Make AI citations point to evidence, not just a source name
Preserve source IDs and page references, attach citations to individual claims and verify that the cited passage actually supports them.
In brief
A citation is useful only when the referenced material exists and supports the nearby statement. A model can produce a plausible document title or link without having verified it.
Use references that a reader can inspect, especially for dates, prices, compatibility and policy exceptions.
Keep the evidence addressable
For a PDF, preserve the printed page or section and distinguish it from the viewer’s page index if they differ. For a webpage, use its specific article or documentation URL. In a retrieval system, keep the source identifier linked to the real document rather than allowing the model to invent new identifiers.
Attach support where it belongs
A paragraph containing a price, release date and technical limit may need more than one source. Do not attach one citation at its end and assume it supports everything. Ask the model to split claims when evidence comes from different places.
Reject plausible but unsupported references
Open the source yourself or through a permitted retrieval workflow. Check whether it refers to the right product and date. If the material is unavailable or does not contain the claim, mark that gap and remove or qualify the claim instead of polishing the citation’s appearance.
Before you act
A valid URL is not proof of support, and a citation to a general homepage is not the same as a specific advisory or product document.
What to do next
- Provide identifiable source passages with URLs or page references.
- Request a citation for each substantive claim and explicit gaps where support is missing.
- Open the references and compare the actual passage, date and scope with the answer.
