Reading Your Results

Checking source quality before trusting a citation

Why citations need a second look

When an AI model backs up a claim with a source, it is worth remembering that the citation itself is generated text, not a guaranteed fact. A model can cite a URL that does not exist, or cite a real page that has nothing to do with the claim it is supposedly supporting. Both failure modes read identically on the page: a confident sentence followed by a link. The only way to tell them apart is to actually check the source.

What the check looks at

The Source Quality view runs each citation through a short pipeline rather than a single pass/fail test.

  • Does it exist? The link is checked to see whether it resolves to a real, reachable page, an unreachable one, or nothing at all.
  • Is it on topic? Assuming the page exists, its content is compared against the claim to see whether it actually covers the same brand, category, or subject, rather than something only superficially similar.
  • Was it used well? Even a real, on-topic source can be stretched past what it actually says, so this checks whether the citation supports the claim it was attached to or only loosely relates to it.

These combine into an overall read for each citation checked, ranging from a well-supported claim to one that could not be verified at all.

What "checked" means in practice

This is a sampled check, not an audit of every citation in a wave. Running the full pipeline against every link a model produces would be slow and expensive, so a handful of citations are checked per run rather than all of them. Treat the result as a spot check on how trustworthy a model's sourcing tends to be for this wave, not a guarantee that every uncovered citation is fine.

Related reading

This same scepticism is worth applying to the numbers built from entities, not just the citations attached to a claim: an entity that was never cleaned up or a co-occurrence pattern you haven't read the underlying responses for can look just as confidently supported as a fabricated citation, until you actually check it.

All help articles · Glossary