You Can’t Write a Bestseller. That’s OK, Because AI Can’t Either
In the age of AI being able to write your novel, we have never figured out what makes someone read a novel.
In 2016, two researchers published a book called The Bestseller Code. Jodie Archer, a former Penguin editor, and Matthew Jockers, co-founder of Stanford’s Literary Lab, fed 20,000 novels into a machine learning algorithm and asked a simple question: can we predict what will hit the New York Times bestseller list?
They could. Eighty to ninety percent accuracy. The algorithm identified 2,799 features strongly associated with bestsellers, from thematic concentration to emotional arc pacing to the ratio of everyday language versus literary vocabulary. It could tell you, with startling confidence, whether a manuscript had the DNA of a hit.
The book was reviewed in The New Yorker, discussed in publishing circles, debated in literary journals. And then it settled into its final resting place on Amazon, somewhere around number 35,000.
The Bestseller Code was never a bestseller. In fact, it didn’t really make an impact, because while interesting, it’s not actionable.
And that is what is going to keep authors who want to produce their own books safe, even as AI floods the market, AI-enabled authorship will still be viable. Because even if an AI wrote a novel, and even if I was a bestseller, we do not, as a society, have the answer to WHY it was a bestseller, only that it was. Which means, we cannot reverse engineer it reliably, and neither can AI.
You Cannot Turn Smoke Back Into Wood
The researchers proved something important. They proved you can reverse-engineer the DNA of a hit. You can extract patterns from themes, pacing, emotional arcs, stylistic fingerprints. The algorithm works. The analysis is real as are the findings.
But you cannot run it in reverse.
Think about combustion. You can study smoke and ash with extraordinary precision. Gas chromatography, mass spectrometry, molecular analysis. You can catalog every compound: the CO2, the water vapor, the particulates, the trace minerals in the ash. From that catalog, you could theoretically describe the wood that burned. Its species. Its age. Its moisture content.
But you cannot make the wood. Combustion is irreversible. The Second Law of Thermodynamics is blunt about this. Some processes only go one direction.
A bestseller is a combustion event. A unique convergence of craft, timing, cultural readiness, audience hunger, distribution luck, and something no one has ever been able to name. The Bestseller Code cataloged the smoke with exquisite precision. It could not relight the fire.
Goodhart put the problem most precisely:
“When a measure becomes a target, it ceases to be a good measure.”
The moment you identify the 2,799 features and optimize for them, those features stop predicting success. You do not get a bestseller. You get a text that games a model. There is a parable about a nail factory: set “number of nails” as the metric, and workers produce tiny useless nails. Switch to “weight of nails,” and they produce giant useless nails. Neither approach produces anything you would actually want to hammer.
After the book came out, Jockers co-founded a startup called Authors.AI. They built Marlowe, an evolved version of the algorithm, informed by over a hundred bestselling authors. The response from early users was (in their words) “overwhelmingly positive.”
But there are no documented cases of anyone using Marlowe to produce a bestseller. Not one. The startup was real, the algorithm was refined, the product shipped, and the silence that followed is the loudest data point of all.
You cannot turn smoke and ash into wood.
The 100 Readers Problem
But, let’s forget the New York Times list. Forget going viral. Forget the algorithm entirely.
If I gave you the best book ever written and said “get 100 strangers to read it and genuinely like it,” how would you do it? Not a hundred thousand. Not a million. A hundred. Most honest people would say: I have no idea.
That is the asymmetry at its most naked. We have solved writing (or at least accelerated it dramatically). We have solved editing, cover design, formatting, distribution, even marketing copy. AI can help with every single one of those things.
But none of those are the actual problem. The actual problem is human attention. Getting a stranger to pick up your book, open it, read past the first page, and care. AI has solved a constellation of problems around that central problem. Every adjacent problem. Every problem except the one that matters.
(I recommend reading up on Wildbow and Worm, how the Martian was written, and how Wool was written - all examples of people finding their first 100 fans in very nontraditional ways that led to huge success.)
The Noise Floor
In one documented study, only 19 of the top 100 bestselling ebooks in Amazon’s Teen and Young Adult Contemporary Romance category were real books. The other 81 were AI-generated spam. Click-farmed gibberish that shot to the top of Kindle Unlimited charts, siphoning page-read revenue from actual authors who wrote actual words.
Amazon now caps self-published books at three new titles per day per author. Content volume coming into distributors is trending roughly fifty percent above normal. The signal-to-noise ratio is not just degrading. It is collapsing. Three books per DAY, people. Per DAY.
