Most AI visibility numbers come from one logged-in screenshot. Ours come from the questions your buyers actually ask, run in clean sessions, repeated, dated and scored on seven metrics. The audit covers ChatGPT, Gemini, Claude and Perplexity. Nightly tracking covers ChatGPT, Gemini, Google AI Overviews and Google AI Mode. Here is the full method, so you can check our work.
Audit: ChatGPT, Gemini, Claude, Perplexity. Nightly: ChatGPT, Gemini, AI Overviews, AI Mode
Audit prompt set
40 buyer prompts
Monthly tracking
25 money prompts on Foundation, 50 on AI Search Growth, tracked nightly
Session type
Logged out or temporary chat
Metrics
7, never one blended score
Full re-run
Audit engines, at day 90
The short answer
To measure AI visibility, run the questions your buyers ask through each AI engine in a clean, non-personalized session, repeat every prompt, and track how often you are named, where you rank, who wins instead and which sources the engine cites. Then tie it to Google Search Console so you can see AI impressions, not just answers.
Clean sessionsLogged-in answers flatter your brand
Rates, not screenshotsAnswers change run to run
7 metricsEach one points to a specific fix
01 · The problem
Why most AI visibility screenshots mislead
AI engines personalize answers using your chat history, your location and the wording you use. Run a prompt about your own category from your own account and the engine has already learned you care about your brand. We tested it on ourselves.
ChatGPT · logged in, founder's accountFlattering
Best GEO agency for B2B SaaS?
1Arobis AI
2Agency A
3Agency B
What the founder sees. Not what buyers see.
ChatGPT · temporary chat, same promptWhat buyers see
Best GEO agency for B2B SaaS?
1Agency A
2Agency B
3Agency C
?Arobis AI not named
The only version we report.
Our test · Sep 10, 2026
Same prompt, same day, opposite result. A logged-in ChatGPT named Arobis AI first. A temporary chat did not name it at all. Since then, every number in our reports comes from logged-out or temporary sessions, and we ask to see session details before trusting anyone else's screenshot. Prompt shown is illustrative of the test.
02 · What we measure
The 7 AI visibility metrics we report
Each metric answers a different question and points to a different fix. We report all seven per engine and per prompt, so you can see exactly where you are winning and losing.
Metric 1 · The headline
Share of AI answers (AI share of voice)
The percentage of tracked AI answers in your category that name your brand. The closest thing to market share in AI search. In our sample client report, the client reached 39% at day 90, the highest of 16 brands.
answers naming you ÷ all tracked answers × 100
Metric 2
Prompt mention rate
How often you are named across repeated runs of one buyer prompt. Tells you which questions you own and which you lose.
runs naming you ÷ runs of that prompt
Metric 3
Average position when named
Named first and named fifth are different outcomes. Buyers read the top of the shortlist.
mean rank across answers that name you
Metric 4
Citation share
How often the engine quotes your own website as a source. High citation share means AI trusts your pages, not only what others say about you.
answers citing your domain ÷ all answers
Metric 5
Competitor win rate
Who gets named when you do not. This is your real competitive set in AI search, and it is often not the list your sales team expects.
competitor mentions on prompts you miss
Metric 6
Source map
The third-party sites AI cites in your category, ranked by how often. Roundups, review sites, forums and media. This becomes the GEO placement plan.
citations per domain, top 25
Metric 7
Google AI impressions
How often you appear in Google AI Overviews and AI Mode, straight from your own Search Console. The one AI metric Google reports directly.
GSC › Generative AI report › impressions
Why we do not sell a single "AI visibility score"
A blended score from 0 to 100 looks tidy on a dashboard and hides everything you need to act on. A brand can score well by being named often but always fifth, or by winning prompts no buyer asks. If a score cannot tell you what to fix next, it is a vanity metric. We show the parts.
The formula
How to calculate AI share of voice
AI share of voice is the number of tracked AI answers that name your brand, divided by all tracked answers for the same prompts, times 100. Work it out per engine first, then across engines, and always on the same prompt set as your competitors so the numbers compare.
1
Fix the prompt set
Use the questions buyers ask before a demo, in category words, never your brand name. Our audit uses 40 buyer prompts.
2
Run each prompt several times per engine
Use logged-out or temporary sessions. Answers shift from run to run, so a single pass can overcount or undercount you.
3
Count the answers that name you
Divide by all answers on that engine to get your share there. Then pool the engines for an overall figure, but keep the per-engine numbers: one engine can carry a share the others don't.
