How books show up in ChatGPT and AI search

The practical opportunity for publishers is not to change a released model, but to improve the information search can retrieve and connect to a reader’s intent.

What I learned from building an AI assistant for wind turbine technicians

A few years ago, while working at BCG (Boston Consulting Group), I architected an AI assistant for technicians at a large renewable-energy company and led the development team that built it.

Our users fixed huge wind turbines worth several million dollars apiece. A technician could be standing at the base of a turbine, or high inside a swaying tower, trying to decode a fault code on a handheld device or describing a strange noise of unknown origin. To troubleshoot, the technicians previously had to access information that was scattered across multiple repositories: lengthy manufacturer’s manuals, service bulletins, maintenance work records from similar sites with the same fault, or even other experts’ recorded walkthroughs of problem resolution.

The AI tool we built had to pull knowledge from all of these sources, understand the correct turbine model, know when new manufacturer guidance superseded outdated manuals, understand site-specific environmental factors such as high dust levels or frequent storms, and then factor in the work that had already been undertaken on that specific wind turbine. The system couldn’t afford to return a merely plausible-sounding or hallucinated answer—it had to respond accurately in a high-stakes, high-pressure situation.

As we iterated over the tool with expert technicians—the people whose expertise we were attempting to scale—one source repeatedly performed well in our evaluations: the company’s internal engineering forums. Experienced engineers had spent years answering questions from other engineers, explaining in plain language how they had solved problems in the field. We found that these exchanges were the most effective material for helping our AI assistant answer the technicians’ questions correctly because of the specific insight they captured and their question-and-answer format.

Our system was judged on how accurately it could help a technician solve a real problem, and the quality of its answers was strongly tied to getting the right evidence to the underlying language model in the right format. We found that the more closely the underlying information aligned with the user’s needs—in both content and format—the better we could help them solve their problem. We also discovered that technicians asked questions in different ways, so we needed to normalize their questions into a form we could match to the underlying solutions.

Why this matters for books

Earlier, as CTO of Bookish, I led the AI work on book search and recommendations. When readers were trying to find a book, a subject or genre could establish a broad match, but we often needed more input about what they were looking for—either in search or to recommend a book interesting enough for them to commit to.

A user typing “self-help” was probably browsing, so they would receive books that were mostly high sellers. That was all our models could work from. But if they typed “books on emotional regulation and nervous system healing,” we would know to return very specific search results or recommendations on calming the body and rewiring fight-or-flight responses.

Similarly, language models are trained to interpret reader requests and follow instructions, and they do this significantly better than traditional search and recommendation systems. With a language model, readers can explain their preferences in ordinary language, qualify them, and describe what they disliked about an earlier recommendation. Readers iterate with models to refine what the model returns, and it’s this back-and-forth refinement that makes them extremely good at understanding what a reader is looking for. But matching that intent to books depends heavily on the information available to the model. If a book exists that is a perfect match to solve a problem or fit a fiction reader’s needs, it will be surfaced only if the model can match that intent with what it knows about the book. If an AI tool can’t match a book’s information to a reader’s intent, it will find the next-closest match or steer the reader to its best guess from its trained knowledge.^{1, 5}

For our technicians, the engineering forums helped because they described both the problem and the circumstances in which a solution worked. For readers, a review that describes a novel’s level of violence can serve a similar purpose: it tells the system whether the book suits a reader who has asked for nothing too graphic. That detail might be found in a publisher’s description, reviews, interviews, reading-group material, or social media posts. Understanding how readers find books and express their intent in AI tools and search—and knowing how to provide information about books that matches this intent—is the key to executing GEO effectively.

An AI model answers you from three sources

It helps to distinguish large language models (LLMs) from the products that provide the customer-facing user interfaces, such as ChatGPT. The model is the “intelligence” layer that generates the response to a message you send, while the surrounding software manages the conversation and gives the model access to tools. While ChatGPT and Google AI Mode don’t have identical implementations, they both combine language models with search capabilities to supplement the knowledge on which they were trained.^{2, 4}

An AI tool’s answer can draw on:

1. What the model learned during training.
2. What the system finds through live search or other tools.
3. What has been said or supplied in the conversation, including relevant past conversations when available.

