Stratify by buyer intent
Separate unbranded provider discovery, comparison, problem-led purchase, educational research and branded entity checks. A brand query such as “Tell me about Company X” should not raise the unbranded discovery rate.
There is no single universal ranking position across AI platforms. A defensible measurement plan defines the questions, records the context and examines repeat observations rather than chasing a score.
Measure AI visibility by testing a defined panel of genuine buyer questions, recording the completed answers, classifying brand mentions, provider recommendations and cited sources, and calculating rates with explicit denominators. Compare like with like across time and supplement those observations with genuine search impressions and business outcomes where available.
An AI chatbot does not offer one universal ranking position for a business. Responses differ by question, platform, language, session, source availability and time. This guide explains how to interpret and calculate the observations. To learn how to capture answers first, see How to Check AI Visibility. The Monitoring service is a different commercial page for ongoing work.
Start with a blank record: Download the free AI visibility observation template (CSV). It contains column headings only—no real or simulated brand results—and is suitable for a spreadsheet.
Different results answer different questions. The denominator must describe the specific population sampled, not an imagined total of every user or model interaction.
| Metric | Calculation or record | Interpretation and limits |
|---|---|---|
| Sampled brand mention rate | Comparable completed responses mentioning the brand ÷ all comparable completed responses × 100. | Reports appearances in the sampled panel; not total AI market share. |
| Sampled recommendation rate | Comparable completed commercial-selection responses explicitly recommending the provider ÷ comparable completed commercial-selection responses × 100. | Exclude general informational and branded control prompts unless reporting them separately. |
| First-party citation response rate | Completed responses citing at least one URL on the brand's site ÷ completed comparable responses × 100. | Count the response once even if it contains several own-domain links. |
| Cited-link source mix | Links in a specified source category ÷ all observed cited links in the same panel × 100. | Make clear whether repeated links, distinct URLs or domains are counted. |
| Competitor shortlist presence | Commercial-selection responses recommending a named competitor ÷ compatible commercial-selection responses × 100. | One response may recommend several competitors; category rates can sum above 100%. |
| Search feature impressions | Reported directly by an eligible verified search property, with its stated dimensions. | Not the same as manually sampled recommendations or third-party AI assistant exposure. |
| Referrals and qualified leads | Actual visit/enquiry events supported by analytics and attribution evidence. | Never infer a sale from a brand mention or source citation alone. |
Rule: when a prompt is not completed, don't mark it as “not mentioned.” Keep completion status separate and disclose what was excluded. Zero, missing, inaccessible and not run describe different situations.
There is no universally valid test count. Start with a manageable range of genuinely different buyer needs, then choose how many repeated runs are feasible for the markets and platforms that matter.
Separate unbranded provider discovery, comparison, problem-led purchase, educational research and branded entity checks. A brand query such as “Tell me about Company X” should not raise the unbranded discovery rate.
Store a prompt ID, wording, date, language, target region and purpose. Record platform and mode when observable, plus run number. Later wording changes should create a new version or cohort.
Three repeated runs can reveal some variability, but are not statistically representative of millions of possible user sessions. Report uncertainty plainly; larger and better-designed samples may be needed for stronger inference.
Compare only sufficiently similar completed runs. Report platform-by-platform and intent-by-intent results before considering any aggregate. Record unavailable or interrupted tests separately.
Do not create thousands of prompts mechanically from a keyword export. Group naturally related phrases into distinct research intents, then use the buyer's actual questions. This supports topical coverage without creating artificial sampling volume.
Hypothetical numbers only—not MentionGauge or client results: suppose eight unbranded commercial buyer questions are tested three times each on one accessible platform, giving 24 compatible completed answers.
| Measured event | Example count | Calculation | What can be said |
|---|---|---|---|
| Completed responses | 24 of 24 planned | 8 prompts × 3 runs = 24 | This defines the tested denominator. |
| Brand mentioned | 6 of 24 | 6 ÷ 24 = 25% | Brand mentioned in 25% of this particular sample. |
| Brand explicitly recommended | 2 of 24 | 2 ÷ 24 = 8.3% | Sampled provider-shortlist rate, not a universal rank. |
| At least one own-domain source cited | 9 of 24 | 9 ÷ 24 = 37.5% | Response-level citation rate; not a recommendation rate. |
| Missing responses | 0 in this example | 24 planned − 24 completed = 0 | In practice, incomplete responses must be reported separately. |
Suppose a separate sample of 40 displayed links includes 12 own-domain links. A cited-link share would be 12 ÷ 40 = 30% of the observed links. It is not the same quantity as the 37.5% response-level citation rate above, and it would change if you deduplicated URLs instead of counting link appearances.
An AI answer can cite a company's website for a technical definition without recommending that company. Likewise, a provider recommendation may appear without a directly observable citation. Keep these events in separate columns.
Some answers display several references, others display none, and a link may support a definition rather than the selection of a provider. Interpret links according to the question and their actual content.
