This is the second article in a series sharing findings from our research into how and when ChatGPT recommends books across thousands of real reader requests. The first is Where AI book recommendations come from.
Key findings
- Publishers are being told that AI discovery rewards question-and-answer formatting, so book pages should be rewritten as FAQs. We tested that against 27,702 ChatGPT answers to real reader requests and the publisher pages those answers cited.
- Fewer than one publisher page in a hundred had an FAQ that matched what ChatGPT wrote.
- ChatGPT reads a book page much the way a reader does, then writes an explanation of why a book suits the person asking. Providing the right answer to a reader’s question is much more important than writing it in question-and-answer format.
Somebody in your organization has probably forwarded the advice by now: to get picked up by AI, rewrite your book pages as questions and answers, add FAQ schema, mark up the headings. Readers ask questions, so your pages should be written as questions with answers.
It’s generally much easier to put existing content into another format than it is to rewrite and improve the underlying content.
We wanted to test the common FAQ advice to see if ChatGPT actually pulls book-related answers from FAQ sources. We put thousands of reader requests to ChatGPT, kept the 27,702 answers that came back complete, collected the publisher pages those answers cited, and compared what ChatGPT wrote against what each page said.
Two publisher pages in a hundred had an FAQ
Across every publisher page ChatGPT cited (book pages, reading guides, subject guides, catalogs and blog posts), 2% had a question-and-answer section. Book pages had the fewest, at under one in two hundred. Most of the cited pages were ordinary book descriptions, themed lists and catalog copy without an FAQ section.
Where there was an FAQ, the answer was almost always from elsewhere in the cited text
Counting FAQs tells us what exists on a page, but not which text ChatGPT drew its answer from. So we compared the wording on each cited FAQ page with the answers.
For every answer paragraph carrying a citation, we took the cited page, separated the FAQ text from the rest of the page, and measured where the distinct content words matched what ChatGPT had written.
Across 7,382 comparisons, the answer matched text outside the FAQ content 72% of the time. The FAQ was matched 9% of the time. The rest tied.
Excluding ties, the closer match was outside the FAQ about eight times out of nine.
FAQs were the source of less than 1% of ChatGPT answers
Two publisher pages in a hundred carried an FAQ. On the pages that did, the FAQ was usually not the passage the answer matched.
Multiply those through and fewer than one publisher page in a hundred had an FAQ that matched a ChatGPT answer.
FAQs were a rounding error in what ChatGPT drew from publisher sites, while ordinary descriptions and lists drove almost all answers.
| Page type | With an FAQ | Cited pages |
|---|---|---|
| Book pages | 5 | 1,098 |
| Reading guides | 20 | 444 |
| Subject guides | 6 | 21 |
| Article and blog pages | 3 | 102 |
| Catalogs | 1 | 53 |
| All cited publisher pages | 35 | 1,718 |
Some of those FAQs were about shipping
The publisher pages that did contain an FAQ varied in content.
Bedford Square’s page for La Maison answers questions about the book, its author and who would enjoy it, with description-grade copy inside the FAQ. Hurst does the same for The Truth About Empire, covering the book’s contents. Learning Pathways Press answers questions about reading level and suitability for English learners on a graded-reader Treasure Island.
Other FAQs handled ordering instead: one answered lecturers’ questions about using the book with a learning management system (LMS), and another told direct-sale customers when their order would ship. Both of those FAQs do what their pages need. A lecturer ordering class sets wants the LMS answer, and a direct-sale customer wants to know when the book ships. Neither one describes the book, and both use the same question-and-answer format as Bedford Square’s. Knowing a page has an FAQ doesn’t tell you whether it describes the book or handles ordering.
What ChatGPT is doing when it answers
We read all 27,702 answers and measured how much explanation followed each book ChatGPT named. The typical answer referenced nine books and wrote about 56 words for each book.
Those 56 words described the book to a particular reader. Tone and mood wording appeared in 29.5% of them, audience and fit wording in 26.2%, then length, comparison titles, content cautions, pace and assumed background in that order.
ChatGPT was writing an explanation of why each book suits the person who asked. The facts that explanation needs are the audience, the assumed knowledge, the scope, the tone and the honest limits. Those facts are either on the publisher’s page or they aren’t. Reformatting existing content as FAQs won’t add value unless it addresses reader questions.
ChatGPT is standing in for a reader
The FAQ advice assumes ChatGPT works like a search crawler: lay the page out the approved way and get picked up. ChatGPT reads a page much closer to the way a reader does, then writes an explanation for the person who asked.
This puts answer engine optimization on the same ground as the rest of discoverability. Keyword work goes wrong when publishers guess at words rather than starting from what readers search for, and jacket copy goes wrong when it’s written to sound appealing rather than to tell someone whether the book will suit them. Across all three, the best results start from understanding what readers are asking for. Write the information that answers the right questions, then decide how to present it.
Reformatting a page that already contains the information ChatGPT can use to answer a reader’s question doesn’t give it new material to work with, and question headings won’t, in and of themselves, supply information the page is missing.
What to do instead
Our earlier study of where AI book recommendations come from covers the diagnostic work in full: collecting the questions readers ask about your subject areas, recording which books get referenced and which pages get cited, then reading those pages. For the format question specifically:
Take one coherent part of your catalog and check whether its pages tell a reader what the book covers, who it’s pitched at, what reading it is like, what it assumes they already know, and why it differs from the four other books they’re choosing between. Where that information is missing or buried, add it. Take it from the book, from editorial knowledge, from existing metadata or from properly attributed reviews.
If readers keep asking a question your pages don’t answer, add an FAQ that answers it: whether a book assumes prior knowledge, whether it’s a standalone, which volume comes first, how the editions differ, whether there’s an audiobook. Those are real recurring questions and a short FAQ handles them well.
The publisher pages ChatGPT drew on were the ones with something specific to say about the book.
Method. We analyzed 27,702 complete ChatGPT answers to book recommendation requests posted publicly by readers, and traced the citations in those answers to the pages behind them. Page findings describe the 1,718 cited publisher pages whose text we collected, so a page with no FAQ in our collection may have added one since. A working FAQ means readable questions with answers following them, which we count separately from schema markup and footer links. The wording comparison measures shared content words against the FAQ and against the rest of the page, using equal window sizes on both sides. The combined figure multiplies the share of cited publisher pages carrying an FAQ by the share of comparisons where the FAQ matched; we computed it on publisher pages alone, on all cited pages, and as a direct page count, and it stayed under 1% each time. The study describes answers and pages observed in October 2026.