This article opens a series sharing findings from our research into how and when ChatGPT recommends books across thousands of real reader requests.
The short version
- We examined the source lists for 27,702 ChatGPT answers to reader-written book recommendation requests. We traced citations in 25,136 answers to 93,736 distinct webpages.
- Publisher websites appeared in 75.6% of source lists. They also appeared in citations attached to 10,440 answers, or 41.5% of answers with traceable citations.
- Individual book pages were cited in 6,610 answers; publisher editorial pages in 3,101. Editorial pages were more likely to recur across questions: 43.5% were cited more than once, compared with 13.2% of book pages.
- The same websites recur, but most individual pages appeared only once. Across the 25,136 answers with traceable citations, 84.2% of distinct cited pages appeared for only one question.
- Knowing which websites appear helps you decide where to pitch. Knowing what a specific page says about a specific book helps you decide what to write about.
Somewhere right now, a reader is describing a book they haven't read yet.
Something devastating but hopeful. A mystery without the gore. A dual-timeline historical novel, like Fiona Davis but darker. An introduction to historical archaeology that doesn't assume a degree in it. A nonfiction audiobook as absorbing as Say Nothing.
They write these requests in public reader forums, in their own words rather than in the language of a catalog.
We analyzed ChatGPT’s answers to 27,702 reader requests. We examined their source lists and traced citations in 25,136 answers to 93,736 distinct webpages. Where we had captured the page text, we also examined what it said about the books.
Our goal was to understand how books become discoverable on AI platforms like ChatGPT - and where publishers should invest limited resources. Should they rewrite book descriptions, create reading guides, seek reviews, or join conversations on forums? Much of the existing GEO advice is conjecture or based on anecdotal evidence. We examined the actual sources appearing in AI answers at scale.
Publisher pages already appear in AI answers
Publisher websites already appear frequently in ChatGPT source lists. Across 27,702 ChatGPT answers, 20,943 included at least one publisher website in their source lists, which is 75.6%. Three quarters.
Publisher websites also appeared in the citations attached to the answers themselves. We identified publisher citations in 10,440 answers, 41.5% of the 25,136 answers with traceable citations. This extended well beyond Penguin Random House: 6,049 answers cited other publisher websites.
That presence extends beyond Penguin Random House. If we exclude the whole Penguin Random House family, 59.5% of answers still contain a publisher source in their source lists. Macmillan properties appeared in 28.7% of answers, Hachette in 21.9%, Simon & Schuster in 9.6%. (Note: don't add these numbers together, as a source list can contain websites from more than one publisher group.)
Publishers tend to assume that AI discovery is happening somewhere where they have no presence in a conversation between readers and a machine that happens without them. The data says the opposite. Publisher-controlled pages are among the most commonly identified sources in this entire environment, alongside retailers, reader communities and review sites.
We examined both the longer lists of webpages returned as sources and the citations attached to the answers. Source-list figures cover 27,702 answers. Citation figures cover the 25,136 answers whose citations we could connect to specific pages. The graphs and tables identify which measure they use. For background on how search results become context for an answer, see How books show up in ChatGPT and AI search.
Book pages and reading guides
Publishers show up in these answers through pages that work in distinct ways, and the split is instructive.
02 / PUBLISHER PAGE CITATIONS
Publisher page types cited in ChatGPT answers.
| Publisher page type | Answers citing this type | Share of 25,136 answers with traceable citations | Distinct cited pages |
|---|---|---|---|
| Individual book page | 6,610 | 9,162 | |
| Editorial pageSuch as a reading guide or themed list | 3,101 | 1,136 | |
| Academic content | 1,429 | 2,240 | |
| Catalog collection or series | 458 | 406 | |
| Author page | 91 | 88 |
A book page establishes what a book is: its subject, audience, edition, and reading experience. A themed reading guide can group several books around a reader’s need and explain how they differ.
Individual book pages were cited in more than twice as many answers as publisher editorial pages. Editorial pages, however, were more likely to appear across different questions: 43.5% of cited editorial pages were cited more than once, compared with 13.2% of individual book pages. Book pages provide coverage across a catalog; a useful guide can serve several related reader requests.
Editorial citations spanned 14 publisher websites. In 952 answers, they included editorial pages outside the Penguin Random House group.
The same websites appear often, but the same pages rarely do.
Here is where a publisher's instinct leads them somewhere unhelpful.
The most frequently cited websites were Reddit, penguinrandomhouse.com, Goodreads, Wikipedia, Book Riot and Google Books. Reddit was cited in 27.7% of answers with traceable citations. When we excluded citations to the original Reddit thread that supplied each question, other Reddit threads still appeared in 19.5%.
Look at a list like that and the plan writes itself. Find the sites at the top, get your books onto them.
Now look one level down.