Open the link and verify the target, the relevant claim, its service applicability and whether it is still available. A link merely appearing near an answer does not prove it caused the model's response.
Separate the brand's own website, independent publications, directories, professional profiles, third-party user content and competitor sites. Avoid counting a company's own profile as independent coverage.
In commercial-selection answers, record every recommended competitor as a separate indicator. A single answer can recommend several businesses, so individual recommendation rates may total more than 100% if added together.
If the interface does not show citations, record that field as unavailable or absent rather than inventing URLs. Do not confuse a cited source with a website referral or a qualified enquiry.
Source investigation is part of our manual research methodology; buying a recurring assessment belongs on the monitoring service page. This guide is for the reader's own calculation and interpretation.
Yes, Google provides a dedicated Search generative AI performance report. Google announced this in June 2026 and states that, as of August 31, 2026, these insights were rolled out worldwide. It reports impressions of links to your verified site's pages in supported Google AI search features, including AI Overviews and AI Mode. It is not the same as a manual survey of ChatGPT or other independent AI platforms.
Reported site impressions, linked pages, countries, devices and date trends, using Google's own reporting definitions. Use URL and time dimensions to understand which pages are appearing; don't assume an impression equals a recommendation or a click.
Google says properties may not see the dedicated report if they lack sufficient generative-feature impressions. Eligibility and indexing can also be affected by the site's Search generative AI control and ordinary Search requirements. Do not invent a zero-impression series when the report is unavailable.
The Google report does not directly measure whether a particular independent assistant cited a brand or why a provider was shortlisted in a sampled response. Reporting units and aggregation may differ between property-level and page-level views.
Keep Google feature impressions, ordinary organic clicks, user analytics, manually captured AI answers and qualified enquiries on separate lines. Link them conceptually to the buyer journey, not through an unsupported single “AI visibility score.”
Official sources: Generative AI performance report (Search), Google's June 2026 announcement, and Search generative AI control. The website's own property data must be checked directly before making performance claims.
A before-and-after measurement can identify an observation change, but it does not necessarily explain it. Preserve a stable cohort and record any changes that make the period comparisons less trustworthy.
| Comparison issue | Practical solution |
|---|---|
| Question wording changed | Label a new prompt version and compare only a stable core group. |
| Platform or mode changed | Disclose the break, show results separately and avoid presenting them as equivalent. |
| Different completion counts | Show each period's numerator and denominator, with missing-run reasons. |
| Branded and unbranded mixed | Segment identity accuracy from new-customer discovery and provider recommendations. |
| Website content updated | Record URL, date, change, validation and subsequent data, without assuming causation. |
| Other sources changed | Note accessible third-party or competitor updates, but avoid claiming access to hidden model selection. |
For a small panel, report the change as a count with context—for example, two recommendations in 24 completed tests in period A versus four in 24 in period B. Don't call that a statistically established improvement or assign it to a single page edit without stronger controlled evidence.
Use the free CSV observation template to record genuine tests in Excel, Google Sheets or your preferred system. It contains no fabricated results or platform data. Each run should occupy its own row.
Enter a prompt ID, exact buyer question, intent, target brand, intended platform, mode, language, region and run number. Keep the prompt version stable across repeats.
Record actual timestamp, completion status, observed mention/recommendation flags, own-domain citations, competitor entities, cited URLs, captured response and reviewer notes. Leave unavailable evidence genuinely blank or explicitly unknown.
The sections above define the measurement units. If you need someone to collect and interpret ongoing source evidence, see our monitoring scope.
Some tools combine prompts, brands and appearances into a proprietary index. Such a score may be useful internally if its definition is disclosed, but it is not an official cross-platform ranking. Require transparent source counts and denominators before comparing scores.
“Share of voice” is used inconsistently. It might mean cited-link share, recommendation share or another defined measure. Our sampled mention rate counts completed answers containing the brand divided by comparable completed answers; its denominator is explicit.
No. A cited page might support a definition, claim or comparison without being offered as a provider. Citations and provider recommendations should be coded as separate events.
Google's dedicated generative AI performance report shows eligible link-impression data for AI Overviews and AI Mode. Availability of visible data depends on the verified property's actual impressions and settings; it is not a ChatGPT visibility report.
Use the same commercial buyer question panel, record each competitor actually recommended, and state whether the result is a mention, shortlist entry or citation. Compare providers using compatible samples, not isolated screenshots.
Zero completed mentions in the selected panel means no mentions were observed in those tests. It does not prove the business is absent from every AI answer. A panel that was never run has missing data, not a meaningful zero.
Choose a cadence that matches the commercial decision, platform access and workload. Preserve prompt wording and log relevant site or platform changes. A calendar interval alone does not guarantee measurable movement.
No. Valid structured data should reflect visible page facts and keyword wording should be natural. Google's AI Search documentation says special AI-only schema is not required, and appearing as a supporting link is not guaranteed.
Request a scoped manual audit to understand how your business appears across relevant AI-assisted search experiences and what can be improved.