Define the measurement decision
Prefer a review matrix over a mysterious single number. Teams can weight fields later, but the raw judgments should remain auditable.
Measurement method
| Step | Action |
|---|---|
| 1 | Resolve the cited URL and store its canonical form |
| 2 | Check that the page supports the specific claim in the AI answer |
| 3 | Classify the source as official, primary research, expert secondary, community, or other |
| 4 | Record publication or update date when freshness matters |
| 5 | Separate missing citation data from a successful answer with no citation |
Worked calculation or observation
A recent official product document may be the best source for a feature claim, while an independent benchmark may be better for a performance comparison. Quality depends on the claim, not only the domain.
Evidence to retain for AI search citation quality
Keep these fields with the decision:
- metric name
- numerator
- denominator
- scope
- platform
- Prompt Group
- success state
- observation date
- formula version
Limits and UnderAI's role in AI search citation quality
A scoring framework does not prove that a source caused an answer or that the same URL will be cited again.
UnderAI preserves citation domains and articles at project and snapshot levels; analysts can apply a transparent quality rubric to those records.