Home › Research › People and personal brands
SEMPITE Research · Category Study · August 2026
We Logged Every Source in 95 AI Book Recommendations. Author Websites Got Zero.
We asked Google’s AI sixteen questions a real reader would ask — by trope, by life stage, by setting — and recorded every source behind every answer. The recommendations are assembled almost entirely from surfaces authors don’t own: Reddit threads, BookTube, Goodreads, the book press, and — encouragingly — very small independent book blogs. What an author publishes on their own site was never once the cited source.
Where AI book recommendations actually come from
All 95 citations across the 11 answers, classified by surface type:
Author-owned websites: zero of 95. Not one citation, across eleven answers spanning romance tropes, senior love stories, Catskills settings and general discovery, pointed at any author’s own domain. The AI builds its book advice entirely from what the rest of the web says.
The individually most-cited domains: YouTube (11 citations), Reddit (10), Goodreads (7), Amazon (6). BookTube videos and Reddit recommendation threads together out-cite every retailer combined — the “what should I read” answer is social before it is commercial.
The more specific the reader, the more likely AI answers
Share of queries in each cluster that triggered an AI Overview:
Trope-language queries — the way genre readers actually describe what they want next — triggered an AI answer every single time. Broad “best books” queries mostly did not. If your book’s marketing doesn’t speak trope, it is absent from the surface where the AI answers with the most confidence.
Every niche has its own gatekeepers
The aggregate hides the most actionable finding. On the Catskills and Borscht Belt setting queries, the AI reached for a specific shelf of sources: Hadassah Magazine, the Jewish Book Council and Kveller — the Jewish literary press — alongside regional outlets. For an author writing in that world, those three mastheads are not nice-to-have coverage; they are the citation path into the AI answer for their entire subject.
The same pattern held in our supplement study, where three review sites carried 77% of answers. The gatekeepers change per niche. The structure doesn’t: find who the AI already cites in your corner, and earn your way onto those specific pages.
The small-blog surprise
Nearly one citation in five went to small independent book blogs — sites like littleinfinite, Read and Wright, Everyday Reading and Love in Panels, plus a Substack. These are surfaces a working author can actually reach: they accept pitches, they review midlist and indie titles, and Google’s AI treats them as citable authorities on what to read. For authors without national press access, book-blogger outreach is not a consolation prize — in this data it out-cites the New York Times.
What this means for authors
The reflex when a book is invisible is to improve the author website. This data says the author website was never the source — and also says that doesn’t make it pointless. Its job is different: to be the legibility layer — the consistent name, bio, series and title information (with Book and Person schema) that lets the AI connect what Goodreads, Reddit, a blogger and a magazine all say about the same person. The citations happen elsewhere; the identity is resolved on your site.
The working order that follows from the data: complete and active Goodreads presence first; then the community surfaces where recommendations actually form (Reddit threads, BookTube/BookTok); then targeted pitches to the small blogs and the niche press the AI already cites in your category. Publishing more pages on your own domain comes last — and our broader AI-visibility data keeps pointing the same direction.
Methodology
Data collected 13 August 2026, United States, English, desktop. Sixteen reader buying-intent queries were distributed across four clusters — trope (“best second chance romance books”), seasoned romance (“romance novels with older main characters”), setting (“books set in the Catskills”), and general discovery (“best romance books 2026”) — and submitted to Google via the DataForSEO live SERP endpoint with AI Overview loading enabled. Eleven returned an AI Overview; the five that did not are excluded from citation percentages.
For each answer we recorded the full reference list — 95 citations across 34 distinct domains — and classified every domain by surface type: social & video (YouTube, Reddit, Facebook, TikTok, Instagram), book press & publishers (NYT, Lit Hub, Kirkus, Penguin Random House, BookBub, niche magazines), small independent book blogs, retailers (Amazon, Bookshop, Audible, Google Books, independent bookstores), reader platforms (Goodreads, Fable) and Google properties. “Author-owned” means a domain operated by an individual author; none appeared.
Limits worth stating plainly. This is a single snapshot of one AI surface — Google’s AI Overview — and results vary by user, time and location; individual queries will not always reproduce. Sixteen queries is a category probe, not a census; we sized it to match the question “does the AI ever cite author sites,” where zero-of-95 is a meaningful answer even at this scale. Visibility is not a sales measure; we make no revenue claims.
Underlying data, including the full query list and per-query citation records, is available on request under CC-BY.
Are you an author (or managing one)?
See how the AI engines describe you and your books today — free, in about 60 seconds.