Sometime in late March 2026, a manuscript about a disgraced marine biologist and a cursed Scottish loch landed in the inbox of a mid-level acquisitions editor at HarperCollins Voyager. Standard submission. Twelve thousand words of sample chapters, a two-page synopsis, a comp list pointing at The Midnight Library and early Tana French. Under normal circumstances, it would have spent six weeks in the slush pile before a junior reader even opened it.
Instead, it got a preempt offer in eleven days. The reason had almost nothing to do with the editor’s gut. It had everything to do with a score: an 87.4 out of 100 on Prism, the AI virality prediction platform that HarperCollins quietly licensed in Q4 2025 for a reported $2.3 million annually. Prism had ingested the sample chapters, cross-referenced 140 million BookTok engagement data points, modeled the manuscript’s emotional arc against tropes currently converting at above-average save-to-share ratios, and returned a projection: high probability of sustained BookTok velocity in the 90-to-120-day post-publication window. The deal closed at $340,000.
Welcome to the acquisition meeting of 2026, where the most important voice in the room increasingly belongs to a model that has never actually read a book — at least not the way humans do.
How Large Is the AI Virality Prediction Market, and Who Controls It?
The market for AI-powered reader intelligence and virality prediction tools inside publishing has grown from an estimated $180 million in annual contract value in 2024 to approximately $1.1 billion by mid-2026, according to figures compiled by Codex Strategy Group, a boutique consultancy that advises three of the Big 5. That growth — roughly 510 percent across eighteen months — is not driven by a single dominant player but by a fractured, fast-moving ecosystem of at least nine commercially active platforms, each competing for exclusive data partnerships with publishers, agencies, and the increasingly powerful class of BookTok-native marketing shops.
The leading platforms, as of this writing:
- Prism (owned by Resonance Labs, Series C valued at $620M): Currently licensed by HarperCollins, Macmillan, and at least four major UK publishers. Specializes in trope-to-engagement mapping and chapter-level emotional volatility scoring.
- Kinetic (Bertelsmann-backed, integrated directly into Penguin Random House’s acquisition workflow since January 2026): Focuses on comp title decay curves — how long a comparable book’s BookTok momentum historically sustained before flattening — to model a submission’s commercial window.
- Vellichor AI (indie, founded by former Goodreads data engineers): Smaller but gaining ground with mid-size imprints and literary agencies; its “reader affect fingerprinting” system has become something of an open secret at Curtis Brown and Janklow & Nesbit.
- StoryPulse (backed by a16z, launched February 2026): The newest entrant and already the most controversial; its real-time TikTok comment sentiment ingestion operates in a legal gray zone that three platform attorneys contacted for this piece declined to discuss on record.
Taken together, these platforms processed scoring requests on an estimated 340,000 manuscripts in the first half of 2026 alone, according to Resonance Labs’ own published usage metrics. That number will seem either extraordinary or obvious depending on how closely you’ve been watching the signal.
What Does an AI Virality Score Actually Measure?
This is the question nobody inside publishing wants to answer too loudly, because the honest answer is simultaneously impressive and alarming. These systems do not predict quality. They predict pattern match — the degree to which a manuscript’s structural and emotional DNA resembles the DNA of content that has previously generated measurable BookTok behavior: saves, shares, duets, comment velocity spikes, and the increasingly important “emotional damage” comment clusters that platforms like Prism have learned to weight heavily as a leading indicator of word-of-mouth purchase conversion.
“We’re not telling publishers what’s good. We’re telling them what’s legible to a specific audience at a specific cultural moment. Those are very different things, and I’d be lying if I said the distinction doesn’t keep me up at night.” — Dr. Amara Osei-Bonsu, Chief Science Officer, Resonance Labs, in a rare on-record interview at the London Book Fair AI Summit, April 2026
Prism’s core model, for instance, assigns weighted scores across seven dimensions: emotional escalation gradient, trope saturation index (comparing a manuscript’s trope density to currently converting versus declining trope cycles on BookTok), prose velocity (measured at the sentence and paragraph level as a proxy for scroll-stopping readability), identity mirroring potential, “obsession inflection” density — meaning the frequency of scenes likely to generate rewatching or re-reading behavior — cover visual compatibility, and projected creator briefability, which is essentially a machine estimate of how easily a human BookTok creator could construct a compelling hook from the material.
The last dimension is perhaps the most revealing about how thoroughly BookTok’s economics have colonized the acquisition process. Publishers are now, at least partially, acquiring for creator utility.
Are Agents and Authors Aware Their Work Is Being Scored Before Humans Read It?
