Visibility is a set of observations
AI search visibility should not be reduced to one unexplained score. A useful framework records whether a brand appears, whether a visible source supports the answer, which pages are used and whether a person reaches the site. Each layer has a different denominator and evidence type.
Define the measurement universe first
Before collecting answers, state the brands, topics, markets, languages, platforms, product modes and dates in scope. Freeze the exact query list for the reporting period. If questions are added or removed, publish the change and avoid comparing the new rate directly with an earlier denominator.
Core answer-layer measures
| Measure | Formula | Required evidence |
|---|---|---|
| Mention rate | Answers naming the brand ÷ eligible answers | Full capture, query and context |
| Cited-mention rate | Answers naming and visibly citing the brand ÷ eligible answers | Capture plus source destination |
| Owned-source presence | Answers citing an owned URL ÷ eligible answers | Resolved canonical source URL |
| Source diversity | Count of unique cited domains and pages | Normalized source register |
| Query coverage | Query groups with a relevant brand/source observation ÷ groups tested | Predefined intent taxonomy |
Keep the commercial layer separate
Observable AI-referral sessions use website sessions as their denominator. Enquiry and qualified-lead rates use approved event definitions. Never divide crawler requests by captured answers or present a brand mention as a visit. Connect layers only when first-party evidence supports the path.
Collection protocol
- Assign a stable ID to every exact question.
- Record platform, mode, account context, market, language and timestamp.
- Save the full answer and all visible sources.
- Resolve source URLs and record their canonical destinations.
- Apply mention and citation labels through a documented QA rule.
- Retain failures and source-free answers in the denominator when eligible.
- Repeat on the scheduled cadence without selecting favourable reruns.
Report uncertainty visibly
Show “6 of 30 eligible answers” beside 20%. State sample size, collection dates and any unavailable platform state. Small changes in a small set should not be described as a trend.
A quarterly report structure
- Executive summary tied to business decisions.
- Scope and frozen method.
- Answer-layer counts and rates.
- Most-used owned and independent sources.
- New, lost and contradictory observations.
- Observable referral and lead activity in a separate section.
- Page-level actions with evidence and owners.
- Limitations and methodology changelog.
How to use the result
Look for repeated gaps across a commercial topic, not one surprising answer. A missing service definition suggests content work; a contradictory company identity suggests entity remediation; inaccessible or duplicate pages suggest technical work; weak independent support suggests authority work. The metric identifies where to investigate, not which tactic to buy automatically.
Primary context
Search and answer products continue to change, so platform documentation and dated live observations should be reviewed together. The framework remains useful because it exposes its units, evidence and limits instead of depending on one interface.
Where to go next
Use the Nexus AI Visibility Framework to place this topic inside a complete measurement system, or review the AI Visibility Audit scope.
13 Jul 2026 — Expanded with an independent structure, examples, implementation guidance and primary references.