Citation

Background

A citation is the link a generative-search system draws between a span of its generated answer and a specific source document, usually surfaced as a footnote, inline reference, or a listed URL. It exists because a language model's raw output is unattributable — the text is sampled from learned weights, not copied from a knowable source — so systems that want to be trustworthy bolt retrieval onto generation and then track which retrieved passage supported which claim. Mechanically, a retrieval-augmented system fetches candidate passages, feeds them into the prompt, and instructs the model to ground its answer in that supplied context and to emit markers (often bracketed indices) tying sentences back to specific passages. A post-processing step maps those markers to the original URLs. Some systems go further and run a verification pass, checking whether the cited passage actually entails the sentence it is attached to, discarding or re-attributing citations that do not hold. A concrete example: ask an answer engine "what port does Postgres listen on by default" and it returns "5432" with a citation to the Postgres documentation page it retrieved. For people building software with AI, citations matter in two directions: they let you audit and trust a model's factual claims rather than accepting them blind, and — because AI answer engines increasingly send users to cited pages — being the source a model cites is becoming its own distribution channel, distinct from classic search ranking.