AI Visibility Metrics: Mentions, Citations and Coverage
Updated
An AI visibility number is useful only when you know what was counted. A brand mention, a citation, a failed request, a Google impression and a website visit describe different events. Treating them as interchangeable produces misleading reports.
This guide defines a measurement worksheet and works through one hypothetical dataset. The example is deliberately small enough to audit by hand. It is not a Ghostd benchmark, a customer result or an estimate of all AI searches.
Begin with the observation contract
For each check, record the question, platform, date, answer context and outcome. Where available, preserve the model, method and response settings too. Store the answer or its evidence references so an observation can be inspected.
Use three distinct states:
| State | What you know | Interpretation |
|---|---|---|
| Usable answer with a brand mention | A returned answer includes the brand under the stated method | Observed presence in this sample |
| Usable answer without a brand mention | The answer was observed and did not include it | Observed absence in this sample |
| Failed, unavailable or uninterpretable answer | The check did not establish presence or absence | Availability limitation, not a confirmed absence |
Do not discard an observed negative result merely because it is inconvenient. Conversely, do not turn an unavailable result into a negative observation.
Our AI visibility definition explains the broader purpose. The calculations below apply to recorded samples, not population-level exposure.
A worked dataset with explicit units
Suppose a fictional audit attempts 120 answer checks in a defined period. Ninety return usable answers and 30 are unavailable. Of the usable answers, 18 mention the brand and nine contain a recorded citation to its site.
For a separate citation-share calculation, suppose the audit records nine own-domain, 12 competitor-B and six competitor-C answer–domain citation units. Count a domain at most once per answer for this example. An answer can cite multiple domains, so these units are not a count of unique people or visits.
| Metric | Calculation | Result | Meaning |
|---|---|---|---|
| Usable-answer coverage | 90 ÷ 120 | 75% | Availability of the attempted checks |
| Sampled mention rate | 18 ÷ 90 | 20% | Brand presence among usable answers |
| Own-site citation frequency | 9 ÷ 90 | 10% | Usable answers citing the brand’s domain |
| Selected citation share | 9 ÷ (9 + 12 + 6) | About 33.3% | Own share of the specified answer–domain units |
The mention rate is not 18 ÷ 120. That 15% figure silently treats 30 unavailable answers as observed absences. Report 20% sampled mentions alongside 75% usable-answer coverage instead.
The 20% mention rate and 10% citation frequency should not be added. They can overlap: an answer may mention the brand and cite its site.
Define citation evidence before calculating a rate
A returned source list, a source retrieved by a tool and an explicit citation in the final answer can represent different stages. Name the stage you actually observe.
If your method measures final-answer citations, do not quietly switch to every retrieved source when final citation evidence is unavailable. If only returned source lists are available, report that observation with its limitation rather than relabeling it as a confirmed citation.
The citation-rate article provides related context. For any calculation, retain the definition with exported data so the percentage does not lose its meaning in a presentation.
Share of voice depends on the selected universe
The example’s 33.3% citation share covers three selected domains and the specified citation units. Adding another competitor changes the denominator, even if the brand’s answers do not change.
A mention-based share is a different calculation. It needs its own unit, competitor selection and rules for multiple mentions. Neither version automatically measures every competitor in the market or every buyer’s exposure.
Read AI share of voice with this boundary in mind. When comparing periods, preserve the competitor set or label the methodology change.
Compare like with like before describing a change
A defensible change report needs comparable recorded observations. Ask:
- Were the questions the same?
- Was the platform and relevant answer context comparable?
- Were both observations usable?
- Are the dates and evidence references available?
- Did the extraction method or question coverage change?
A new model, a different question set or improved citation extraction can change a number without changing the brand’s underlying presence. Label those as measurement changes until comparable evidence supports another conclusion.
Even a valid comparison does not prove why an answer changed. “The newer recorded answer did not cite the site” is narrower and more defensible than “a competitor took our traffic.”
Keep Google impressions and site traffic separate
Google’s current Generative AI performance report describes impressions for supported Google Search features, including AI Overviews and AI Mode. The report’s page and property aggregation can differ, and unavailable values may become zeros in exports.
Those impressions are not this guide’s sampled mention rate. They are also not referral visits or conversions. Use appropriate site analytics for visits and outcomes, and preserve attribution limitations.
The AI Search statistics resource demonstrates how a dated study’s scope affects interpretation. Do not use a historical study’s click rate as your site’s measured rate.
A reporting note you can copy
“During the stated period, we attempted ___ checks for ___ questions across ___. ___ returned usable answers. The brand appeared in ___ of those usable answers. Citation frequency uses ___. Competitor share covers ___. The results describe the recorded sample, not all users or searches. Unavailable observations were reported separately.”
Fill every blank before reporting a result. If a field is unknown, make the unknown visible; do not replace it with a plausible assumption.
For evaluating whether a tool preserves this information, use the visibility-tool checklist. For deciding what to improve next, use the content optimization workflow.
Editorial note: This update was prepared with AI assistance. All numerical examples are hypothetical. The original contribution is the observation contract, explicit denominator worksheet and reporting template; none is a promise that a particular edit will improve rankings or traffic.
About the Author

Todd Wandtke
Founder & CEO, Ghostd
Todd is the founder of Ghostd and a specialist in generative engine optimization (GEO). He helps B2B brands build visibility inside AI-powered search engines like ChatGPT, Perplexity, Gemini, and Claude—so they get found before the click ever happens.
Connect on LinkedIn