Query tests proposed
What makes an AI citation useful?
A citation can be easy to find and still fail to support the claim next to it. This desk study proposes scoring visibility and claim support separately.
An AI answer says a museum is free on Sundays and links to its visitor page. The page lists Sunday opening hours but says nothing about admission.
- Visibility
- The source appears in the answer.
- Claim support
- The source does not verify free admission.
- Visibility
For each query, record whether the answer cites the page.
- Claim support
For each cited claim, record whether the source supports the statement.
- Compare
Keep the two scores separate so a high visibility score cannot hide weak support.
Question
When an AI answer cites a page, can a reader verify the claim it appears to support?
Method
The desk study assembled a register of 53 entries and 24 query concepts, each with proposed Dutch and English wording. It proposed two scorecards, one for source visibility and another for whether the cited evidence supports the associated claim.
Earlier work
In June, the Anna work introduced selective research triage and bounded follow-up. August work added source and grounding labels and exposed a counting error: result blocks had been treated as queries. These were separate engineering and research efforts. The October study proposed scoring visibility and support separately.
What the study produced
A test design that can distinguish a frequently cited page from a page that provides strong support. The two scorecards let reviewers record disagreement between those outcomes.
Limits and next check
The query tests were proposed. A follow-up should run a fixed bilingual panel, record which sources appear and have reviewers independently judge support at the claim level.
Source study · Agentic citations research report, 15 October 2026
Related reading · ALCE citation evaluation. A benchmark for assessing citation quality.
Evidence and next check
Public evidence review is pending. The next step is to run the fixed Dutch and English query panel and have independent reviewers score claim support.