joyce appears in 11 scenes.

• She speaks over 1,148 words.

In other words, more than half of the verbal space is anchored to one character.

That’s not just storytelling.

That’s structural concentration.

Character centrality, dialogue dominance, emotional arc distribution — these are measurable.

And when measured systematically across thousands of episodes, patterns emerge.

*Importantly: this tool does not train on, scrape, or exploit scripts using large language models. It does not “learn from” copyrighted material. Instead, it measures structural, emotional, and dialogue-based features within scripts to generate analytical insights across television, film, podcasts, and other narrative media.*

The future of cultural economics will not be built on dashboards that tell us what already happened.

It will be built upstream — where narrative architecture quietly shapes long-term economic outcomes.

We are only beginning to instrument that layer.

And that changes everything.

NLP Library Definitions

Emotion and psychological engagement metrics

Sd_Scaled: The standard deviation is scaled across all emotions for comparative analysis, derived from the sum of standard deviations per act. This approach provides insight into emotional variance across different story segments. It is based on a transportation measurement that assesses the speed and depth of a viewer’s immersion in a story—both mentally and emotionally. This metric captures how quickly audiences lose awareness of their surroundings and become absorbed in the narrative. By dividing the standard deviation of overall emotion scores by the mean, it quantifies shifts, variability, and intensity of emotions throughout the experience.

Anger: The level of anger present in an act.

Surprise: The level of surprise present in an act.

Disgust: The level of disgust present in an act.

Sadness: The level of sadness present in an act.

Neutral: The level of neutrality present in an act.

Fear: The level of fear present in an act.

Joy: The level of joy present in an act.

Positive: The amount of positive sentiment present in an act.

Engaged: High psychological involvement or emotional investment, characterized by increased use of personal pronouns, emotional words, and cognitive processing. Linguistic and narrative complexity metrics

Word Count: Total number of words in an act.

Analytic: The presence of analytical, formal, or logical discussion.

Clout: The extent of social status, confidence, or leadership discourse.

Authenticity: The presence of honest, non-filtered, or unregulated discussion.

Tone: A measure of positivity; scores below 50 indicate negativity.

WPS (Words per Sentence): The average number of words per sentence in an act. Advancing predictive content analysis: a natural language processing and machine learning…

Six Letter Word: Percentage of words with more than six letters.

Dic (Dictionary): Percentage of words captured in the dictionary. Cognitive processing metrics

Cogprocess: Aggregate measure of words indicating active information processing and mental activity, including causation.

Insight: Words that indicate realizations or understanding.

Cause: Words that signal causal relationships between elements.

Discrep: Words that indicate counterfactual thinking (e.g., should, could, would).

Tentative: Words that indicate uncertainty or possibility (e.g., maybe, perhaps).

Certain: Words that indicate absolute statements (e.g., always, never).

Differ: Words that indicate differentiation between concepts (e.g., but, else). Perceptual processing metrics

Perceptual: Aggregate measure of words describing perception (e.g., look, heard, feel). See: Words associated with visual perception.

Hear: Words associated with auditory perception.

Feel: Words related to tactile sensation. Motivational drives metrics

Drives: Aggregate measure of different motivations expressed in an act.

Affiliation: Words associated with social relationships (e.g., ally, friend).

Achieve: Words related to success and accomplishment.

Power: Words related to dominance and hierarchical structures.

Reward: Words related to receiving benefits or prizes.

Risk: Words related to danger or uncertainty. Spatial and temporal metrics.

Relativ (Relativity): Aggregate measure of words describing spatial relationships (e.g., area, bend, exit).

Motion: Words related to movement (e.g., arrive, go, car).

Space: Words indicating spatial direction (e.g., down, in).

Time: Words related to time duration (e.g., end, until, season).

The show arrived at a very specific cultural moment. At this time, there was peak cable bundling, early social media, pre-streaming fragmentation. For a generation of young viewers, it provided something uniquely stabilizing: a shared narrative ritual. You watched it at the same time as your classmates, downloaded the songs, debated the double life. It was monoculture with a pop soundtrack.

Now, twenty years later, Miley Cyrus returning for a Hannah Montana anniversary special on Disney+ is not simply a nostalgic reunion. It is a case study in how platforms convert collective memory into measurable engagement. It is meant to reignite fond memories for the halcyon days of yesteryear. For Gen Z, who has been bludgeoned by high university prices, a slightly fickle job market, and houses priced out of reach, adulthood has increasingly become hard to dance to.

