Summary: How to help your books get found by AI. Reflections after the Frankfurt Book Fair 2026 on the discoverability of books: how a title gets into an AI answer, what changes when the reader does not click, why a publisher's website and ONIX need to be sources of truth, and a diagnostic exercise with ten books from the catalog.
I came back from the Frankfurt Book Fair with my head full of ideas. I am still digesting the conversations, the talks and the meetings, trying to connect all of it to what we live with in the Brazilian publishing market. One question has taken up a good part of that thinking: the discoverability of books.
The word is awkward, but the question is simple. We publish, produce, distribute and promote. Even so, how does a reader find a book they do not even know exists?
Part of that discovery already happens in conversations with AI assistants. People use these chats to research topics, compare options and ask for recommendations, including what to read. That goes through ChatGPT, Gemini, Perplexity and the conversational experiences inside Google search itself. The reader can explain what they are looking for, add preferences and keep asking until they reach a suggestion that interests them.
This is the context in which GEO becomes relevant. The acronym came up in many conversations about artificial intelligence in Frankfurt, but the question it raises for a publisher is simple: how do I get these systems to find my book and understand when it might be a good recommendation?
I took that question to the LatAm Stage in the talk Metadata's new reader isn't human. It's AI. Before the reader even opens the cover, a system may already have consulted the book's data, compared its characteristics with those of other works and chosen to mention it.
That is why I keep insisting on care for the publisher's website and for the metadata it puts into circulation. The page presents the book to the public; the metadata carries that presentation to distributors, bookstores and the systems that help with discovery. The two need to tell the same story, with specific, reliable, up-to-date information.
How a book gets into an AI answer
Imagine someone asking: "I want a Brazilian novel set in the countryside, with complex family relationships, to discuss in a book club." The person named no title and no author. They described a wish. For a book to enter that conversation, the information available about it has to allow that connection.
That kind of presence is what GEO, short for Generative Engine Optimization, is about. When they consult the web, tools such as ChatGPT, Gemini and Perplexity, as well as Google's AI features, can gather information from several sources to build an answer, sometimes with links attached. Not every chat answer involves a web search. When that lookup does happen, though, the public content of the website and of other channels can help shape the answer. So it is worth telling apart whether the tool is searching for sources or answering without consulting any pages at that moment.
We already know SEO, search engine optimization, which helps a page get found on Google and other search engines. There is also AEO, aimed at direct answers. These approaches overlap, and SEO remains necessary. With GEO, another concern comes to the fore: how will the information about the book be understood and used inside an answer?
This is where editorial work makes a difference. A vague blurb, a generic classification and a nearly empty page offer little for that match. A description faithful to the work, with context and consistency across the metadata, makes it far easier to understand what that book offers.
Being cited, receiving a visit and selling a book are different outcomes. It is worth tracking each of them to understand what that presence is bringing to the catalog.
What changes when the reader does not need to click
For years we grew used to one route: Google showed the results, the person clicked and landed on a website. When the answer appears in the search itself, part of that journey no longer happens. For anyone who produces content, a hard question arises: how do you sustain the work if part of the audience no longer reaches the page?
A Pew Research Center study, published in July 2025 and based on the browsing of 900 adults in the United States, helps put the question in proportion. In searches with an AI summary, users clicked a traditional result in 8% of cases. Without the summary, the rate was 15%. The study portrays that group and that period, but it gives us a good reason to pay attention to the shift.
Google presents its AI features as new possibilities for discovery and has to fit advertising into that experience. Websites and advertisers are also trying to work out what happens to the paths of attention, purchase and revenue.
For us publishers, this raises a very concrete concern. A book may be presented before the reader ever reaches the publisher's or the bookstore's website. What information is being used in that presentation? Where did it come from? Does it represent the work well?
What a small bookstore can teach us
One of the presentations that struck me most in Frankfurt took place at the International Supply Chain Seminar run by EDItEUR, the organization responsible for ONIX and Thema. Robin Fang, of Fox & Fable Booksellers, spoke about using AI to detect demand for books.
An independent bookstore's account made concrete a discussion that sometimes feels distant. By making information available in a form the systems could access, it was winning space in the answers alongside large retailers. The question was how to become a useful source both for those who recommend and for those who buy.
Think of the knowledge a small publisher holds about a subject, an author or a community of readers. That knowledge has value, but it needs to appear on the pages and in the published data. If it stays only in editorial conversations, it never reaches the person looking for something to read.
