A side-by-side comparison of Citation Precision and Citation Recall. Understand how correct support for cited claims differs from exposing the source documents used to produce a response.
Citation Precision measures metric that measures the share of citations in a model output that correctly support the claims they are attached.
Context: Most relevant when evaluating whether model citations are accurate and useful for verification.
Citation Recall measures metric that measures the share of source documents used to produce a response that are actually cited in the response.
Context: Most relevant when evaluating whether a response discloses the main evidence sources behind it.
| Aspect | Citation Precision | Citation Recall |
|---|---|---|
| What it measures | Citation precision measures whether citations correctly support the claims they are attached to. | Citation recall measures whether the source documents used to produce the response are actually cited. |
| Best use case | Best for checking claim support and preventing decorative or misleading citations. | Best for checking evidence coverage and whether relied-upon sources are visible to the reader. |
| Failure mode | Low precision means citations do not support the claims they appear to verify. | Low recall means important source documents were used but not exposed in the response. |
| Threshold sensitivity | Precision can fall when extra citations are attached without claim-level relevance. | Recall can rise when more sources are cited, but that may reduce readability or precision. |
| Common mistake | Counting citations as correct without checking the claim they support. | Assuming a response is well-supported just because it cites many documents. |
In practice, citation precision protects the reader from false support, while citation recall protects the reader from hidden evidence. Strong systems need both.
Treating every citation as valid without checking support.
Optimizing for many citations instead of useful citations.
Ignoring uncited source documents used by the system.
Evaluating citations separately from the claims they verify.
Use Citation Precision when the goal is to evaluate whether each citation actually supports the claim it is attached to. It is central for grounded answer quality, legal research, compliance summaries, and RAG evaluation.
For AI assurance evidence, citation metrics should be tied to groundedness, traceability, and review workflows. NIST AI RMF-style evaluations should avoid relying on citation count alone and should test both support quality and source coverage.
Yes. The citations that appear may be accurate, but the response may omit other source documents that were used to produce the answer.
Yes. The response may cite many or most source documents, but some citations may not actually support the attached claims.
Citation precision is critical because unsupported citations can mislead readers. Citation recall is also important where auditability and evidence coverage matter.
No recently viewed comparisons yet.