Measurement · OWNER’S GUIDE
Measure AI visibility without inventing an AI score
Measure one defined question at a time. Technical access, observed citations, factual accuracy and qualified inquiries are different outcomes. Show the underlying counts and the limits of the sample instead of compressing them into a mysterious score.
Write the denominator before collecting results
Illustrative numbers only — not Tarawa or customer performance.
| Measure | Example calculation | Interpretation |
|---|---|---|
| Access coverage | 9 readable pages / 10 planned pages | One page remains unavailable; this is not an indexing rate. |
| Unbranded citation observations | 3 answers citing the domain / 12 completed observations | 25% in this defined question sample, not market share. |
| Checked fact accuracy | 18 correct assertions / 20 evaluated assertions | Two assertions need correction or further investigation. |
| Qualified inquiries | 4 qualified / 7 confirmed requests | An internal qualification rate, using an agreed definition. |
Record unavailable and skipped observations separately rather than silently excluding inconvenient outcomes. If the denominator changes between periods, explain the change before showing a percentage comparison.
Keep the observation conditions visible
Choose a small panel of real customer questions. Include branded facts, unbranded needs, policy questions and questions the sources cannot answer. Record product, displayed model or mode, locale, date, exact prompt, whether search occurred and cited URLs. Repeat in fresh sessions if repeatability matters.
Supplying the website URL is useful for testing retrieval, but it cannot establish unaided discovery. An API test is labeled as an API test. A response without web search is not quietly mixed into a search-enabled citation sample.
Bing’s citation report and Google’s generative-search impressions have their own definitions. Keep those reports separate from your manual question panel. Bing: AI Performance reporting; Google: Generative AI performance report.
Turn the evidence into a decision
Start with factual errors and broken customer paths. If an answer cites an old policy, identify the source and determine whether it needs correction. If the answer is accurate but inquiries are weak, investigate offer fit and the destination experience before rewriting everything for “AI.”
For a before-and-after review, preserve the question set and record other changes, such as model releases or advertising activity. A small sample can guide investigation; it cannot isolate causation on its own.
Download the blank observation log (CSV). The next step is to agree which decision the measurement should inform and what evidence would be enough to change it.
Sources and review
Primary references checked November 11, 2026. Platform rules can change; check the linked instructions before acting.
How we research and maintain this content · Send a correction