Last updated: November 18, 2026
How we test Book2Pub books
How we test the books Book2Pub makes: real paid runs, every page read, KDP print checks on every file, and what failed along the way.
Book2Pub is only worth paying for if the books it makes would pass KDP and satisfy a reader. So before anything goes live we make real books with the real AI models, pay for them, read them, and fix what is wrong. This page says how, with the numbers. The complete test log, including every failure, is kept with the source code.
What we test
- Real paid runs, not demos. Between 15 and 18 November 2026 we ran 13 rounds of test books on the same AI models customers use: novels, non-fiction guides, a memoir from notes, picture books, coloring books, and every puzzle type. Together they cost about 25 US dollars in model fees, and every cost is logged.
- Every manuscript read in full. Each prose book was read start to finish by a separate reviewer (an AI model given only the brief), and we checked its findings against the text: plot holes, contradictions, repeated scenes, facts that change between chapters, invented details in non-fiction.
- Every print file checked against KDP's rules. Page size, bleed, margins for the page count, fonts embedded, images at 300 DPI or more, blank pages, cover size and spine width. A file that fails any check is never offered for download.
- Every puzzle solved by code. Each sudoku has exactly one solution, each maze exactly one path, each hidden word and number appears exactly once, and each crossword grid is read back square by square.
- The KDP calculators as the reference. Cover sizes for paperback and hardcover match KDP's own cover calculator to a thousandth of an inch on the samples we compared.
What failed, and what we changed
Testing is only useful if it finds problems. A few of the ones it found:
- Books came out shorter than ordered. A 60-page novel printed 36 pages because the type held more words per page than the plan assumed. We changed the typesetting and planning; the next 60-page novel printed 62 pages and a 100-page novel 94.
- Blank pages that KDP would reject. Short guides needed more than 10 blank pages at the end. Such files are now refused before export, and the book is refunded.
- Chapters that contradicted each other. When chapters were written side by side, names and facts drifted. Every chapter now reads the whole book written so far, and an outline check runs before writing.
- Crossword books that reused themes. A 120-page crossword book used only 28 themes. Theme names are now chosen first and written in parts; the next run had 1,075 different answers out of 1,175.
- Covers with made-up text. An image model painted a fake title onto a cover. Cover prompts now forbid writing, and the real title is typeset separately.
- Long novels that cost more than they sold for. Re-reading the book on every call was expensive; with prompt caching a 100-page novel now costs about 1.40 dollars in model fees.
What we do not claim
- AI-written books still need your read before you publish. We tell you what to check, and the listing kit tells you how to answer KDP's AI disclosure questions.
- We do not promise sales or rankings. A book that passes KDP's checks still needs a good niche, a good cover and a fair price.
- We have no customer reviews yet, so we show none.
When we test again
Every change to how books are written, laid out or checked gets a new round of real books before it ships. The dates and results are added to the log.