Citation Accuracy — AI Glossary

Reference details

Related terms

• Grounding
• Hallucination
• Context Precision
• Context Recall
• Evaluation Harness
• RAG (Retrieval-Augmented Generation)

How it works

1. Answers are decomposed into individual claims, because a citation attaches to a claim rather than to a paragraph and a single answer usually makes several.
2. Each claim is paired with the specific passage its citation points at, resolved to the passage rather than the document, since a document-level pointer cannot be checked.
3. A judge — human or model — reads the claim against that passage alone and assigns support, partial support, or no support, with no access to the rest of the corpus.
4. Withholding the corpus is the point: a judge allowed to search will confirm claims that are true in the world but unsupported by the passage cited, which is the exact failure being measured.
5. Scores are aggregated per claim and then per answer, because an answer is only verifiable if every one of its citations holds.
6. The result is reported separately from answer correctness, since the two are independent and a system can move one while leaving the other unchanged.

Example: Why per-citation accuracy overstates per-answer verifiability

Suppose each individual citation is correctly attached 90% of the time, which sounds like a healthy system. An answer carrying three citations is fully verifiable only if all three hold, and if the failures are roughly independent that is 0.9^3 ≈ 0.729 — so about 72.9% of answers survive a reader checking every citation, against the 90% the per-citation metric reports. The gap is 90 - 72.9 = 17.1 points, and it exists entirely because the unit of measurement is not the unit of trust. The effect worsens exactly where you would least want it to. A longer, more thorough answer carries more citations, so a five-citation answer is fully verifiable 0.9^5 ≈ 0.590 of the time, meaning roughly 41.0% of such answers contain at least one mis-attached citation. The more carefully the system shows its work, the more likely a diligent reader is to find a pointer that does not support what it is attached to — which is why the common shortcut of scoring only the first citation per answer is not a mild approximation. On a five-citation answer it measures 90% where the answer-level figure is about 59.0%.

Commonly confused with

Citation accuracy is not groundedness, and keeping them apart is the whole reason to measure it. Groundedness asks whether a claim is supported by anything in the retrieved context; citation accuracy asks whether it is supported by the passage the answer actually points at. An answer can be fully grounded and badly cited — the support exists two chunks away from the pointer — and that failure is invisible to a groundedness score while being precisely what breaks a reader who clicks through. It is also not answer correctness. A right answer with a wrong citation and a wrong answer with a convincing citation are both common, and the second is more dangerous, because the citation is what a user checks instead of the claim.

When to use it

Reach for it when:

• On any RAG surface where users can follow citations, since the pointer is what they verify instead of the claim.
• Measured per claim and aggregated per answer, because an answer with one bad pointer is not partially trustworthy.
• With the judge restricted to the cited passage, which is what separates mis-attribution from mere incorrectness.
• Alongside groundedness rather than instead of it, as the two can move independently.

Reach for something else when:

• Inferred from retrieval recall, which says the right passage was fetched and nothing about where the answer pointed.
• Scored on the first citation only, which reported 90% against about 59.0% at the answer level above.
• With document-level citations accepted for passage-level claims, where the pointer is too coarse to check.
• As a proxy for answer quality, which it is deliberately independent of.

Where this appears on ModelRefs

• measure retrieval recall and grounding

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Citation Accuracy — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path