How do AI systems choose which sources to cite?
AI answer systems generally run a search, retrieve a set of candidate pages, select passages that appear to answer the question, and generate an answer citing the pages the passages came from. Selection therefore favours pages that are retrievable for the query, self-contained enough that a single passage answers the question, and corroborated elsewhere. Citation is a retrieval outcome, not a ranking reward, and it varies between runs of the same question.
Three stages matter. Retrieval decides which documents enter the candidate pool — largely a search problem. Passage selection decides which chunk is read, favouring direct definitions, tables and short answers over narrative build-up. Generation decides which of the read sources are named; models frequently use a source without citing it, and cite sources that merely corroborate. Our own sampling shows the named source changes across repeated runs of an identical prompt.
A pricing page that states a number in a sentence tends to be cited; a pricing page that hides the number in an image or a gated form is read past, even when it ranks well in conventional search.
Anyone trying to be found inside AI answers should treat this as three separate problems with different fixes — indexability, passage clarity, and third-party corroboration — rather than as a single ranking score to optimise.
- That being cited once is a stable position. Repeat the prompt and the citation set often changes.
- That schema markup causes citation. It helps machines parse a page; it does not decide retrieval.
- Answer systems do not publish their retrieval pipelines, so all external accounts are inferred from observation.
- Results differ by account, region, and model version, so no sample generalises fully.