Nostalgia as Digital Infrastructure

In digital spaces, nostalgia is no longer passive remembrance. It is algorithmically activated memory, lying in cyclical wait like cultural cicadas, resurfacing every 10 to 20 years to remind consumers of timelines, life stages, and the markers of progress or stagnation. Before consumers fully register it, these cultural touchstones fade again, returning to the collective vault, only to reemerge when platforms determine the moment is ripe. What makes this powerful in streaming environments is that platforms can now quantify that continuity, as they measure completion rates, rewatch behavior, and cross-generational co-viewing. Therefore, Nostalgia is no longer anecdotal, it is data.

Platforms like Disney+ understand that nostalgia operates as both emotional glue and economic leverage. When viewers revisit a show from adolescence, they are not just rewatching content. They are re-engaging with a prior version of themselves. In an era of identity fluidity and constant digital reinvention, nostalgia provides continuity. Instead of social identity theory, in which consumers may compare themselves to others it is, it is more temporal identity theory, comparing ourselves across time.

In this sense, the Hannah Montana anniversary special is less about reviving a character and more about reactivating a cohort, and enabling them to temporarily shed their current identities and take on the cultural cloaks of their former selves.

From Cable Event to Platform Ecosystem

Teen media has undergone a structural transformation over the last two decades.

In the cable era, shows like Hannah Montana thrived on scarcity and schedule. The Disney Channel controlled distribution, advertising inventory, and audience flow. The model was vertically integrated but temporally constrained. Moreover, it was a purveyor of cultural tastes, styles, and tones. Language choice, jean selection, food selection and other pieces of daily life were determined in part by this channel for its constituents, much like MTV was for teens and college age adults.

Streaming reverses the temporal constraint but intensifies platform integration. Disney+ doesn’t just air the anniversary special, it surrounds it with the full catalog, curated playlists, behind-the-scenes features, and recommendation loops, enabling viewers to enter into a nostalgic ecosystem, carefully curated with transportive content. The viewer doesn’t “tune in.” They enter an ecosystem. In finance terms, nostalgia reduces volatility. It stabilizes engagement by tapping into pre-existing emotional capital.

Generational Identity and the Performance of Memory

Anniversaries like this are not random, as they serve as low-risk engagement spikes. Unlike launching a new IP, reviving a beloved property carries built-in audience awareness and emotional investment, and exists as a portfolio diversification through familiarity. Importantly, platforms are not merely banking on sentiment, but on intergenerational transfer. Gen Zers (and some younger millennials) who grew up with Hannah Montana now introduce it to younger viewers (though current birth rates among younger consumers suggest this may be an uphill battle). What was once teen programming becomes family content. In this sense, an anniversary revival becomes a stress test of the IP itself. Paired with recommendation algorithms and adjacent catalog surfacing, it also serves as a bellwether for how early-2000s content can be repackaged and repositioned for contemporary audiences.

Anniversary specials also reveal something about audience behavior today: viewers increasingly perform nostalgia publicly. Social media turns memory into mystified spectacle. TikTok edits, Instagram reels, “I can’t believe this was 20 years ago” posts are full-throated expressions, and are not private reflections. They are communal signaling devices, signifying belonging in the Gen Z cohort.

When Miley Cyrus says that Hannah Montana “shaped my life and the lives of so many fans,” she is acknowledging that the show functioned as a shared developmental text. Viewers grew up alongside her. Her rejection of the persona in the 2010s mirrored their own desire to outgrow adolescence. Her return now mirrors a broader cultural comfort with integrating past selves rather than disavowing them.

In a fragmented media landscape, nostalgia becomes a shortcut to community. It creates temporary monoculture moments inside otherwise individualized algorithmic feeds.

Rather than choosing one identity over another, her journey reflects the possibility of holding both in tension, much like the duality she navigated on Hannah Montana. The difference is that she is no longer playing a character; she is reconciling the public artist with the private individual. For a generation raised online, where public and private selves constantly intersect, this duality feels deeply familiar. Much of this dynamic echoes Erving Goffman’s dramaturgical theory of “front stage” and “back stage” life. Our front-stage personas are performed socially, on platforms, in professional spaces, and in curated digital feeds, while the backstage is where the performance drops and the self can exist without audience scrutiny. Social media, however, has increasingly blurred these boundaries. For those whose lives are fully mediated online, the backstage narrows. The risk is not simply exposure, but erosion, and perhaps a gradual loss of the private space necessary to construct an unperformed self. Gen Z has grown up more outspoken, more comfortable embracing vulnerability, and with fewer perceived boundaries between online and private selves.

