An AI citation is useful when the referenced source exists, the user can access it and the cited passage supports the associated claim. A plausible title or working URL is not sufficient. Evaluate citations at claim level and preserve the relevant source version so reviewers can understand what evidence the system actually used.
Let’s plan your AI citation checks.
Bring your workflow and the decisions you need to make. We’ll help you clarify requirements, review the trade-offs and define a practical next step.
Define the citation contract
Specify what the answer presents: document title, section or page, version and a link to the relevant location where supported. Separate a source retrieved during generation from a source that genuinely supports the final wording.
For a hypothetical internal policy assistant, the answer may combine a current policy with an obsolete FAQ. The citation interface should make version and authority clear rather than giving both sources identical visual credibility.
Use a citation evaluation worksheet
| Check | Reviewer question | Failure example |
|---|---|---|
| Existence | Does the source identifier resolve? | Invented file or page |
| Access | Can this user open it? | Citation reveals an inaccessible document title |
| Support | Does the passage support this claim? | Related topic with different conditions |
| Scope | Are qualifications preserved? | Exception omitted from the answer |
| Currency | Is the applicable version used? | Superseded policy presented as current |
| Attribution | Is synthesis distinguished from quotation? | Generated wording shown as exact source text |
Record results separately from general answer quality. An answer can be correct through coincidence while citing the wrong evidence.
Design for uncertainty
When support is incomplete, qualify the answer or say that the available sources do not establish it. Do not manufacture a citation to make the response appear complete. A useful next step might be a link to an authorised owner or an explanation of the missing evidence.
If the source is a scanned document, verify page mapping after extraction. OCR text order and PDF page labels can differ. Test a source update so citations do not silently point to an unrelated passage in a new version.
Make review efficient
Let an authorised reviewer inspect the relevant passage near the claim. Preserve source references and evaluation labels without retaining unnecessary sensitive content in general logs.
Use a varied test set: direct lookup, comparison, exceptions, conflicting sources and questions the collection cannot answer. Include cases where the correct behaviour is to decline a definitive claim.
Citation questions
Does adding more citations make an answer more trustworthy?
Only when they provide relevant support. Extra unrelated references can make verification harder.
Can citation quality be assessed automatically?
Automated checks can find selected failures, but representative human review remains useful for meaning, qualification and authority.
Use the AI evaluation dataset guide and grounding guide. Discuss the evidence contract with Drema's generative AI team.