Back to the real world: less than one percent of all published titles sell more than 5,000 copies. Across creative platforms (YouTube, Spotify, Twitch, publishing) ninety-nine percent of revenue accrues to less than one percent of creators. The distribution is not a bell curve. It is a cliff: a skinny head of a few hits, a dwindling middle, and a seemingly infinite tail of obscurity. Most books end up spine-out on an end-cap and then gathering dust in boxes in the author’s garage, or worse, their covers ripped off and they are destroyed by the publishers who cannot warehouse them any longer.
The unsettling part is this: every book in that tail might be just as good as the ones in the head. Nassim Taleb called this the “silent graveyard,” the evidence we never examine because it never made noise. We study the bestsellers and construct elegant theories about why they succeeded. But we never study the thousands of books with identical craft, identical emotional arcs, identical thematic sophistication, that simply never caught fire. The algorithm studied the smoke of fires that ignited. It never looked at the identical kindling that did not.
The difference between those two piles was never in the wood, but we don’t know what it is. We can identify it, but we can’t force it.
The One Mechanism That Works
BookTok has generated over 190 billion views of book-related content. In certain months, BookTok-linked titles accounted for more than twenty percent of all fiction unit sales in the United States. Rebecca Yarros. Sarah J. Maas. Emily Henry. Colleen Hoover. Fantasy and romance surging, driven almost entirely by short-form video recommendations.
And BookTok is the proof. Not the solution, the proof. It is again only an observation, not a recipe. Because BookTok, at its core, is a person holding up a book and crying. That is the mechanism. A human being, visibly moved by something they read, sharing that emotion with other human beings who are looking for something to feel. It cannot be manufactured. It cannot be prompted. It cannot be optimized.
The most powerful book discovery mechanism on the planet runs on human emotion about a book that no algorithm has read. You cannot engineer that reaction. You cannot A/B test your way into it. The most an author can do is put the book in someone’s hands and hope.
That “hope” is the asymmetry that AI can’t help with yet, and maybe never will be able to.
Plus, even BookTok concentrates rather than democratizes. The same titles circulate, the same authors trend, romantasy dominates. Cross-genre fiction, unconventional narratives, literary science fiction (which is what I write) does not fit the keyword-and-category system that makes the recommendation engine turn. BookTok may spread influence more broadly (you do not need a massive following to create a viral recommendation) but it narrows attention onto a thin band of the shelf.
Again, BookTok is evidence, not instruction. Knowing it works doesn’t help you know why it works, meaning you can’t be a queenmaker.
The Precious Art Problem
Here is something uncomfortable that needs saying: the idea that most books are precious works of art is undeniably false.
That is the reader’s problem, not the author’s. Most commercial fiction is formulaic, built on tropes that publishers fill their catalogs with because those tropes sell. Occasionally a black-swan book redefines the market, and then everyone races to copy it. The publishing industry has always been a pattern-matching machine dressed up in the language of literary curation. The publishing industry is bad for writing, but great for marketing.
AI will be able to produce books that are experientially indistinguishable from the average airport-store or Barnes and Noble bestseller. That is probably true now. It will be more true soon. And the constant background noise in every conversation about AI and creativity is the accusation that the machines can write your novel for you.
But that accusation misunderstands what authorship actually is. The author’s problem was never “how do I produce 80,000 words of readable prose.” The author’s problem is getting a vision rendered the way you imagine it. Staying in the driver’s seat while the scaffolding assembles around you. Having perfect recall of your own continuity across hundreds of scenes. Keeping motivation alive through the years-long slog of drafting.
I wrote four books the hard way from 2018 to 2021. I know exactly where the friction lives: the overhead, the motivation gaps, the continuity errors that creep in at scale, the sheer impossibility of holding a 400,000-word universe in your head while writing page 847. Those are the problems that kill books. Not lack of prose generation.
So I am building a tool to solve them. Fictioneer is not a writing assistant. It is a Socratic muse, an editor with perfect memory, an advisor who knows your canon better than you do. You stay in the driver’s seat. The tool accelerates, enhances, and improves the process without taking away authorship or ownership. Think of it as an instant, always-available crew of writing helpers, the kind of staff a publisher provides for a bestselling author. Fictioneer is that staff. But the art is still yours.
Why am I building an entirely new piece of software to do this? Because I can’t solve the overhead problem. Life moves too fast, the burdens of being an adult, a parent, a breadwinner, and having a day job means that writing a novel the traditional way would take me decades, not years. The only reason I got the books published I did was because of COVID lockdowns.
For me, in my position in life, willpower, grit, waking up at 4am, dedicating time each day to my art: it’s still not enough. I simply cannot sacrifice more of my time to actually produce books at a pace that matters, regardless of how much discipline I have. The only reason I was able to do what I did was because the lockdown created a unprecedented amount of free time, likely to never be repeated in my life (or at least I fucking hope not!).
