Nearly one in four good books
is invisible to AI.

For each book we generated 8 questions from its subject matter alone, never its title, so the questions could not be shaped to flatter the result. We then asked those questions cold and recorded the books that came back.

The assistant was never told which book we were testing, in either step. Whether a book appeared was then checked mechanically against its answers, because asking a model whether it mentioned something invites it to agree with you.

Answers came from what the assistant already knows, with web search switched off, which is what a reader gets when they simply ask. One assistant was tested, so treat this as a sample of how AI recommends books rather than a census of every tool.

One book turned up twice in our source library under two records, so we measured it twice by accident. The two readings came out 13 points apart, which is a fair indication of how precise a single reading is. Read these numbers as a range rather than a decimal.

Measured 2026-07-30. Books were drawn from a library of several thousand nonfiction titles, ranked by reader rating, with fiction excluded. 101 records covered 100 distinct books.