Twenty years later, many of us are all leading Hannah Montana-esque lives.

Broader Cultural Implications

The return of Hannah Montana also reflects something larger about contemporary media culture. We are living in a moment where past and present collapse into each other. Streaming libraries flatten time and everything is simultaneously available. That changes how identity is constructed, and has stalled how cultural trends are anticipated, elevated, monetized, and faded. Instead of moving linearly through eras, audiences curate their own temporal playlists. There is no one single source for pop culture, and whatever enters into being trendy can quickly fall out of vogue.

Nostalgia, then, is not regression. It is recombination with the present. It creates a reflective distance between the present and past, and enables us to recognize the thru-lines of our own lives, along with continuities and changes. Amid these boundaries, new interpretations may emerge.

Platforms understand this deeply. They are not just distributors of stories, but exist as custodians of cultural memory. When they activate an anniversary, they are effectively saying: this moment mattered, and let’s measure how much it still does.

And judging by the anticipation surrounding Miley Cyrus’ return, it still matters quite a bit.

Everyone is talking about Disney’s next CEO.

What’s missing is the more important conversation beneath it.

What does this choice say about the direction of entertainment itself?

Dana Walden, my personal selection, has built her reputation on deep, durable relationships with creatives. She understands storytelling, talent, and how value is created early in the entertainment value chain at the level of scripts, characters, and narrative worlds. She operates at the front of the flywheel.

Josh D’Amaro, by contrast, has enjoyed blockbuster financial momentum running Disney’s parks and experiences. His success is downstream: monetization, pricing power, operational scale. He sits at the end of the flywheel, where value is harvested rather than created.

I do wonder, outloud, if Iger and board thought that, with the glacial decline of the celebrity (I dare you to list five whose last names are not Sweeney, Chalamet, Ortega, Zendaya...ugh I think I lost my own dare), that really what consumers will care about, and continue to care about, are experiences. This comes up against more high-price attractions and experiences, as wealthy people are moving toward experiences as a status symbol (https://lnkd.in/eDxD_6_X). Everyone can enjoy a Chalamet movie. But not everyone can enjoy high-priced theme park experiences (https://lnkd.in/efayHiYT).

Therefore, are Walden’s relationships with creatives not as valuable as they might’ve been a decade ago?

As wealth concentrates and consumption shifts, experiences, not goods, and not even stars, are becoming the clearest status signal. They are scarce, exclusionary, and impossible to pirate or infinitely scale.

That’s why the organizational dynamic here is so interesting. Walden now leads creatively (soon to be chief creative officer in a few months) at the beginning of the system, while D’Amaro controls the most powerful monetization engine at the end of it. In effect, the head reports to the tail.

That inversion tells a larger story.

The era of celebrity doesn’t disappear, but it dims as the primary profit engine. In a fractured entertainment marketplace, money is no longer made by commanding mass attention. It’s made by monetizing smaller, segmented audiences through experiences they can’t all access.

That may be the real entertainment shift Disney is quietly betting on.

And this is where the conversation becomes more interesting.
If experiences are the new monetization engine, then what happens to the creative inputs that power the flywheel in the first place?

------------------------------------------------------------------------------------------------------

For decades, television scripts were treated as purely creative artifacts, important for storytelling, but resistant to systematic measurement. Meanwhile, advances in machine learning and natural language processing (NLP) have transformed how researchers study text in domains like marketing, politics, and finance. This article argues that television dialogue deserves the same analytical treatment. Scripts are the first concrete input in the TV value chain, long before casting, marketing, or scheduling decisions are made. If audience reactions are shaped by narrative structure, emotional intensity, and linguistic style, then those signals should already be present, at least in part, within the script itself.

Over the next few months, I’m going to introduce my most recent study that was published in the Journal of Marketing Analytics (but this is going to be WAY more fun). This study positions itself between these worlds by asking a simple but provocative question: can we predict audience viewership before a show airs by analyzing how characters speak, emote, and interact across episodes? Doing so reframes scripts not as artistic black boxes, but as early data assets for understanding engagement and demand.

