Readers are beginning to describe the books they want instead of searching for them. That small behavioral shift could have big consequences for publishers.
Finding a book used to involve quite a lot of wandering around the internet.
Imagine you are looking for a novel to take on holiday. You might search Google for “best historical novels set in Japan,” open a few listicles, check Goodreads, read some reviews, visit Amazon, and gradually narrow the options down.
Now you can simply ask ChatGPT: “Recommend me a historical novel set in Japan. I want something literary but fast-moving, preferably under 400 pages, and nothing too depressing.”
A few seconds later, you have a shortlist. No ten blue links, no eight browser tabs, and no need to work out which keywords will produce the right results.
This does not mean Google is about to disappear. But it does suggest that something more subtle is changing: the long, messy journey between “I want something to read” and “I know which book I want” is increasingly happening inside an AI conversation.
For publishers, that could matter enormously. For years, digital discoverability has largely been about ranking well enough to be found. Conversational AI introduces a different problem: what happens when there is no page of search results to rank on, and the AI simply names three books?
From searching for books to describing them
Traditional search works particularly well when you already know roughly what you are looking for. You type “books about behavioral economics,” “best new fantasy novels,” or “biography of Steve Jobs,” and Google returns pages that appear relevant. You then do the work of comparing them.
Generative AI changes that relationship because readers no longer have to reduce what they want to a handful of search terms.
Someone can ask for “a science fiction novel with the political complexity of Dune, published within the last five years, under 400 pages, and with less emphasis on warfare.” The system can consider those preferences together and return a shortlist.
Search finds you places to look. Generative systems try to answer the question outright. Research into AI-driven discovery suggests this move towards conversational, constraint-based queries is one of the most important differences between conventional search and generative recommendation.
The change also expands the language readers can use to discover books. They can search through moods, situations, and combinations that do not fit neatly into conventional keywords:
“Something optimistic about climate change.”
“A management book without Silicon Valley examples.”
“A thriller that is tense but not violent.”
“A history book for someone who normally hates history books.”
Google can already handle increasingly complex searches, and its own products are becoming more conversational. The important change is therefore not that one technology can understand natural language while another cannot. It is that conversational systems make this style of discovery central to the experience.
Instead of asking readers to figure out the right search, they invite readers to explain what they actually want.
Google spreads attention. AI concentrates it.
That difference becomes even more significant when we consider how many books a reader actually sees.
Google has traditionally distributed attention across a fairly large surface. Even if a particular book does not rank first, it might appear fifth or eighth. It might be discovered through a review, a Reddit discussion, a bookseller page, an interview, or a “best books of the year” article.
A generative answer compresses that landscape.
Ask for recommendations, and you may receive three books. Perhaps five.
The catalogue remains enormous, but the recommendation surface becomes much smaller. That is convenient for readers. It could be unforgiving for publishers.
The internet gave publishing something close to an infinite shelf. Conversational AI may give the reader a shelf with three spaces.
This concentration is already visible in the way generative systems respond to highly specific prompts: rather than presenting a broad field of possible sources, they tend to synthesize information and return a handful of named recommendations.
In Google, being seventh still means being visible.
Inside an AI answer, being seventh may mean not appearing at all.
But Google is not going away
That is why “Will ChatGPT replace Google?” is a useful question, but probably the wrong way to think about the longer-term shift.
Google itself is incorporating generative answers. Retailers are building conversational assistants into their own platforms. Amazon’s Rufus, for example, interprets natural-language shopping questions using information from Amazon’s catalog, product pages, reviews, and other marketplace data.
The boundaries between search engine, chatbot, and retailer are already beginning to blur.
AI has not replaced the checkout either. Readers may use it to work out what they want, then head to Amazon, Bookshop.org, a publisher’s own site, or another retailer to buy it, much as they always have. The strongest current role for conversational AI appears to be earlier in the journey, where it helps people compare options and narrow down choices.
So the likely future is not a clean transfer of power from Google to ChatGPT.
It is more likely that search becomes more conversational, retailers become better at answering questions, and AI assistants become another route into the same commercial ecosystem. Readers may move between all three without giving much thought to which technology is doing what.
For publishers, the more important point is that the discovery journey itself is being compressed.
AI may send less traffic, but better traffic
This creates a strange measurement problem.
If readers increasingly use AI to discover products, we might expect publishers and retailers to receive large amounts of traffic from ChatGPT and similar systems. So far, that is not happening.
Much of the research phase remains inside the AI interface. This fits with the wider rise of “zero-click” behavior, where users get enough information from a generated answer that they never visit the original source. A study on AI-enhanced search found lower click-through rates when generated summaries appear above conventional results.
ChatGPT referrals remain tiny compared with conventional search. But the visitors who do arrive may be further along in their decision-making.
An e-commerce analysis has reported conversion rates several times higher than ordinary organic traffic in some cases, although there is not yet comparable evidence showing the same effect specifically for book retail.
The explanation is fairly intuitive. A traditional buyer might move between Google, review sites, Reddit, Goodreads, and retailer pages before choosing a book. Each stop generates traffic, even though most of those visits are part of the research process rather than an immediate purchase.
