Review the exact claim against the cited passage. A relevant page can still fail to support the sentence beside its citation. Agencies and content reviewers should distinguish source relevance from actual support before repeating an AI answer in a client report.
Use a claim-level rubric
Split the answer into statements small enough to check. “The platform supports exports and has the lowest price” contains two claims and may require different evidence.
Assign one of five labels: supported, partly supported, contradicted, not established, or unavailable. Add the passage or section that led to the label. “Unavailable” means you could not inspect the evidence; it is not the same as false.
The ALCE research paper evaluates citation quality separately from answer correctness. The worksheet here is an editorial review method, not a reproduction of its benchmark or a current provider performance estimate.
An illustrative review
Suppose an answer says, “Acme includes unlimited seats on every plan,” and cites a pricing page that says only “Team plan: unlimited seats.” The source supports a narrower claim. Mark the original statement partly supported and record the missing plan scope.
Do not quietly rewrite the answer in the evidence field. Preserve the observed sentence, then put the corrected wording in a separate reviewer field. That makes the discrepancy inspectable.
A practical review sequence
- Save the prompt, provider surface, collection time and full answer.
- Identify the citation attached to the selected claim.
- Open the original source, following redirects while retaining the returned URL.
- Compare the entity, action, conditions, geography and date.
- Record a support label and a short justification.
- Ask a second reviewer to adjudicate consequential disagreements.
Keep numerical claims separate from qualitative claims. A source saying a product is “popular” does not establish a percentage, sample size or market position.
Report the denominator
If you inspect ten claims, report support among those ten reviewed claims. Do not describe the result as the factual accuracy of every answer in the panel. State how claims were selected and how many sources could not be opened.
Source content can change after collection. Record the observation time and review time separately. Keep only the excerpts needed for review and respect source rights; do not republish complete third-party articles.
Use the citation quality guide for frequency and source-type context. Your next action is to select one consequential claim from the citation tracking workflow, open its source and record whether that source establishes the full statement.
Fact-checked: 17 October 2026. Examples are illustrative; no customer results are reported.