The framework draws on three strands of theory. First, peak-end theory suggests audiences remember experiences based on emotional highs and endings rather than averages, making emotional spikes in dialogue especially consequential. Second, seriality and engagement theory emphasizes continuity, escalation, and momentum across episodes—features that are often encoded linguistically through pacing, repetition, and evolving tone. Third, narrative and dialogue research highlights how conflict, resolution, and character voice shape attention and identification. Together, these perspectives justify a close reading of dialogue features, such as emotional intensity, linguistic complexity, and conversational structure, as drivers of sustained viewership.

Methodologically, the study analyzes over 25,000 television episode “scripts” spanning decades, genres, and network contexts. I used an open sub caption website (because TV scripts are SUPER hard to find in bulk) to use closed caption dialogue. Each “script” is segmented by narrative acts and processed using validated NLP dictionaries and metrics. Features include emotional dimensions (e.g., anger, sadness, surprise), linguistic complexity (e.g., word count, words per sentence), cognitive processing cues, and indicators of psychological engagement. These measures are designed to capture not just what is said, but how it is said, reflecting pacing, clarity, tension, and emotional flow within and across episodes.

To translate dialogue into predictions, the study uses a range of regression and machine-learning models, including OLS, Lasso, Ridge, Elastic Net, Gradient Boosting, and XGBoost. Each model offers a different way of detecting patterns between script features and audience behavior. The models are trained to predict next-episode viewership, which mirrors how audiences actually make decisions: they watch one episode and then choose whether to continue. To avoid overfitting or relying on noisy variables, feature selection is guided by random forest importance scores, helping the models focus on the most meaningful dialogue characteristics while reducing overlap among similar measures. Model performance is evaluated using cross-validated R² (which measures explanatory strength, I’ll get to that) as well as MAE and RMSE (which measure prediction error). This layered approach combines interpretability with predictive power, allowing the analysis to identify which dialogue features matter most, and under what conditions, without treating the algorithm as an opaque black box.

To just look at a few of the top TV series, based on this, here is a cut out from the study itself:

WOW. I know. For our purposes, just focus on the left-most column: Test CV R-squared.

Before diving into what the table shows, it helps to understand what that number actually means. “R-squared” is simply a score that tells us how much of the change in next week’s viewership can be explained by analyzing the script’s emotional tone and language patterns. A higher number (closer to 1) means the script does a better job predicting what happens to ratings. For example, a score of 0.80 suggests that about 80% of the ups and downs in audience numbers are connected to patterns found in the dialogue. A score of 0.20 means the relationship is much weaker.

The “test” part means the model is evaluated on episodes it hasn’t seen before. In other words, it’s not being graded on homework it already memorized — it’s being tested on new material. The “cross-validated” part means this testing process is repeated multiple times, each time holding out different episodes, to make sure the results aren’t a fluke.

So when you see a high Test CV R-squared, it means the emotional and linguistic patterns in the script reliably help predict what audiences will do next — even when the model is applied to new data. In short, the table is asking a simple question: how predictable is audience behavior if we only look at what’s written in the script?

------------------------------------------------------------------------------------------------------

When we ranked the top 100 television series based on how well our model could predict next-episode viewership from script dialogue alone, one thing became immediately clear: not all shows are equally predictable. For some series, the emotional tone and language patterns embedded in the script strongly forecast what happens to ratings next week. For others, the relationship is much weaker. In other words, some shows behave like steady engines, while others behave like roller coasters.

Long-running sitcoms and procedural dramas tend to be the most predictable. Shows like Seinfeld, Home Improvement, CSI, and Law & Order follow relatively consistent storytelling formulas. Characters behave in familiar ways. Emotional rhythms repeat. Conflicts rise and resolve within expected boundaries. That structural stability makes it easier for a model, and arguably for audiences, to anticipate what comes next. When narrative structure is consistent, audience behavior becomes more consistent too.

By contrast, highly serialized or shock-driven shows are much harder to forecast. Series that rely on major twists, tonal shifts, or evolving story arcs don’t follow the same repeatable emotional patterns. A dramatic character death, a surprise betrayal, or a season-level narrative pivot can dramatically alter viewership in ways that simple dialogue patterns cannot fully capture. These shows generate excitement and buzz, but also volatility. Predictability declines as narrative risk increases.

Another important factor is scale. Some shows simply have much larger audiences than others, and that naturally increases prediction error in absolute terms. A small miss in a show averaging 20 million viewers looks very different from a small miss in a show averaging 2 million viewers. So differences in model performance don’t necessarily mean the storytelling matters less, sometimes it just reflects audience size and fluctuation.

