The AI Bookshelf Report: AI builds its shelf from Wikipedia, not Amazon.
Where AI assistants really get the books they recommend — and what every author should do about it.
When you ask ChatGPT, Gemini or Perplexity to recommend a book, they build the answer from reference and community sites — not from shops. Across 188 books in 16 genres, we counted every source the models cited. Wikipedia was cited 2,200 times and Goodreads 1,493 times; Amazon — the world's largest bookstore — was cited just 134 times. To be recommended by AI, a book has to exist on the sites AI trusts, not only on its Amazon listing.
Where AI's book recommendations actually come from
We logged every source the three models cited as they recommended books — roughly 21,900 citations in total. Ranked by how often each source appeared:
Total citations across 188 books × 16 genres × 3 engines. Preliminary, June 2026.
The three pillars: reference, community, video
AI doesn't recommend books from a shop window. It recommends them from the open, structured, well-linked corners of the web:
Wikipedia is the spine
Wikipedia was the single most-cited source and the only one to appear across all 16 genres. A book (or author) with a solid, well-sourced Wikipedia presence is far more "knowable" to a model than one without.
Goodreads and YouTube carry the conversation
Goodreads (1,493) and YouTube (1,373) sit just behind — the models lean heavily on community reviews and creator video. Where readers talk about and film books, AI listens.
Amazon is a shop, not a source
At just 134 citations, Amazon behaves the way the models treat it: a place to buy, not a place to learn what's worth recommending. Optimising only your Amazon listing leaves you invisible to AI discovery.
Some genres are far easier for AI to recommend
Not every shelf is equal. We measured how readily the three models actually name a genre's books — Surfio's AI Discovery Score (0–100) — and averaged it across each genre. The spread is wide. Even the most AI-visible shelf, personal finance, averages just 54/100; the hardest, crime thrillers, averages 28. A personal-finance author is roughly twice as likely to be named by AI as a crime-thriller author asking the same kind of question.
Average AI Discovery Score (0–100) per genre · selected rows from the 16-genre table · June 2026.
The niche kingmakers
Beyond the universal giants, each genre has a source that punches far above its weight — one site AI turns to again and again for that one shelf. The striking part is how concentrated they are: for most of these, nearly every citation the model gave them landed on a single genre. Get featured here and you're talking directly to the model's source list.
Nearly every time AI cited the ALA, it was naming a young-adult title. The library establishment quietly shapes the YA shelf.
AI's default for business books: 96% of HBR Store citations landed on the business shelf — almost a pure-play kingmaker.
A specialist sport-science press. Every single one of its AI citations landed on the health-and-fitness shelf.
A niche SFF magazine most readers have never heard of — and AI's most over-indexed source for epic fantasy.
A personal-finance media site, not a bookshop — and 100% of its AI citations were for money books.
A single cookery newsletter the models reach for first when recommending cookbooks — wholly concentrated in one shelf.
What this means for your book
AI discovery is winnable — but the work happens off Amazon. Based on the data, the highest-leverage moves:
- 1Become knowable to Wikipedia. A well-sourced author or book page is the closest thing to a master key for AI visibility.
- 2Own your Goodreads presence. Complete profile, real reviews, correct metadata — it's the community layer AI reads.
- 3Get on the lists AI trusts. Pitch the curators (FiveBooks) and your genre's kingmaker rather than chasing Amazon ranking alone.
- 4Show up in video. A book mentioned on YouTube — BookTube, interviews, trailers — feeds a channel the models lean on heavily.
How we ran this
Surfio probes the AI assistants readers actually use with the questions readers actually ask, then records which sources each model cites. This report aggregates that data across our curated genre set.
- Models
- ChatGPT, Gemini and Perplexity (web-grounded). Claude and Amazon's Rufus are being added to the live tool.
- Sample
- 188 books across 16 genres (≈12 per curated genre).
- Volume
- ≈21,900 cited sources logged and de-duplicated.
- Counted
- Every source a model cited while recommending books; spam/dead domains filtered (≥2-book trust gate).
- Dated
- Preliminary snapshot, June 2026. AI answers shift over time; figures will be re-locked as the dataset grows.
- Run by
- Surfio, the AI book-visibility tool, powered by publishing.co.uk.
Share the findings
All six graphics are free to republish with credit to Surfio (powered by publishing.co.uk) and a link back to this report. Click any graphic to open it full-size, then save or embed.
Wikipedia vs Amazon: the 16× headline
AI Discovery Score by genre
Why your Amazon page is invisible to AI
The hidden single-genre kingmakers
The AI book-authority pyramid
Four ways to get recommended by AIPress enquiries and the full dataset: publishing.co.uk.
Frequently asked
How does AI decide which books to recommend?
It doesn't read your sales page — it assembles an answer from sources it trusts. In our data the dominant sources were Wikipedia (2,200 citations), Goodreads (1,493) and YouTube (1,373). A book that's well represented on those sites is far more likely to be named than one that exists only on Amazon.
Does Amazon matter for AI book recommendations?
Far less than most authors assume. Across 188 books, the three models cited Amazon just 134 times — about one-sixteenth as often as Wikipedia. Amazon is where readers buy; it is not where AI learns what to recommend.
How do I get my book recommended by ChatGPT?
Build presence on the sources AI cites: a well-sourced Wikipedia page, a complete Goodreads profile, placement on curated lists (FiveBooks and your genre's specialist site), and mentions in video. See the four-step checklist above.
Which AI models were tested?
ChatGPT, Gemini and Perplexity, all with live web access. Claude and Amazon's Rufus (“Alexa for Shopping”) are being added to the Surfio live tool, after which we'll publish an updated, five-engine edition.
Can I see how AI sees my own book?
Yes. Surfio runs the same probes against your specific title and shows which engines name it, at what list position, and which books win instead. Start with a free check.
See how AI sees your book
Surfio asks the assistants what readers ask — and shows you exactly where your book stands, and how to climb. Free to start.
Run a free AI visibility check →Preliminary findings, June 2026. Figures are drawn from Surfio's live citation data and will be re-locked with a dated methodology note as the dataset grows beyond 12 books per genre. Not affiliated with or endorsed by OpenAI, Google, Perplexity, Amazon or Wikimedia.