The experiments discussed at the seminar, including work by Otterly.ai, reinforced that concern. Some tests found marked differences in how often complete pages were cited compared with uninformative ones, along with gains linked to information about awards and to simpler page addresses. These are results from specific situations, not predictions of what will happen with any given catalog. The lesson I take from them is the value of making available what we know about our books.
Limits also showed up in solutions presented as shortcuts. In one of the experiments cited, llms.txt, a file proposed for presenting website content to AI systems, received only 84 visits out of more than 62,000 bot requests in 90 days. That result helps put its priority in perspective.
Structured data, such as markup that uses the Schema.org vocabulary to identify book, author and ISBN, deserves technical attention, but its effect varies between systems. Likewise, pages mass-produced by AI behaved differently on Google and in other assistants. None of these solutions does away with the need for a reliable presentation of the catalog.
Before anything else, I would look at your website and ask a few basic questions. Does the page explain the book? Is the data correct? Is there specific information, or only adjectives that would fit any title?
Metadata is an editorial and sales decision
This is where we need to change the way we look at metadata. It has to stop being seen as a tedious, bureaucratic chore, done only to get a record released, and become part of editorial and sales decisions.
When we describe a book, we are answering important questions: who is my audience? What might spark their interest? In what situations does this reading make sense? What themes, places and experiences does the work address? What sets it apart from other titles in the catalog?
For a work of fiction, that may include the tone of the narrative, the conflicts and the reading experience. For a non-fiction book, it may be the question it helps clear up or the problem it teaches you to tackle. This is the kind of information that brings the record closer to the way a person asks for a recommendation: "I want a light read", "I need to understand this subject" or "I liked this author, what can I read next?"
Those answers help build the blurb, define the classification and guide promotion. They also give the bookseller something to recommend the work with, and give the systems information to match it to a search. Filling in the record becomes an opportunity to think about how that book will find its readers.
That calls for conversation between editorial, marketing, sales and whoever looks after the systems. Those who read and followed the work have a lot to contribute. So do those who talk to booksellers and readers. The team responsible for the record needs to be able to turn that knowledge into useful information and keep it up to date.
This is not about inventing an audience to widen reach, nor about promising what the book does not deliver. The strength of this work lies in representing the work well. Metadata that is complete but contradictory can confuse a person and a system alike.
The website needs to show what the publisher knows
The publisher's website should be a reliable reference on its own catalog. Readers, booksellers, journalists and automated systems need to find there information consistent with the metadata distributed to the other channels.
Imagine a page with only a cover, a price and two generic sentences. "A moving narrative that will surprise the reader." It may catch the eye, but it says little. What is the story about? Where does it take place? Why would anyone be interested in it?
There is no need to turn the blurb into a treatise or give away the ending. It is enough to give the presentation text the same care we give to the design or the proofreading. And to make sure the rest of the page helps the visitor get to know the book and choose it.
A few measures help a great deal:
- Have a page for each book, clearly identifying the editions and formats available. Include the blurb, the author, the translator where there is one, the ISBNs and any information relevant to the choice.
- Publish the information as accessible text. A data sheet hidden in an image or available only after some interaction may not be retrieved by some systems. The technical team needs to check what the bots can actually read.
- Preserve page addresses. If a platform change alters the URLs, set up redirects to keep the path to the works intact.
- Present awards, reviews and press mentions with context, dates and links that let the information be verified.
- State the accessibility features before the purchase. Structured navigation, image descriptions and other characteristics of the EPUB should be described accurately.
- Check consistency between the website, the distributor and the bookstores. The data for one and the same edition needs to match in all of those places.
It is also worth showing the relationships within the catalog: author pages linked to their works, series in reading order, titles related by subject and backlist books that are still reachable. These connections help whoever is browsing and make each work's context more explicit to the systems that consult the site.
Google's own guidance on AI features in search keeps the focus on traditional good practice, without requiring any special markup that would guarantee a place in the answers. There is consistent work to be done, starting with the quality and the availability of the information.
ONIX puts that knowledge into circulation
If the website is the publisher's public reference, ONIX helps carry the information along the book supply chain. I like to think of it as a shared record card, with agreed fields and agreed meanings. The publisher sends, the distributor receives and the bookstore interprets, without each company having to guess what the other meant.
Title, ISBN and price get attention because they are essential to the sale. But there is other data that helps explain why someone would be interested in the book, and it deserves a closer look.
We can start with the audience. "For elementary and middle school teachers who want to work on financial literacy in the classroom" says far more than just "education". In ONIX, that description can be recorded in AudienceDescription. To fill it in well, we need to think about the reader and about the content the work actually offers.
