AI search citation reporting often starts with a deceptively simple question: did the engine cite us? That is a useful visibility signal, but it says little about whether the engine selected appropriate evidence, omitted important sources or cited material that does not support the answer.
AI search citation precision and recall provide a more disciplined way to assess this. Precision asks whether the citations shown were relevant and suitable. Recall asks whether the answer surfaced the relevant evidence it should reasonably have cited. Used together, they expose a common problem in answer-engine measurement: a visible citation can coexist with a poorly evidenced answer, while a strong source can be retrieved but never become visible to the user.
This framework is designed for recurring evaluation across prompts, engines and content categories. It is not a claim that every query has one objectively correct source list. The practical job is to construct a defensible relevance baseline, record uncertainty, and make the scoring rules stable enough that trends mean something.
Start by defining the measurement unit
Do not score a whole answer as one undifferentiated citation event. An answer may make several claims, cite several URLs, and attach a citation to a sentence that it only partly supports. The unit of analysis should reflect that reality.
For each prompt and response, capture four linked objects:
- Prompt: the exact user request, locale, date, interface and relevant settings.
- Atomic claim: a verifiable proposition in the answer, such as a product capability, definition, threshold or recommendation.
- Citation: the displayed URL, publisher or source family and its placement in the response.
- Source-role judgement: whether that source is authoritative, relevant, current and sufficient for the claim it is attached to.
Keep source, page and source-family views separate. A government department may publish several pages that are individually relevant, but they remain one publisher family for concentration analysis. Conversely, two syndicating outlets should not automatically be counted as independent evidence. Clean canonical URLs and deduplicate copied pages before scoring; this is where a rigorous citation-data cleaning process pays for itself.
Build a relevance baseline before looking at the engine
Precision and recall require a comparison set. For each prompt, create a relevance baseline: the sources and claim-evidence pairs that a reasonable answer could use. Build it independently of the response under review where possible. If annotators see the model’s citations first, they will tend to ratify them.
For a narrow factual prompt, the baseline may contain one primary source and one credible corroborating source. For a recommendation or complex YMYL query, it may include official documentation, independent testing, expert guidance and sources representing material trade-offs. Relevance is not synonymous with a high-ranking page, a familiar publisher or the client’s domain.
Label each candidate source as required, acceptable, helpful but optional, or out of scope. A required source is necessary because it is primary evidence, the only current authority, or directly substantiates a central claim. Acceptable sources can support the same claim adequately. This graded design avoids treating ten interchangeable pages as ten missed obligations.
Document why every baseline item belongs there: claim supported, evidence type, publication or update date, jurisdiction, and known limitation. A reusable dataset should also preserve prompt versions and adjudication notes. For a fuller setup process, see how to build an AI search evaluation dataset.
Score citation precision at source and claim level
At source level, citation precision is the share of displayed citations judged relevant to the prompt and appropriate for the answer:
Source precision = relevant displayed citations / all displayed citations
A binary label is adequate for simple tracking, but a weighted score is usually more informative. Give each citation a relevance score from 0 to 1 based on topical fit, evidential authority, freshness where it matters, and whether the cited page actually supports the stated point. Then divide the sum of citation relevance scores by the number of citations displayed.
This prevents an answer with five weakly related links from looking equivalent to one with five directly supporting links. It also prevents a useful but secondary explainer from being treated as equal to a primary standard when the question asks for a rule or specification.
Source precision alone can still flatter an answer. Consider a response with one excellent citation attached to a paragraph containing four factual assertions. The citation may be relevant to the topic but support only one assertion. That is why claim-level precision matters:
Claim citation precision = supported cited claims / claims presented as supported by citations
Assign each cited claim one of four labels: fully supported, partly supported, unsupported, or contradicted. In my experience, “partly supported” is the most operationally useful label. It catches answers that turn a qualified source statement into a universal recommendation, or cite a page for a feature while adding unverified performance claims.
Define in advance whether partly supported claims receive 0.5 credit or must be split into smaller atomic claims. Splitting is more work but produces clearer diagnostics. Pair this with summary-fidelity measurement when omitted caveats or changed meaning are the concern.
Measure recall without pretending the reference set is complete
Classic recall is the share of relevant items retrieved:
Source recall = relevant baseline sources cited / relevant baseline sources
In AI search, that denominator is rarely complete. There may be unknown but suitable sources, regional variants, pages blocked from inspection, or several acceptable alternatives. Reporting a single definitive recall figure without qualification creates false precision.
Use one of three approaches, chosen by query type:
- Required-source recall: the proportion of required baseline sources cited. Best for primary documentation, official rules and high-stakes factual claims.
