1. Define the question set
Group prompts by job: category discovery, comparison, alternatives, implementation, and local or regulated intent. A large unstructured list is difficult to interpret.
Measurement reference
AI search analytics that explain what changed
AI search analytics is not a single visibility score. It is a repeatable way to connect buyer questions, AI answers, recommendations, citations, competitors, and the content that influenced the result.
The useful question
“Which buyer questions did we lose, which competitor won, and what evidence did the answer use?”
A dashboard becomes actionable when every movement can be traced back to a prompt cluster, an answer classification, or a source pattern.
Metric dictionary
Six metrics, six different decisions
Avoid collapsing presence, recommendation, citation, and accuracy into one flattering number. Each metric answers a different operational question.
| Metric | What it asks | Signal | Next decision |
|---|---|---|---|
| Prompt coverage | How many of the buyer questions you care about are represented in the tracked set? | Coverage | Add missing intents before changing copy. |
| Mention rate | How often does the brand appear at all for the tracked answers? | Presence | Separate recognition from recommendation. |
| Recommendation rate | How often is the brand presented as a suitable choice rather than merely named? | Commercial visibility | Read the rationale and buyer fit. |
| Recommendation share | What portion of observed recommendations goes to your brand versus named competitors? | Competitive position | Use the same prompt set and sampling rules. |
| Citation quality | Which pages and external sources support the answer, and are they current and relevant? | Source influence | Fix the source gap, not only the target page. |
| Answer accuracy | Are the facts about pricing, features, audience, geography, and limitations correct? | Brand safety | Track wrong facts as a separate failure mode. |
Operating loop
Make the data explainable before making it bigger
The goal is a clean experiment, not a giant prompt count. Keep the question set stable long enough to distinguish a content change from normal answer variance.
2. Freeze the sample
Record providers, markets, language, date range, prompt wording, and run frequency. If those change together, the trend is not explainable.
3. Classify outcomes
Mark each answer as absent, mentioned, shortlisted, recommended, cited, or inaccurate. Do not count every mention as a win.
4. Inspect the evidence
Save the answer text, cited URLs, competitors, and rationale. The most useful finding is often the source or fact that caused a competitor to win.
5. Ship one measurable change
Update a page, source profile, comparison, proof block, or technical element. Keep the change narrow enough to evaluate.
6. Re-run and annotate
Compare the same prompt cluster after a meaningful interval and log what changed. AI answers vary; a single run is not a verdict.
Do not report these as the same thing
A crawler visit is not a citation. A mention is not a recommendation.
Technical access, search visibility, answer inclusion, recommendation, citation, and conversion are related but separate signals. Keeping them separate is what makes an analytics program credible.
- Keep provider, country, language, and prompt intent visible.
- Store the exact answer and cited sources for important changes.
- Report accuracy and competitor displacement alongside share.
- Annotate page releases and external source changes.