With conversational AI, much of that comparison can happen before the reader leaves the chat. By the time someone clicks through to a retailer or publisher, they may already have decided that a particular title fits what they want. Some research on AI-referred traffic has found that these visitors spend more time on destination sites and browse more pages than ordinary visitors.
That leaves publishers with an uncomfortable possibility: AI could send fewer visitors while still influencing more decisions.
If that happens, raw website traffic becomes a less complete measure of whether a book is actually being discovered.
So how does ChatGPT decide which books to recommend?
This is where the publishing side becomes much more important.
An AI assistant does not objectively read every book ever published before deciding which three are best suited to a reader. It works with the information available to it.
That can include publisher websites, retailer listings, reviews, author pages, media coverage, structured metadata, citations, and discussions elsewhere online. It needs enough evidence to understand that a book exists, what it is about, and why it might match a particular request.
This is one reason Generative Engine Optimization, or GEO, has begun attracting attention. SEO asks how to make a page perform well in conventional search. GEO is concerned with whether information is clear, trustworthy, and structured enough to be retrieved and used inside a generated answer.
Early research suggests that verifiable statistics, attributed quotations, and reputable citations can improve visibility in generative search environments, while old-fashioned keyword stuffing can be ineffective or even counterproductive.
For publishers, however, the most interesting lesson may be much less fashionable: metadata matters.
ONIX and other structured publishing data already tell booksellers, distributors, and libraries about a title’s subject, contributors, format, rights, accessibility features, and other characteristics. Richer metadata can also make it easier for machines to understand whether a particular book matches a complicated request.
Consider a specialist history title published by a small or university press in 2016. It has a two-line description, one broad subject category, an author biography written a decade ago, and almost no recent discussion online. The book is still available. Its scholarship may still be excellent. But from the perspective of a system trying to answer a detailed recommendation request, there may simply not be enough useful information around it.
The book exists in the catalogue. It may barely exist in the machine-readable world surrounding the catalogue.
That gives publishers a practical problem. A marketing team looking at its backlist may need to ask not only which books still sell, but also which books are sufficiently described, reviewed, and connected to other reliable sources to be understood by a recommendation system.
For some titles, improving discoverability may be less about running a new campaign and more about repairing the digital record around the book.
AI likes books that are already visible
Better metadata does not solve everything.
AI recommendation systems also inherit the patterns of the information available to them. Books that are already widely reviewed, quoted, discussed, and linked naturally generate more signals than books that barely appear online.
That can become self-reinforcing.
A popular book attracts more reviews and coverage. Those mentions give AI systems more information about the book. The book becomes easier to retrieve and recommend, which can generate still more attention.
Meanwhile, a strong but obscure title may have fewer reviews, fewer links, fewer reader discussions, and thinner metadata. The system has less evidence to work with, even if the book would actually be the better recommendation.
Research on LLM-based recommendations has found exactly this kind of popularity bias, with well-known titles and highly visible cultural products enjoying an advantage over niche alternatives.
This raises a much more interesting question than whether AI can recommend books:
Can AI recommend books we do not already know?
That matters especially for midlist authors, independent presses, regional publishing, specialist academic titles, and books published in languages or markets with a smaller online footprint.
These are often the books that benefit most from human discovery.
A bookseller can place an overlooked title in someone’s hands. A librarian can remember the perfect book from six years ago. A reviewer can make a case for a debut that has no existing audience.
A recommendation system that mainly follows the strongest available signals could do the opposite. Instead of expanding discovery, it could quietly reinforce what is already visible.
For publishers, particularly smaller ones, that may be the real strategic concern. The question is not merely how to make books legible to AI, but whether a discovery system built on existing digital evidence can ever give underexposed books a fair chance.
The search journey is changing
For the last two decades, publishers have gradually learned how to make books easier to find online. They improved descriptions, supplied keywords, built author pages, maintained metadata, gathered reviews, and worked to ensure that retailers and search engines had enough information to surface a title.
None of that is becoming obsolete.
What is changing is the context in which that information is used.
A reader may no longer encounter a publisher’s description because it ranked highly in Google. The same information might instead contribute, indirectly, to an AI deciding that a particular title belongs in a three-book shortlist.
That makes the quality of the digital information surrounding a book more important, not less. It also makes third-party discussions, reviews, and authoritative references harder to dismiss as mere promotional extras.
The shift will probably be uneven. Some readers will continue to browse bookshops. Others will find books through TikTok, newsletters, Goodreads, podcasts, librarians, or friends. Google will remain an enormous part of online discovery.
But conversational AI adds something different: the ability to turn a vague desire into a recommendation without requiring the reader to do much searching at all.
That is why ChatGPT does not need to replace Google to change book discovery.
It only needs to replace part of the journey.
Readers will still search, browse, and ask other people what to read. Increasingly, though, some will begin by describing exactly what they want to a machine and asking it to choose.
Google taught publishers to compete for the first page.
AI may force them to compete for something much smaller: a place inside the answer.