AI-made books on Amazon

There's a video promising €2,300 a month from publishing children's stories made with AI on Amazon. I tried it exactly as shown, with the usual rules: everything gets told and the sales get shown, even if they're zero. The experiment is already over: I did it from 12 January to 8 March 2026, eight weeks, and now I'm telling it in four chapters. This is the first: the promise, the niche, the rules and the start.

I found the video over the Christmas holidays of 2025, with hundreds of thousands of views. Anyone who has spent ten minutes searching for how to make money with AI knows the format: a smiling person shows a screenshot of their sales dashboard, says they can't draw or write and that they publish children's stories made with the machine. "€2,300 a month," they say. "And you can too."

I'm not linking to it, because there are dozens just like it and I'm not interested in the presenter, only in the pattern. I admit the promise has something tempting about it. A children's story is short, the tools to write and illustrate it exist, and Amazon lets anyone publish. What could go wrong?

Once you strip away the background music, the promise has four pieces, and each one can be checked:

And there are things the video doesn't say. It doesn't say how many books are behind that figure, or how much is spent on advertising, or whether that €2,300 is sales or profit, or whether the real business is the book or the course advertised at the end. Income without subtracting the costs isn't income, it's a headline.

Amazon KDP (Kindle Direct Publishing) is Amazon's platform that lets anyone self-publish their books, as ebooks and as print-on-demand paperbacks. You don't need a publisher or an agent: you upload the book, fill in the listing and within a few days it's on sale.

Does it accept AI-made books on Amazon KDP? Yes, but you're required to disclose it when publishing. If the text, the illustrations, the cover or the translation were generated by AI, even if you touched them up afterwards, you have to tick it on the publishing form (checked in September 2026).

The nuance that confuses people most is the difference between "generated" and "assisted" content. If the machine produced the first draft or the first image, it counts as generated and has to be disclosed. If you wrote the book and the tool only helped you correct or polish it, it's assisted and doesn't need disclosing (checked in September 2026). That disclosure isn't shown on the book's listing, but skipping it can end with the book taken down or the account closed.

In my case there was no doubt: the stories were generated and I disclosed it. The whole point of the challenge was to repeat what the video promises, and the video says nothing about writing the story yourself.

The first week, from 12 to 16 January, was for research, about nine hours spread over five afternoons. Here AI did have a clear role.

I used Claude to organise what I was finding: I gave it the listings of about 60 stories from the bestseller lists in the children's category, with titles, descriptions, prices and number of reviews, and asked it to group themes, ages and formats. And with NotebookLM, which I already use for this kind of thing, as I explain in What I really use NotebookLM for, I cross-checked what eight videos and articles said about the same business to see where they agreed and where they contradicted each other.

They agreed on everything easy: the story writes itself, the illustrations come out on their own, it's uploaded in no time. And they contradicted each other on the only thing that matters: sales. One talked about "one book that works pays for the rest", another about "publish twenty so that one takes off", and none of them showed a breakdown of expenses.

Then, by hand, what the machine can't do for me: open the sample books. I saw dozens of stories with almost identical covers, the same soft watercolour style, big-eyed animals and interchangeable titles. And many of those published only a few weeks earlier had zero reviews and a sales rank that, according to Claude and the estimation tools I looked at, amounts to selling little or nothing. A word of caution about that last point: Amazon doesn't publish sales, and every estimate based on rank is just that, an estimate.

The conclusion of the research already hinted at the ending: publishing a machine-made book is easy, and selling AI-made books is another story. Precisely because anyone can generate a story in an afternoon, the shop is full of similar titles.

With the research in front of me, I ruled out anything generic ("stories for children") and went for something narrower. I settled on bedtime stories for children aged three to five about the first times that are scary: sleeping away from home, the first day at school, the first visit to the doctor.

Why that one? Because in the listings I analysed, stories about specific fears had more reviews than those on general themes, and because they're books bought by a parent with a problem, not by someone browsing. And because in my family there's a four-year-old willing to act as test reader in exchange for having things read to him. That was, without exaggeration, the best tool in the whole experiment.

If you don't define success before starting, in the end everything looks like a success or a failure depending on the day. So I wrote it down on 14 January:

• Goal: sell at least 100 copies across all the books and cover the expenses.
• Timeframe: until 8 March 2026: eight weeks from the first day of research.
• Number of books: three stories, not one, because the promise talks about a catalogue.
• Spending cap: €200 between tools, printed proofs and advertising.

Eight weeks is short for a book, I know. But the promise talks about income per month, not per year, and I wanted to see what happens in the first month, which is when most people give up.

And the protocol, also fixed before starting so I wouldn't change it along the way:

• Same process for all three books: the same tools, the same kind of instructions and the same layout template, so that the differences came from the book and not from the method.
• Hours logged as I went: in a spreadsheet, with the start and end time of each session and what was done in it, not reconstructed from memory at the end.
• Sales taken from the KDP dashboard: every Monday, copying the previous week's figures.
• Every expense with its receipt, even if it's two euros for a printed proof.
• No buying reviews and no cheating to make the numbers look good.

The video goes from "open your account" to "publish" in a single cut. In real life, opening the KDP account took me an afternoon, that of 15 January.

You have to fill in the payment details, where Amazon will pay the royalties, and a tax interview, a questionnaire about your residence for tax purposes. I got stuck on it for a while because I didn't know whether to answer as a private individual or as self-employed, and Claude helped me understand what I was being asked, although I checked the decision with my accountant, which is what they're there for. It's one of those things best sorted out before selling the first copy, not after.

That same week I did two more things I was grateful for later:

• I read KDP's help pages on paperbacks and generated content, with Claude alongside to ask about whatever I didn't understand. It's not beach reading, but it saved me a rejection or two.
• I asked Claude for ten story ideas within the niche and went for a hedgehog who doesn't dare sleep away from home. The family's four-year-old approved it with an enthusiasm I've never seen him show for anything I write.

On Friday 16 January I closed the first week with nine hours of research, two of paperwork and zero euros spent. I had the niche, the rules, the account and a hedgehog. For the weekend I wrote down a sentence I'm now embarrassed to reread: "Saturday, book finished".

The only way I know to find out what's behind these videos is to repeat their steps one by one and write everything down, especially what goes wrong. It's the same method I used later, in the summer, with another promise in Can I build a business that pays me a salary?. That series is already published, and its conclusion is quite similar to the one coming here.

The rules of the series are the blog's rules:

• All the numbers get published: hours, euros spent, copies sold, reviews. Even if they're zero or negative. And broken down: royalties (what Amazon pays after deducting printing), advertising, what's left and the time invested.
• What the machine did and what I did are kept separate in each chapter. Without that there's no way of knowing which part of the promise is true.
• A series once started gets finished, however it turns out.

I had it flagged as something that may no longer be worth it: when everyone is publishing generated stories, the gap that used to exist narrows. In the next chapters you'll see whether that was the case, with the numbers in front of you.