Brand mention rate is the percentage of valid measured answers that name the business. Owned-site citation rate is the percentage of answers with inspected source links that cite the business’s website. The denominators can differ: an answer may be available while its source links remain unassessed. Report the numerator and denominator for each rate. Recommendations, prominence and factual accuracy are additional observations, not substitutes for either rate.
A mention is different from a source citation. An answer can name or recommend a business without linking to its website. An owned-site citation links to the business’s own website; a third-party citation links to another publisher. Record those links and inspect their destination. Missing, blocked or unassessed observations are reported as gaps, never silently entered as zero.
What does each AI visibility report actually measure?
Start with the question you need answered. A citation count, an impression, a session and an enquiry describe different events. Put them beside one another with their source and period; adding them into one “AI visibility score” hides what happened.
| Evidence source | What to record | What it cannot establish |
|---|---|---|
| Your direct AI tests | The exact question, product, settings, date, answer and inspected source links. Score mentions and owned-site citations independently. | How every customer sees the business, or its share of all AI answers. You chose the questions and conditions. |
| Google Search Console: generative AI | Impressions of links in AI Overviews and AI Mode, with page, country, device and date breakdowns. | A separate click or conversion total, or a count of ChatGPT/Gemini-app answers. These impressions are already within the Web search report. |
| Bing Webmaster Tools: AI Performance | Reported citations and cited pages across supported Microsoft and partner experiences; the selected period and available grounding-query sample. | Recommendation position or an exhaustive list of customer questions. A grounding query is a retrieval phrase, not necessarily the question a person typed. |
| GA4: acquisition | Recorded sessions and events under the chosen source/channel definition. The AI Assistant channel covers recognised assistant referrals; Google AI Overviews and AI Mode belong within Organic Search. | Every AI-influenced visit, the preceding answer, or qualified demand. A recorded visit and a genuine commercial enquiry need separate evidence. |
| Your enquiry and sales records | Received enquiries, qualification, won work and contribution after delivery costs. Keep test submissions and spam separate. | That AI caused a sale merely because a customer also visited from an assistant or said they had used one. |
Two easy Google reporting mistakes: do not add generative-AI impressions to Web impressions, and do not force page-row totals to equal the property chart. The report uses different aggregation for pages and properties. Dates are in Pacific Time. Preserve the exported period and filters. Google also notes that unavailable values can export as zero, so check the interface before interpreting a downloaded zero as measured absence. Read Google’s definitions.
Why not just measure rankings or traffic?
Rankings describe search placement and analytics describes recorded visits or events. Direct AI tests capture answers to an agreed panel of questions. Available native reports add a different view: Google’s generative-AI feature report and Bing AI Performance describe activity within their published scopes. Use these sources together while retaining each report’s dates, definitions, coverage and limitations. Neither a test mention nor a platform impression is a customer enquiry.
What makes the measurement trustworthy?
Discipline. AI answers vary between sessions, so a single run is an observation rather than a reliable trend. Use a fixed prompt set and a fixed direct-test engine set: ChatGPT, Gemini, Claude and Perplexity. Record the product, search setting, date and location, and use fresh sessions with previous conversation and personal memory excluded wherever the product permits. Platform-reported evidence is kept separate: Google now provides dedicated generative-AI impression reports alongside overall Search performance. Bing Webmaster Tools AI Performance data, where available, has its own scope and is not mixed into the direct-test score.
How do you build the prompt set?
Build the panel from actual customer questions where available, then agree its size, services, locations and exclusions. Mark every prompt as branded if it names the business and non-branded if it does not. Hold the panel stable for comparison and version any changes. If a group has not been tested, report it as untested. Presence on non-branded test questions is useful visibility evidence; it is not a measurement of real customer discovery.
How do you score an answer?
For each completed answer, record whether the business is named, whether it is recommended, its prominence and any factual issues. Inspect the source links separately and record owned-site and third-party citations. Keep competitors’ mentions distinct from their source links too. Each score must point to the dated answer and assessment evidence, with unavailable or unassessed fields clearly marked.
| What you actually checked | Mention | Owned-site citation |
|---|---|---|
| The answer names the practice; its inspected sources contain no link to the practice’s website. | Yes | No |
| The answer does not name the practice; an inspected source links to its website. | No | Yes |
| The name appears only inside a third-party source page; the answer body does not name it and no source links to its own website. | No | No. Record the third-party evidence separately. |
| You read the answer but could not inspect its sources. | Assess from the answer. | Unassessed, not No. |
A worked example: two 50% rates with different coverage
Illustration only, not an Answari or client result. Plan eight observations. Six answer bodies are available and assessed for mentions; three name the business. Source links have been inspected for four answers; two link to the business’s own website.
- Mention rate: 3 ÷ 6 = 50%. Mention coverage is 6 of 8 planned observations, or 75%.
- Owned-site citation rate: 2 ÷ 4 = 50%. Citation coverage is 4 of 8 planned observations, or 50%.
- The gaps: two mention assessments and four citation assessments remain unknown. The equal rates do not mean equally complete evidence.
Using eight as both rate denominators would produce 37.5% and 25%, silently treating missing assessments as negative findings. Keep the planned count for coverage and the assessed count for each rate. When nothing has been assessed, show Unassessed, not 0%.
Run this calculation separately for questions that name the business and questions that do not. If comparing competitors, apply the same definition to every business. Several can appear in the same answer, so their mention rates need not add up to 100%.
branded and non-branded results reported separately, with coverage and counts.
Brand mention rates across competitors need not total 100% - a single answer can name more than one business, so treat each share as that business’s own presence rate, not a slice of one pie.
The mentions-and-citations protocol, end to end
- Agree the questions, products, settings, repeat schedule and scope before capture. Mark branded and non-branded prompts.
- Run every question on every engine in fresh sessions, on a cadence agreed for the panel.
- Assess each captured answer for mentions, recommendation, prominence and accuracy. Inspect source links separately and retain the evidence.
- Compute brand mention rate and owned-site citation rate using their assessed denominators. Show counts, coverage and unavailable observations; keep branded and non-branded groups separate.
- Compare against your baseline and your competitors, question by question - because visibility is won and lost per question, the per-question view tells you where to work next.
- Corroborate with AI-referral attribution in analytics: report identifiable visits and enquiries from assistant surfaces separately. Visits do not establish qualified demand, and attribution gaps mean analytics will not capture every discovery journey.
What does good movement look like?
Useful progress may include corrected facts, clearer source links, more consistent mentions or attributable enquiries. There is no required sequence, and a website change followed by a different answer does not establish causation. Compare repeated observations under recorded conditions, include unchanged or worse results, and report changes in test coverage. Commercial outcomes require their own evidence.
This protocol is exactly what an Answari engagement runs: the Visibility Audit establishes the baseline, and the monthly Watch reports movement against it. To see a single-question preview of your own numbers, start with the free visibility check.
Keep a record you can check later
Use the free AI visibility scorecard to plan a small test, enter the observations you have actually checked and download the record. It calculates from your entries; it does not query AI products or scan a website.
Save the original answers and sources alongside your export. Before comparing periods, confirm the same questions, products, settings and repeat schedule were used. Record additions, outages and missing assessments. An apparent gain caused by dropping difficult questions is a change in the test, not demonstrated improvement.
Measuring a law firm or accountancy practice
Keep separate question sets for the services and places you actually cover. See how we scope AI visibility for solicitors and AI visibility for accountants. The examples explain which business facts and service pages to review alongside the captured answers.