- Claim-evidence recall: the proportion of required claims that have at least one adequate visible supporting citation. Best for answers with many valid source alternatives.
- Bounded recall: report a lower bound from confirmed relevant sources and an upper or sensitivity range that accounts for plausible alternatives or unresolved items.
Claim-evidence recall is often the fairest headline metric. If three independent pages adequately establish the same non-controversial fact, the engine should not be penalised for failing to cite all three. But it should be penalised if it misses the only source supporting a key limitation, eligibility rule or exception.
Record “not assessable” rather than forcing a zero when citations are absent, inaccessible or not mapped to claims. Separately record uncited but accurate claims. They may reflect model knowledge, retrieval that was not exposed, or unsupported generation; they do not count as citation recall. The distinction is central to measuring uncited influence and attribution gaps.
Separate retrieval quality from answer visibility
A citation evaluation can only observe what the interface exposes unless the engine supplies retrieval logs. Treat these as different stages:
| Stage | Question | Useful measure |
|---|---|---|
| Eligibility | Could the page be accessed and considered? | Technical accessibility and retrieval-eligibility checks |
| Retrieval | Was relevant evidence retrieved? | Retrieved-source or passage recall, where logs exist |
| Selection | Was it chosen for the response? | Selection rate among eligible or retrieved sources |
| Visibility | Could a user see its citation and connection to a claim? | Displayed citation precision and claim-evidence recall |
Without logs, call the result visible citation recall, not retrieval recall. A source absent from a response may never have been retrieved, may have lost a selection competition, or may have informed the answer without being cited. These have different remedies. Improving passage structure may help retrieval; making a claim more directly substantiated may help selection; clearer page-level evidence may improve citation attachment. Google’s guidance on AI features in Search is useful context for what can be surfaced, but it does not provide a universal citation-ranking formula.
Run tests at passage level when possible. A page can be eligible while the relevant passage is ambiguous, buried or stripped of its qualifiers. That is a content diagnosis, not proof that the entire domain lacks authority.
Use a worked case to reveal overlap and concentration
Suppose six responses to a policy query contain eight family-level citations after URL normalisation: six cite Google Search Central and two cite an independent publisher. Google Search Central’s citation share is 6/8, or 0.75; the independent publisher’s share is 2/8, or 0.25.
The two-family Herfindahl-Hirschman Index is therefore 0.75² + 0.25² = 0.625. That indicates concentrated citation selection in this small sample. It does not mean the answers are wrong: a primary source may properly dominate a policy question. It does mean that high aggregate precision should not be mistaken for broad evidential diversity.
Now add claim labels. If Google supports the core policy rule but the independent publisher is the only cited source addressing implementation limitations, source recall may look healthy while claim-evidence recall for limitations is poor. The remedy is to report recall by claim category, not just one portfolio average. Check whether the same source families recur across engines too; apparent multi-engine corroboration can be less independent than it looks.
Make the framework repeatable, not merely rigorous on paper
Use a fixed prompt panel, capture raw responses and citation destinations, and retain screenshots or exports with timestamps. Models, indexes and interfaces change. Without provenance, month-to-month differences may be artefacts of query rewriting, locale, a model update or an annotator’s changed interpretation. Track rewritten prompts where observable, because a rewritten query can change the relevant baseline before any source is selected.
Have two reviewers label a meaningful sample of claim-source pairs. Resolve disagreements with written decision rules, especially around authority, partial support and acceptable substitutes. Evaluator agreement is a measurement result in its own right, not administrative overhead.
Report precision, required-source recall, claim-evidence recall, citation concentration and an uncertainty note together. Segment by intent, topic risk, engine and response format. A single blended score is tempting for dashboards, but it conceals the difference between a concise factual answer and a broad comparative recommendation.
FAQ and conclusion
What is a good AI search citation precision score?
There is no universal target. Compare engines, topics and time periods using identical rules. For sensitive or rule-based queries, demand stronger source authority and full claim support than for low-risk explanatory queries.
Can citation recall be measured when an engine hides retrieval logs?
Yes, but call it visible citation or claim-evidence recall. It measures surfaced evidence, not everything the system retrieved or used internally.
Should my own domain be in the relevance baseline?
Include it only when it genuinely supports a required or acceptable claim. A baseline is an evidence standard, not a list of desired brand outcomes.
Conclusion: Citation precision tells you whether displayed sources deserve their place; recall tells you whether essential evidence made it into view. Score both at source and claim level, disclose incomplete reference sets, and keep retrieval, selection and visibility distinct. That discipline turns citation tracking from a count of links into an auditable assessment of answer evidence.