The second and third sources contribute to the model’s context: information available to the model when it generates a response to the user. This can include search-result snippets, the current user conversation, a document the reader has uploaded, or details and memories the product has stored from earlier chats.^{5, 6}

1. What the model learned during training

A language model’s own knowledge comes from material available when it was trained. A model may have learned about an author, a book’s subject, or relationships between titles from book metadata, reviews, Reddit, book blogs, and other sources. It can “know” about a book through material discussing it, even if the book’s text itself was not included in training.⁷

What the model learned is fixed for that version of the model. A language model has no automatic visibility into metadata updates, newer editions, major awards, announced adaptations, or connections to current events. This also means that an inaccurate description or misleading account picked up from public discussion cannot be corrected simply by changing the source page. A provider may train a later model on newer material, but the publisher cannot control when that happens or how the book is represented in the training data.^{7, 8}

For a specific model version, its knowledge is a point-in-time snapshot. It can’t be updated directly, so the only way to provide updated or corrected information—outside what a reader writes or pastes in—is through live search.⁹

2. What the system finds through live search

When ChatGPT receives a message, it may answer directly or reason about the reader’s request before replying. In a reasoning mode, the model can consider what information it needs, decide to search the web, review the results, and search again before it answers.¹⁰

Not every ChatGPT answer involves a web search, but search is becoming more prominent in AI-tool responses. Google AI Mode is slightly different from ChatGPT: it is a search experience first and uses information from Google Search to answer detailed questions with a language model.^{2, 4}

Although we call this “live search,” the tools generally search a stored index of webpages that the company has previously crawled. A newly published page, metadata update, or blog post can take time to become available through an AI tool’s search. Some tools can visit a live page during a session when given or discovering its URL, but ordinary search discovery depends on the page being crawled and indexed.^{11, 3, 12}

When a web search does happen in an AI tool, the system can retrieve information about a new book release, the latest author interview, content about a book distributed as part of a publicity campaign, or connections between a backlist book and something in the news. This information can be much more current than what the model learned during training. The retrieved information then augments the context available to the model, which then generates its response to the reader. This pattern is called retrieval-augmented generation, or RAG, and we used it extensively in the wind-turbine assistant build.⁹

A new book can therefore appear in an answer even when it was not part of the model’s training. The system needs to find enough useful information about it to make that possible, provided the relevant pages have been crawled. It doesn’t have to wait for the next model release to return up-to-date answers to readers.⁹

One conversation turn can lead to several searches

Under the hood, handling a reader’s book request during a conversation can involve more than simply copying the reader’s words into an underlying search system. Suppose a reader asks:

I need a practical book about caring for a parent with dementia, written by someone with clinical experience, not a memoir. Is there anything recent that covers technology we could use at home?

The system might search for caregiving guides, investigate the authors’ qualifications, and look for recent editions or reviews discussing home-care technology.

Google describes this as query fan-out: issuing related searches across subtopics and sources. OpenAI also says ChatGPT Search can rewrite a prompt into multiple targeted searches and run further searches after reviewing the first results. The user’s phrasing may not be the best search phrasing for finding all the information needed to answer the question properly. In our wind-turbine AI tool, we built a similar capability that rewrote variations of a technician’s question and intent by matching their phrasing with how the problem might more commonly be phrased. This approach dramatically improved the system’s ability to surface the right context and provide the correct answer.^{4, 2}

The publisher’s metadata does not need to guess and repeat a reader’s request word for word, as different sources may supply different parts of the answer. A product page might explain the book’s scope and edition, an author biography might establish clinical qualifications, and a review might explain how practical the advice is or how the book differs from a memoir. Repeating “practical dementia guide” in the metadata does not tell the system whether the author has clinical experience or the book covers technology for use at home.

Here, all three information sources can contribute: trained knowledge helps the model interpret the subject, search brings in current books and evidence about their suitability, and the conversation supplies the reader’s intent and focus.

For fiction, a similar process might seek information about tone, romance, pace, or content. In these cases, a review, Reddit post recommending books, or reading-group guide may contain the detail that matters to the reader even when the product description doesn’t.

3. What has been said in the conversation

Consider a reader who begins with:

I want a historical mystery set somewhere cold.

Then adds:

Make the detective an older woman.

And finally:

Nothing too graphic.

The last message makes little sense by itself. To answer the user effectively, the system needs to understand what was said earlier in the conversation, much as a person uses short-term memory. If I walked up to you and simply said “Nothing too graphic,” you might wonder what was wrong with me.

AI tools supply relevant conversation history as context, so each new turn can change which books are surfaced in an answer. In the example above, a title that matched the setting and detective may no longer be suitable once the reader adds the preference for a low level of violence.⁵

The same applies to nonfiction: the books returned early in a conversation may change if a reader then refines their request to ask for a beginner’s guide, primary research, a clinician’s perspective, or a recent edition.

Readers can express these requirements in conventional searches too. A conversation lets readers add detail gradually without restating the whole request, so the relevant need may be spread across several messages rather than contained in one keyword or question.

