In plain terms
Footnotes for AI answers. A good answer does not only say “the return period is 30 days”; it adds a marker that leads to the page and paragraph where that is written. The reader can then check the claim in seconds without having to take the model's word for it.
Why it matters
Citations turn an answer from something you must trust into something you can verify, which is what makes AI usable in legal, medical, financial and audit work. They are no guarantee. A model can cite a real source that does not support the sentence, and a model answering from memory can invent a reference that looks perfect. Citations make checking fast; they do not make it unnecessary. On the web they have a second meaning for businesses: being cited in AI answers is the new form of visibility.
Example
A legal assistant answers a question about a limitation period and gives two citations. The lawyer clicks the first and sees the exact clause, highlighted. The second leads to a clause about a different type of contract: the sentence built on it is wrong. The check took half a minute. Without citations the error would have gone into the memo.
Most often confused with
Citations vs. Grounding
Grounding is the practice of answering from supplied sources; citations are the visible trail that proves it. An answer can be grounded without showing citations, and it can show citations without being faithfully grounded. You want both: answers built on sources, and links that let a person confirm it.
Under the hood
Citations are produced in three ways: the model is given numbered passages and asked to mark which one supports each claim; an API returns structured citations with the exact quoted span from the supplied documents; or a separate step attributes sentences to sources after the answer is written. Granularity ranges from the document to the page, the passage and the exact sentence. Two measures: citation precision (does the cited source really support the claim?) and citation recall (does every claim that needs support have a citation?). Known failures: fabricated references when no retrieval took place, a real source attached to the wrong claim, and reliance on a weak source. Good interfaces show the quoted text and link directly to its location.