Largely, no. And this is where the industry’s discomfort becomes acute.
Submissions sent to major imprints are typically scored automatically upon intake, before a human editor opens the document. None of the Big 5 publishers contacted for this piece confirmed this practice on record. Two agents — both senior, both representing commercially successful BookTok-era authors — confirmed it off record, one of them describing the process as “the slush pile getting a credit score before the loan officer sees your face.”
“I had a client whose manuscript scored low on Kinetic’s trope cycle index because the enemies-to-lovers arc resolved in chapter nine instead of the model’s preferred chapter-twelve-to-fourteen window. The editor liked the book. She got overruled by a committee citing the score. This is real. This is happening.” — A senior agent at a major New York literary agency, speaking anonymously
The Writers Guild of America East, which has been expanding its author advocacy work following the 2025 AI disclosure battles, issued a formal inquiry to the Association of American Publishers in May 2026 requesting transparency standards around algorithmic screening of submitted manuscripts. The AAP has not yet responded publicly. At least two BookTok-native authors — including one whose debut sold 400,000 copies through TikTok Shop in 2025 — have told this publication they had no knowledge their agents were running Vellichor pre-submission scores before shopping their second books.
How Are AI Virality Scores Changing the Economics of the Deal?
The financial distortion is measurable and growing. Analysis of deal data compiled by Publishers Marketplace across Q1 and Q2 2026 shows that manuscripts receiving documented high-confidence scores from major platforms are closing at average advances 34 percent higher than comparable genre submissions without platform endorsement — a premium that did not exist in 2024 data. At the upper end, “Platinum tier” Prism scores (90 and above) have correlated with deals that exceeded initial editor ceiling offers by an average of $180,000, as acquirers competed partly on the strength of the score rather than exclusive editorial conviction.
The flip side is darker. Manuscripts scoring below platform thresholds — even those championed by experienced editors — are increasingly struggling to get to offer stage. Three acquiring editors at different imprints described, independently, a version of the same meeting: an editor arrives with a manuscript they believe in, a score appears on the shared screen, and the conversation shifts from “is this a good book?” to “can we trust our instinct against the model?” In a risk-averse post-2025-returns-crisis publishing environment, the model is winning that argument more often than it should.
“The score is becoming a liability shield. If a book fails after a high Prism score, no one gets fired. If a book fails after an editor overrode a low score because they ‘believed in it,’ someone gets fired. That calculus has already changed the room.” — Lena Vasquez, former senior editor at Simon & Schuster, now running her own imprint consultancy, speaking publicly at BookExpo AI track, June 2026
What Does This Mean for the Books That Will Define 2027?
This is the non-obvious thesis, and it requires sitting with a slightly uncomfortable idea: AI virality prediction tools are not just reflecting BookTok culture. They are beginning to generate it — by systematically selecting which manuscripts receive the marketing infrastructure, creator seeding budgets, and retail placement that create the conditions for virality in the first place. A book that scores an 88 on Prism gets a $200,000 creator campaign. A book that scores a 61 does not. The 88-scored book goes viral. Prism’s predictive accuracy improves. The model learns to prefer the features of the 88-scored book even more strongly. The cycle closes.
What this produces, over a publishing cycle or two, is a feedback loop that narrows the emotional and structural range of commercially supported fiction in ways that are not yet fully visible because we are only eighteen months into meaningful platform adoption. The books acquiring editors are being quietly steered toward in 2026 will be the dominant BookTok titles of 2027. And those books will, disproportionately, share DNA — not because readers demanded that homogeneity, but because a model trained on past virality is, by definition, optimizing for the past.
There are counter-pressures. Vellichor AI’s affect fingerprinting system was specifically designed, its founders say, to reward emotional novelty rather than pattern replication — to score manuscripts higher when they produce sentiment signals that diverge from existing BookTok emotional templates. StoryPulse claims its comment sentiment engine can detect “emerging affect” before it scales, theoretically allowing it to identify the next emotional register before the current one peaks. Whether these design philosophies survive contact with commercial pressure from the publishers paying for access is, at best, an open question.
What is not an open question: the $1.1 billion flowing into these platforms has already redrawn the invisible architecture of which stories get told. Agents are pre-scoring manuscripts before submission. Editors are using scores as political cover. Marketing budgets are flowing toward algorithmic confidence rather than editorial conviction. And somewhere in a slush pile right now, there is a manuscript that would have broken BookTok in ways no model can currently predict — precisely because it doesn’t match anything the model has seen before.
That book is getting a 54. And a form rejection is three weeks away.