Across the 25,136 answers with traceable citations, those citations point to 17,124 different websites and 93,736 individual pages. Of those pages, 78,969 appeared for exactly one question. That is 84.2%. The most repeated page turned up in 156 of the 25,136 answers, which is 0.6%. The typical answer cites four websites and five pages.
So the brands recur while the pages rotate underneath them. Learning that Book Riot ranks high tells you where a pitch might land. But it doesn't tell you anything about what the page would need to say, because the Book Riot page cited for one reader question is rarely the one cited for the next. There is no list of twenty pages to get onto.
There is a very long tail of specific pages, each answering a specific reader want.
The same website can hold a book description, a review, or an argument against the book
A web domain tells you where a web page lives, but it doesn't tell you about what is written on it, or who wrote it. Two examples from our data make this point especially clear.
A reader asked for a travel romance set in Japan as a gift. ChatGPT recommended a book about a travel blogger in Tokyo, with food, sightseeing, and a reunion with someone from her past - and cited the book’s Goodreads page. It’s tempting to assume reader reviews shaped the recommendation. But if you open the page: the matching details appear in the book's description, above the ratings. In this case, Goodreads supplied the book copy. The citation alone does not show that reader reviews drove the recommendation.
Second case. A reader asked for a story with a completely platonic relationship between a vampire and a human. The answer named Bloodsucking Fiends by Christopher Moore, noted that it is funny, and then explained that its central vampire-and-human relationship is romantic, so it would not suit. The source cited for that judgment was Simon & Schuster's own book page, where the description presents the romance plainly.
The publisher page did its job. It described the book accurately enough that a reader looking for something else could be told to look elsewhere. It also shows that publisher citations can include books being ruled out. A book can be cited because it does not meet the reader’s requirements.
When the page provides the evidence
In two more cases, the cited pages directly supported the answer.
A reader wanted an introduction to historical archaeology, with an interest in power, memory and preservation. The answer recommended Historical Archaeology in a Nutshell by Adrian Praetzellis and cited the Routledge book page. That page describes first-time students as the audience, states that the chapters assume little prior archaeology, and lists the thematic subjects: race, colonialism, identity, diaspora, religion. Ordinary publisher copy, carrying exactly the detail a reader needs to know whether a book meets their needs.
Another reader wanted an emotionally demanding nonfiction audiobook, having loved Say Nothing. The answer positioned Empire of Pain as a close match and cited a Pan Macmillan guide to nonfiction audiobooks. The guide calls the book the true story of the family behind America's opioid crisis, names the Sacklers and OxyContin, and ties the drug to the scale of the harm. The entry also emphasizes that the book is as gripping as a thriller. The page describes the book's subject and its reading experience in the same terms the reader used. The answer compares it with Say Nothing, which the cited guide does not mention.
Both pages contained specific information relevant to the reader’s request.
What this means for publishers
These findings raise practical questions: what information belongs on a book’s page, whether the existing description is enough, and when a new publisher page is worth creating. Before creating new content, check whether your existing book pages explain who each book is for, what readers can expect, and how it differs from similar titles.
Start from what readers ask for. Choose one part of your catalog and read the reviews its books receive on Amazon, Goodreads, and other sites. Look for what readers hoped to find, what they valued, and how they describe the reading experience. Then explore recommendation threads in book communities such as Reddit. Read both the requests and the replies: what is the reader looking for, which titles do others suggest, and why do they say those books fit? Use those details and the readers’ own language to make your book pages clearer about who each book is for and what readers can expect.
Read the individual AI answers. For each reader question you ask ChatGPT, record which books get referenced, how the answer describes them, which sources are cited, and what those pages say about each book.
Fix the information you control. Check that your book pages state the audience, the scope, the approach and what distinguishes the book from similar titles. Where several of your titles could suit the same reader, create a guide that explains the differences. In our data, individual book pages were cited across more answers, while editorial pages were more likely to recur across different questions.
The information behind the answers
In the examples above, the cited pages contain details that help explain whether a book fits the reader’s request. Publisher descriptions and guides already provide some of that information, including a book’s intended audience, subject matter and reading experience.
This all puts publishers in a stronger position than the fear around AI discovery suggests. The reader questions are visible, in public, in the readers' own language. The recorded sources and citations are visible too. And three quarters of the source lists we examined included a publisher website.
Publishers already have a substantial presence in the sources behind AI book recommendations. The opportunity is to make those pages more useful: clear about who each book is for, what readers can expect, and how it meets their particular needs.
Method. We analyzed the retained source lists for 27,702 ChatGPT answers to book recommendation requests posted publicly on Reddit. We traced citations in 25,136 of those answers to 93,736 distinct webpages. Each website or page is counted once per answer in its respective analysis. Captured text was available for 18,775 of the cited pages; the page-content examples draw on that text. Source categories describe the websites we could classify, rather than the authorship of every passage they contain. The requests reflect readers who ask for recommendations publicly. The study is an October 2026 snapshot of observed answers and sources; it did not test the effects of changing publisher content.