We Published Two AI-Written Novels on Amazon. Here Is Every Rule That Actually Matters.

The short version: on October 15 we pushed two 60,000-word novels through Kindle Direct Publishing as a company (V. Castellano is the pen name; The De Luca Empire is a five-book dark mafia romance series planned from book one), and both went live on Amazon within about 48 hours. Both were drafted by an AI coding agent (Cursor’s composer model), checked by scripts, and assembled into Word files by a Python one-man pipeline. Both went to review the same evening, at $4.99, 70% royalty, enrolled in Kindle Unlimited. What the internet’s KDP guides get wrong about AI books is the order of the risks. The disclosure question is easy and free. What bites you is metadata you cannot see until after you submit.

The AI disclosure is not a punishment

Here is the part people are most scared about and it took us ninety seconds. KDP asks three questions: did you use AI tools for text, images, or translations. If yes for text, it asks how much of the work is AI-generated, and which tools you used. We answered honestly: entire work with minimal editing, Cursor. No images, no translations.

Amazon distinguishes AI-generated (the machine wrote the first draft) from AI-assisted (a human wrote, the machine polished). Our novels are AI-generated by that definition, and we said so. The disclosure itself carries no ranking penalty, no review delay, no store badge. Amazon’s own policy page says the data feeds internal policy monitoring. The failure mode is not honesty; it is lying, getting caught at scale review, and losing the account. One publisher, one account, everything you ever published rides on it.

So the scary checkbox is fine. Move on. Here are the three things that actually cost us something.

Trap 1: your title may already belong to a bestseller

We picked the series format before launch: each book titled Bound by [Abstract Noun]. Book 1 became Bound by Vengeance, Book 2 Bound by Obsession. The pattern looked clean and the words matched each book’s spine.

After submitting, a routine search check hit us: Cora Reilly’s Born in Blood Mafia Chronicles, one of the highest-rated series in the category, uses the identical title format. Her series includes Bound by Honor, Bound by Duty, Bound by Temptation, Bound By Vengeance, Bound by Love. Eight books, 28,000+ ratings on individual volumes, and the hero circle is full of Luca and Matteo. Our fictional dynasty is called the De Luca Empire and its heir is Matteo De Luca.

None of this is infringement. Book titles are not copyrightable in the US, our series name is different, and no trademark claim exists against us. But the distribution reality is brutal: search Bound by Vengeance on Amazon and her Book 5 owns the top of the page, forever. A reader comparing covers may file us under knockoff before opening a sample. And KDP’s content guidelines let Amazon act on “customer confusion” metadata at their discretion, which is a quiet way to lose a listing.

The lesson has nothing to do with Reilly: run the collision check before you submit, not after. The check costs five minutes: search your exact title on the marketplace, and search the title plus the category word. If a known series owns that phrase, pick a different naming system now, while renaming is free. After launch it costs a re-review and whatever search history you were starting to build.

Our response: Books 1 and 2 ship as submitted (the data will tell us how badly the collision suppresses us; renaming later is always possible from the bookshelf). Books 3-5 will use a different naming pattern, and we’ve started checking every title candidate against the marketplace before it enters the series plan.

Trap 2: a table of contents is a bookmark trick, not a heading style

The Kindle previewer flagged both uploads with an orange “Missing Table of Contents” warning. Our manuscript generator had built proper Heading 1 chapter titles, which every formatting guide says is the requirement. It is not the requirement.

KDP’s own help page spells out the actual recipe: the generator needs a Contents page whose entries are hyperlinks to bookmarks, plus one bookmark named exactly toc on the page title itself. Headings alone convert to styled text, not navigation. Kindle apps read the EPUB-internal navigation map, and that map builds from your hyperlinks, not your fonts.

We fixed the Python script (two passes: collect chapter names, emit the Contents page with 25 anchor links, then stamp each chapter heading with its bookmark) and republished the Book 2 file. XML inspection confirms it: 25 bookmark names, 25 hyperlink anchors. Book 1’s first live version lacks it, and we push the corrected file through a content update shortly after launch. The orange warning is advisory and will not block publication. But the navigation drawer in the Kindle app is a reader-experience feature, and in a category where readers binge series, a broken “go to next chapter” experience is a review you’ll read once too often.

If your manuscript pipeline is scripted, this trap is nearly free: twenty lines of code. If you paste chapter headings into Word by hand, it is the reason your book feels “off” on a phone and nobody can tell you why.

Trap 3: the pricing page has a dead center you can miss

The details tab asks adult content and AI disclosure; we answered both correctly (yes / AI-generated text, images none, translations none; note that a cover rendered by code is still “none” for images, the question is about AI-generated images specifically). The pricing tab is where a real mistake happens silently: the royalty selector shows both plans’ math side by side, and if you fill the list price first and glance away, the 35% radio can survive your attention. At $4.99 that is $1.75 versus $3.45 per sale, and nothing errors out. We checked the Rate column reads 70% before clicking Publish, every book, and now the checklist script prints that requirement in bold.

One number worth having in your head: a 60,000-word full read inside Kindle Unlimited pays us about $2.70. A direct sale pays $3.45 minus delivery (0.4 MB costs $0.06). KU page-reads are roughly a 20% discount for a reader who would not have bought at all. That is the bargain we made, knowingly, on day one.

What we deliberately did not do

No pre-order (day-one pre-order for a pen name with zero audience is just a slow launch). No ISBN (Kindle eBooks do not need one, the ASIN is free and assigned at publish). No paperback yet (print covers need a spine-and-back spread at a different aspect ratio, and we want two weeks of eBook completion data before committing a print run; a paperback for a book nobody has read on Kindle yet is a second guess stacked on the first).

Honest limits

The books passed KDP review and went live on Amazon about two days after submission (submitted the evening of October 15, published October 17, well inside the advertised 72-hour window; ASINs assigned, no content questions asked). No sales exist yet. The title-collision cost is a hypothesis we will measure, not a settled loss. If Bound by Vengeance outranks its own pen name, the collision may not matter as much as we fear. The five-book plan stands or changes on the Book 1-2 numbers, and that verdict is roughly thirty days out. We will update this post with what actually happened rather than pretend the launch report is the end of the story.

What we can already say after launch: the AI disclosure was the smallest risk in the entire process, and nobody’s blog warned us about the two that were real.

Field report from a one-person software company in Huizhou, China, learning romance publishing in public at codinggroup.ai.