Citation Accuracy Testing
Citation accuracy testing verifies that when an AI system cites a source — a document, a page number, a URL — that source actually exists, actually contains the claim being attributed to it, and is the real source that was retrieved, rather than one that sounds plausible but was fabricated.
Citations are one of the most convincing forms of AI hallucination, because a fake citation looks exactly like a real one — a correctly formatted reference to a paper, page, or URL that simply doesn't exist or doesn't say what it's credited with saying. Users tend to trust cited answers more, which makes an inaccurate citation more dangerous than an uncited wrong answer, not less.
Testing means checking three separate things that are easy to conflate: does the cited source exist at all, does it actually contain the claim (not just related content), and is it the source that was actually retrieved and shown to the model, rather than one it generated from memory. The last check matters especially for RAG systems — a model can cite a real, correct-sounding source it was never actually given, which passes a naive "does this URL exist" check but still fails the system's actual grounding guarantee.
