How to Measure Citation Accuracy in RAG Agents (2026)
A citation makes an answer look verifiable. That is exactly why a wrong one is dangerous: it manufactures confidence. An agent can cite a document that genuinely exists, in a corpus it genuinely retrieved from, attached to a claim that document never makes — and a reader who trusts the citation format will never check. Citation accuracy is the metric that catches the gap between "there is a reference here" and "the reference supports this".
Quick answer: Citation accuracy measures whether the sources an agent cites actually exist, are correctly attributed, and genuinely support the claim they are attached to. Score those three axes separately — existence, attribution, support — because they fail independently and a single number hides which one broke.
What is citation accuracy, and why isn't groundedness enough?
Groundedness asks a whole-answer question: is this answer supported by the retrieved context? Citation accuracy asks a pointer-level question: does this specific reference resolve to a passage that supports this specific claim?
An answer can pass the first and fail the second. If the retrieved context contains the supporting fact in paragraph four, but the agent cites paragraph two, the answer is grounded and the citation is wrong. For a casual reader that difference is invisible. For a clinician or a lawyer following the reference, it is the whole ballgame — see how to measure RAG groundedness for the complementary metric.
Why do agents cite the wrong source?
Three recurring mechanisms, in rough order of severity:
- Fabricated references — a plausible-looking identifier, author, or case number that does not exist. Trained on citation-shaped text, models produce citation-shaped text.
- Misattribution — a real, retrieved source attached to the wrong claim, often because the claim came from a neighbouring chunk and the citation drifted.
- Unsupported attribution — the cited passage is real and relevant but doesn't actually entail the claim; the agent synthesised across sources and pinned the result on one of them.
How do you score citation accuracy?
Score three axes separately, and never average them into one figure:
- Existence — does the cited identifier resolve against your corpus index? This is deterministic and cheap; check it first, because a fabricated reference makes the other two axes moot.
- Attribution — is this the source the claim actually came from, or a neighbour?
- Support — does the cited passage entail the claim attached to it?
Existence is a lookup. Attribution and support need an entailment judgement, so use an LLM judge — and validate that judge against human labels periodically, because an unvalidated judge drifts and takes every downstream number with it.
Report per-claim results, not one score. "Citation accuracy 0.7" is unactionable; "two claims cite passages that don't support them, one reference doesn't exist" tells you what to fix and how badly you're failing.
Which domains need this most?
Wherever someone will act on the reference. Medical agents, where a citation supports a dosage or contraindication. Legal agents, where a fabricated case citation has professional consequences. Financial agents, where a misattributed figure moves a decision. In these domains, a wrong citation is worse than no answer, because no answer prompts a human to look it up — while a wrong citation stops them from looking.
That is also why citation accuracy belongs in the same suite as abstention: an agent under pressure to produce a reference will invent one. Test both together, or you optimise one into the other — see how to test RAG agent abstention.
How do you gate it in CI?
Run citation checks against a private golden set on every change to prompts, models, chunking, or retrieval. Start with the deterministic existence check as a hard gate — any fabricated reference should fail the build outright, no threshold — then treat attribution and support as scored metrics with a regression threshold. Keep the case set out of public logs so it keeps discriminating over time.
Related
- How to measure RAG groundedness — the whole-answer metric this pairs with.
- How to test RAG agent abstention — why a pressured agent invents citations.
- Medical RAG groundedness pack · Legal — verified suites for the domains where this bites.
- RAG groundedness benchmark — discriminating power measured on a real reference panel.
- Capability packs — the full category.
FAQ
Citation accuracy measures whether the sources an agent cites actually exist, are correctly attributed, and genuinely support the claim they are attached to. All three can fail independently.
Groundedness asks whether the answer is supported by the retrieved context as a whole. Citation accuracy asks whether each specific pointer resolves to a passage that supports its specific claim. An answer can be grounded overall while citing the wrong paragraph.
Common causes are fabricated references that look plausible, citations copied from a neighbouring chunk, and citations attached to a synthesised claim that no single source makes. Fluency makes all three hard to spot by eye.
Score three axes separately: existence (does the cited source exist in the corpus), attribution (is it the source the claim came from), and support (does the cited passage actually entail the claim). Report them separately, not as one number.
Medical, legal, and financial agents, where a reader will act on the citation. In these domains a fabricated or misattributed reference is worse than no answer, because it manufactures false confidence.
Existence can — check the cited identifier against your corpus index. Attribution and support usually need an entailment check, so use a validated LLM judge and confirm it against human labels periodically.