Presentation texts have their place too. TextContent gathers text types such as description, table of contents and supporting material. Other information has fields of its own, such as the biography attached to the author or another contributor. Putting each piece of content in the right place helps whoever receives the data to use it.
Reviews and third-party mentions can be recorded in CitedContent; awards, in Prize. If a book has won an award, that information needs to travel with it, with the correct prize name, year and result. It makes no sense to wait for a reprint to update the record.
Thema, also maintained by EDItEUR, organizes subjects and offers qualifiers that add context. For a novel set in Curitiba in the 1970s, place and period can help bring the work closer to a reading interest. It is worth using all the specificity the classification allows and the content justifies.
In day-to-day work with catalogs, I see that the problem often lies in the gap between what the team knows about the book and what reaches the system. Sometimes the form does not help. At other times, the information gets lost on the way to the store. Improving metadata means looking at that whole route.
There is also a decision about how that content gets used. Allowing a page to be found in searches and authorizing its use for AI training are different matters. There are platform-specific controls, and developments in ONIX for signaling restrictions on promotional texts and images, discussed by the International Publishers Association. It is a conversation to bring in the teams responsible for rights, systems and distribution, bearing in mind that a signal on its own does not guarantee that every bot will respect it.
A shared responsibility
A publisher can put together a careful record and discover that much of it never appears at retail. That is why every link in the chain has to take part. Management systems should make the relevant fields understandable; aggregators and distributors need to carry the data without losses; bookstores need to present it in a useful way.
Industry bodies can also contribute with guidance, examples and indicators. I see room for concrete questions: how many titles have a useful audience description? How many record their awards? How much of what the publisher sends actually reaches the reader?
I do not have those numbers for Brazil. I wish I did. They would help us set priorities and track improvements, bringing technology, sales and editorial together around a problem that matters to all of them.
Where I would start
My suggestion is to pick ten books from the catalog. They can be important backlist titles, works that remain relevant but get little attention, or books whose presentation is incomplete.
For each one, I would bring together the people who know the work and start with the audience: who is this book for, and what might lead that person to take an interest in it? Then I would revise the blurb, check the classification and gather awards and reviews with their sources. I would then compare the record with the page on the website and with what appears in the main stores. That exercise alone reveals a lot.
A checklist from Supadu, The Publisher AI Visibility Checklist, proposes a useful exercise for this diagnosis. First, search for the exact title and the author's name, checking whether the system identifies the right book. Then ask questions the way a reader who does not yet know the work would: "which book do you recommend on this subject for this audience?" or "what can I read after a book like this one?" I would repeat the exercise across different assistants, using web search when available.
It is not enough to note whether the title showed up. I would check whether the description is correct, which edition was suggested, where the information came from and whether the links lead to the right pages. I would record the questions, the answers and the date so the observation can be repeated after the adjustments. Answers vary, and this follow-up does not by itself prove the cause of a change, but it does make it possible to identify concrete gaps in how the catalog is presented.
I have written before about metadata as the digital soul of the book. I still see in this work a concrete opportunity for publishers. There is too much knowledge about the books inside the teams for the internet to receive only a cover and a generic paragraph.
That is the reflection I brought back from Frankfurt and want to keep developing. You, the publisher, know your authors, know how to choose what to publish and know how to build a catalog. That knowledge needs to circulate with the same quality. When someone looks for the kind of reading your book offers, the information that makes it findable needs to be available.
The publisher's website and metadata are an excellent place to start.
Further reading and references
- GEO: Generative Engine Optimization, academic paper presented at KDD 2024.
- AI features and your website, guidance from Google Search Central.
- Clicks in searches with and without AI summaries, Pew Research Center study, July 2025.
- Public GEO experiments, Otterly.ai.
- I Ran 30+ GEO Experiments So You Don't Have To, presentation by Thomas Peham.
- International Supply Chain Seminar by EDItEUR, the program referred to in this account of Frankfurt.
- The impact of AI search on advertising strategy, Brainlabs.
- Discussion on AI and website traffic, AI Weekly.
- Updates from EDItEUR, ONIX and Thema, International Publishers Association.
- Metadata as the “Digital Soul”, Fernando Tavares, PublishNews.
- The reader of your book's metadata is now a machine, Fernando Tavares, Booknando.
- GEO: from good metadata to a book that AI can find, Fernando Tavares, Booknando.
- AI Mode, Google.
- How does Perplexity work?, Perplexity.
Originally published on LinkedIn on 20 October 2026.