A Thousand Matches
Kevin Kelly (co-founder of Wired) wrote an essay in 2008 called “1,000 True Fans.” The premise is simple: you do not need a bestseller. You need a thousand people who care enough about your work to buy everything you make.
For most of publishing history, reaching those thousand people required permission. A publisher had to believe in you. A bookstore had to stock you. A reviewer had to notice you. The gatekeeping was not just about quality. It was about access to the distribution system itself.
Summed up: books are optimized to fill gaps in publishers catalogs, not in readers’ minds.
See, this is where AI actually changes things, though not in the way most people discuss. AI is not going to write your bestseller. But it can lower the barriers that kept you from reaching your thousand true fans. The same tools that let spam flood Kindle Unlimited also let a serious, committed author produce polished work without a publisher’s resources. Cover design, formatting, editing assistance, marketing copy, metadata optimization: the entire apparatus of professional publishing is becoming accessible to individuals.
The question is not “can AI write a bestseller?” (It cannot.) The question is: can you produce work that is so potent and so specifically yours that it finds its thousand? Not by reverse-engineering what sells, but by going deeper into what only you can make?
That reframe shifts the entire paradigm. The Bestseller Code tried to map the smoke of mass-market combustion. Kelly’s framework says: forget the bonfire. Light a thousand candles. Different fuel. Different physics. Different outcome. The entire paradigm around novels and books is the evolutionary descendent of mass-market publishing and treating authors like racehorses who either win, or are sent to the glue factory.
Where This Leaves a Novelist
I write science fiction with political intrigue and found-family dynamics, all modeled on real geopolitical power structures. The Dauntless Gambit is my series, and I believe in it. Not because I think it is perfect, but because I know what it is trying to do, and I know it does that thing well. It is art, and it is MY art rendered in a way that fulfills the vision.
I can write the books, I am writing the books. I can build platforms (and I have built them). I can use every piece of technology available to me. I can summon the courage to try BookTok and Reddit and newsletters and whatever else the current wisdom says an independent author should be doing.
But I do not think anybody knows how to actually get books in front of people and get them reading.
Not a hundred thousand. Not even a thousand. A hundred. A drop in the bucket. Zero profit. If you told me right now that I had to get a hundred people to read something I wrote and genuinely like it, strangers, not friends, I would have no idea where to start.
And I say this as someone who is radically pro-AI, who has built AI tools, who believes in the technology with her whole chest. I am not a Luddite wringing her hands about the machines. I am a technologist who has looked at the discovery problem from every angle I can find and arrived at a conclusion that is aggressively uncomfortable:
This is not a technology problem. It is not a craft problem. It is not a marketing problem. It is something else. Something underneath all of those. Something that has to do with how human attention works, how trust gets built between a reader and an unknown author, how meaning travels from one person to another through channels that no platform has mapped and no algorithm can model.
And I think the honest thing to say is: I do not know what that something is. And I do not think AI will tell me.
The Fire Anyway
So I am left with this.
An algorithm that could predict the fire with ninety percent accuracy. That could catalog its composition, model its behavior, identify its signatures. That could not light one. A startup that tried to productize the prediction and produced only silence. A publishing ecosystem where the noise floor is rising so fast that even being seen is becoming the hard part, never mind being read, never mind being loved.
And a novelist who knows all of this, who has read the research, who has built the tools, who has studied the landscape with the same obsessive attention she brings to everything else in her life…
And writes the book anyway, with AI helping on every part of the overhead.
Not because I have solved discovery; I have not. Not because I expect the next platform or the next technology or the next clever hack to finally crack the code, because there is no code to crack.
That is the whole point. The Bestseller Code proved it: the code exists, and knowing it does not help.
You write the book because the book demanded to exist. Because the story was yours and no algorithm could have produced it and no analysis could have predicted it and no optimization could have improved upon the strange, specific, irreducible thing that happens when a particular human being sits down and writes a particular story at a particular moment in their life because she just has to find her 1000 true fans.
That is the asymmetry. Analysis goes one direction. Creation goes in another. They are not inverse operations. They are different things entirely.
And maybe that is not a flaw in the system. Maybe the fact that you cannot engineer resonance is precisely what makes resonance valuable. If you could manufacture it, it would not be worth manufacturing.
The smoke does not become wood. The ash does not become fire.
But someone, somewhere, is still striking the match. Not to start a bonfire. To light a candle. And then another. And then another. And then nine-hundred-and-ninety-eight more.
A thousand is enough. After all, if our reach doesn’t exceed our grasp, what’s a heaven for?