The bigger takeaway is this: scripts matter, but they matter differently depending on the type of show. Stable, formula-driven series produce more reliable audience patterns. Complex, prestige, or shock-heavy series produce more dramatic swings. From a business perspective, this suggests that television portfolios benefit from balance, dependable “steady earners” alongside high-risk, high-reward narrative bets. Storytelling isn’t just art; it also carries measurable patterns of stability and volatility.

There will be more to come! The next article will delve into more results from this article!

As the champagne bottle fizzes have faded (though elegant bubbles are surely still popping somewhere) and our collective focus has slowly climbed out of the soporific, indulgent two weeks following Christmas and New Year’s, it feels like we’re all resetting—again.

Some of us have had to pull ourselves away from the idolatry of new gifts, distractions, video game consoles, movie marathons, and whatnot, just to get back at it—and maybe even attempt a resolution or two.

For me, that reset has taken the form of cross-training. Over the past year, I’ve started leaning into it more deliberately, working on all parts of my body with greater care as I enter my final year of being in my 30s and begin the gentle descent from the peak…morbid, right? Maybe it’s the brisk Connecticut cold that’s warranted such visceral language.

More importantly, I’ve been meaning to talk to you for a long time—well, frankly, all of you.

There’s no shortage of cultural analysis of media and entertainment (I’ve contributed to plenty of it) across publications, newsletters, and news outlets. But there’s a noticeable absence of economics, machine learning, and other quantitative tools, methods that may sound soporific at first (though this is the second time I’ve used this word!), yet can meaningfully illuminate how these industries actually function and evolve.

For years, I wanted to start a blog, but I didn’t quite have a reason. I teach at the Darden School of Business, and it’s taken me nearly six years to clarify what I truly want to focus on: the measurement of art. Don’t get me wrong—I remain fascinated by studying people. After all, where does one even begin to unspool the psychological mystery behind why something “just works”? We take for granted how so many elements just work, but we don’t stop to consider just how difficult it might be to get something to work in the first place, right?

We’ve studied people, markets, and firms for centuries, yet I’m not convinced we’ve applied the same high-touch, systematic quantitative lens to art itself. Perhaps there isn’t a clean way to do so. Who the hell thinks about Bayesian statistics while looking at a Jackson Pollock—considering the probability distribution of paint flung from an alcohol-infused hand?

Still, I’m motivated by persistent questions: how we price content, what content means to us, and how—at a deeper, qualitative level—it shapes and advances society. Why does anything become dank, slap, dope, lit, goated, or even bussin’ for that matter? Can a blend of media theory, business thinking, math, and a little curiosity help illuminate those murkier corners?

Part of what’s pushed this further is that I’m currently a part-time MS in Economics student at Purdue’s Daniels School of Business. During my PhD in Media at the University of Florida—and even earlier during my M.A. at Syracuse’s Newhouse School—it became clear, perhaps a bit too late, that I should have pursued a PhD in business while studying media. That realization sent me down a different path: multiple machine learning bootcamps (all pre-ChatGPT), auditing three econometrics courses, and steadily building technical fluency (If all of this sounds impressive and a bit braggadocious, I’m currently getting a beat-down in calculus).

Along the way, one thing crystallized. This is the way forward: to marry quantitative methods with culture—to combine economics, machine learning, and media studies in an effort to distill what truly moves us. I’m convinced that this kind of intellectual cross-training is essential if we want to understand the nuances of entertainment in an era where formats, platforms, and business models increasingly blur together.

To start this year, and to hopefully maintain its focus, I’ll be distilling insights from a repository of roughly 25,000 TV episodes paired with Nielsen data and financial filings, work that’s currently underway, to better understand how content might move the needle, and whether there are new ways to think about the value of art itself. There may be other commentary peppered along the way, of course, but TV scripts and movie scripts (working on that too currently) are arbiters of change. Why? Dialogue! Our actions, our movements, the very fibers of our beings are built around dialogue. It’s why when I write business cases, they’re full of dialogue. No one hands people a manual at work, on a date (hopefully), or in other spheres of life. So much of what we do is guided and inspired by the language we elect to use. I’m fascinated by story arcs, emotional tenets, and thinking about financial implications in the stock market for TV firms, and at the box office for movies.

So join me. Really. I am doing this for academic work, but I am eager to build this into something that creatives, business leaders, and others can benefit from. There’s no catch here (unless you’d like to have one with me).

At the very least, we might be able to find some of the 2025 champagne left.