As with people, AI tools such as ChatGPT can also have long-term memory outside the scope of an active, in-the-moment conversation. ChatGPT can retain information from earlier conversations and bring it into the context of a later one. It might know that a reader prefers hard science fiction over science fantasy and adjust its suggestions without being told again. This information is supplied as context rather than learned by retraining the model for that person.⁶

Discovery does not always begin with a book request

A user doesn’t have to ask “What should I read?” for a book to make it into an AI tool’s answer.

A reader exploring how to manage a growing business might be introduced to a book’s framework, or a discussion of a historical event could lead to competing accounts of the event. A question about caring for a relative might lead to a guide for further reading. Similarly, a novel can enter conversations about adaptations, historical periods, social questions, school courses, or book-club choices. An author interview, teaching resource, essay, or detailed review may explain a connection that a catalog description leaves unstated.

An answer could use an idea discussed in an article about a book yet cite only the article and not identify the book. Or it could name the book without linking to a page where the reader can learn more, borrow, or buy. Publishers need to check whether books in AI-tool responses are named and linked, not just whether pages discussing them have been cited.

What publishers can influence

Publishers can’t directly edit a released model’s trained knowledge, but they can improve the information that current searches pull into an answer. This includes what is published about a book—either directly or by other outlets or social media—what the book covers, who it is for, what distinguishes it, and, where relevant, why its author is qualified to write it. Reviews, interviews, and other independent coverage can also provide information for the AI tool to retrieve, including details missing from the publisher’s own content and metadata.

Google explicitly says that its AI search features do not require special AI text files or special structured-data markup.⁴

Short-term visibility and longer-term preparation

Publishers can test whether changes to descriptions, author pages, and other material make suitable books appear more often in current search-backed answers. The descriptions, reviews, and interviews available today may also contribute to what future models learn.⁷

This work doesn’t need to wait for a new edition or publicity campaign. A backlist title may already answer a question readers are asking, but its metadata or other published content about the book might not connect with related reader intent and questions. Publishers can address these omissions and test whether the book appears in conversations where the gaps exist.

A decade ago, when publishers began adding metadata keywords, relevant books could become visible in searches where they had previously been absent. There was a window of advantage for early movers to have their books indexed with keywords before the competition caught up. A similar window may exist now with GEO. Publishers who can identify reader intent for their books and ensure that relevant content matching that intent is indexed and surfaced within AI tools may gain an early, outsized advantage.

What GEO visibility is worth

AI-based recommendations or book mentions are often trusted more than traditional search and recommendation engines for book discovery because they are so effective at understanding a reader’s intent. A surfaced book might lead to a retailer or library rather than the publisher’s own website, so publisher-site traffic metrics alone won’t capture the value of book discovery through these tools. A mention is not a click, and a click is not a sale. But a book that doesn’t appear when it has a right to win within that conversation has lost the opportunity to convert to a sale.

Before deciding what content to rewrite for GEO, first establish what information the AI tool already has about the book and what it needs to answer the reader’s question.

Kadaxis will be releasing a GEO visibility tool in the coming months. Please reach out through our contact form if you are interested in becoming an early user.

The next article, Where books disappear from ChatGPT and AI search, examines the different reasons a relevant book may be missing: access, discovery, weak evidence, or failed attribution.

---

About the author

Chris Sim is the founder of Kadaxis. He previously served as CTO of Bookish, where he led AI work on book search and recommendations, and later, while at BCG, architected an AI assistant for technicians at a large renewable-energy company and led the development team that built it.

Sources

1. OpenAI, “Aligning language models to follow instructions.” Research on training models to follow user instructions using human feedback.
2. OpenAI, “Searching the web with ChatGPT.” Covers search use, targeted query rewriting, follow-up searches, and the absence of guaranteed placement.
3. OpenAI, “Web search.” Documents tool-based retrieval, reasoning-driven search, and opening pages while researching a response.
4. Google Search Central, “AI features and your website.” Covers AI Mode, query fan-out, search eligibility, and technical requirements.
5. OpenAI, “Conversation state.” Explains supplying conversation history and managing context across turns.
6. OpenAI, “Memory FAQ.” Explains optional use of retained context to personalize later conversations.
7. OpenAI, “How ChatGPT and our foundation models are developed.” Describes training sources, learned parameters, and data selection.
8. OpenAI, model documentation, illustrated by GPT-5.2. Model pages state a knowledge cutoff; this is not a title-by-title account of training coverage.
9. Patrick Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” The foundational paper describing generation using trained knowledge together with retrieved information.
10. OpenAI, “Reasoning models.” Describes reasoning within a model and how it supports planning and tool use.
11. Google Search Central, “In-depth guide to how Google Search works.” Explains crawling, indexing, and serving search results.
12. OpenAI, “Overview of OpenAI Crawlers.” Distinguishes search and training crawlers and their independent access controls.