4
Do the same for every competitor
Same prompts, same runs, same denominator. In our sample client report the client reached 39% at day 90, the highest of 16 brands, and that ranking only means something because every brand was counted on the same answers.
For the wider set of KPIs, formulas and a tracking template, see our guide to AI visibility metrics.
03 · How we do it
Our 6-step measurement process
The same process runs in every audit and every monthly report, so results compare cleanly over time.
1
Build the prompt set with you
We map how your buyers search: their job, their problem, the alternatives they consider. Then we agree on your money prompts, the questions asked right before a demo, and track that set: 25 on Foundation, 50 on AI Search Growth.
40 prompts in the audit
25 or 50 money prompts by plan
2
Capture a day 0 baseline
Every prompt is run on every engine before we change anything on your site. Without a real before, there is no honest after.
3
Run in clean conditions and repeat
Logged out or temporary chat, location stated in the prompt, US Google settings, each prompt repeated. One answer is an anecdote. A rate across runs is a measurement.
4
Score the seven metrics
Share of answers, mention rate, position, citation share, competitor wins and the source map, per engine and per prompt.
5
Tie it to your Google data
We read your Search Console, including the Generative AI report, and your analytics. We split branded from non-branded and filter out paid traffic, so AI wins show up as real demand.
6
Report, fix, re-measure
A monthly report that leads with a one-line takeaway per section, lists what shipped, and sets next month's targets. Full re-run on ChatGPT, Gemini, Claude and Perplexity at day 90.
The prompt decides the result. We never test branded prompts like "Is Arobis AI good?". We use the category words your buyers use, across five intent types. Examples below use a project management tool.
Shortlist"Best project management software for remote agencies"The money prompt type
Problem"How do I stop missing client deadlines across 20 projects?"Before they know the category
Alternatives"Best alternatives to [market leader] for small teams"Switching intent
Comparison"[Tool A] vs [Tool B] for client reporting"Late-stage decision
Constraint"Project tool with time tracking under $15 a user"Real-world filters
Field note
One word can change the whole shortlist. In our tests, adding "agentic" to a software prompt pulled engines toward AI-agent frameworks and app builders instead of the category we were measuring. Buyer language beats marketing language, every time.
05 · Test conditions
The test conditions behind every number
Every result in our reports states these conditions. If a condition changes, we say so in the report.
Condition
What we do
Why
Session
Logged out, or a temporary chat with memory off
Removes personalization from your own history
Location
Country named inside the prompt; Google set to US, English, non-personalized
Our Oct 1, 2026 test: changing Google's country setting did not change the AI Mode shortlist. Naming the country in the prompt did.
Repetition
Each money prompt run several times; results reported as rates
AI answers vary from run to run
Wording
Buyer category words, no brand names, no leading phrasing
Leading prompts produce the answer you wanted
Dating
Every capture saved with engine, date and session type
Engines update often; undated results cannot be checked
Traffic
Paid AI traffic, like your own ChatGPT ads, filtered out by UTM
In one client account, ads made up about 200 of 230 monthly "ChatGPT" visits
06 · Coverage
The AI engines we cover, and how
Each engine exposes different data, so we measure each one the way that engine allows.
Engine
How we measure
Cadence
ChatGPT
Nightly tracking in the client portal plus logged-out manual checks
Nightly
Google AI Overviews
Nightly tracking in the client portal plus Search Console Generative AI report
Nightly
Google AI Mode
Nightly tracking in the client portal plus Search Console impressions
Nightly
Gemini
Nightly tracking in the client portal plus manual clean-session checks
Nightly
Perplexity
Manual checks; session type noted where sign-in is required
Audit + day 90
Claude
Manual clean-session checks
Audit + day 90
ChatGPT, Gemini, Claude and Perplexity are fully re-run at day 90. AI Search Growth clients get weekly re-checks.
07 · Google data
What Google Search Console can and cannot tell you
Search Console now has a Generative AI report. It is the only first-party AI data Google gives you, and it is easy to misread.
What it shows
Impressions in AI Overviews and AI Mode
Which of your pages Google's AI pulls from
Trends month over month, from your own property
What it does not
Clicks from AI answers, only impressions
Real searches for long question-shaped queries. Many are generated by Google's AI behind the scenes (query fan-out) and cannot be reproduced by googling them
Anything about ChatGPT, Claude or Perplexity
Our own data
About 35% of our own Search Console impressions came from AI-generated queries in September 2026. Reported as normal rankings, they make average position look worse and CTR look broken. We separate them in every report.
08 · Our rules
What we never count
Rules are only useful if they cost you something. These cost us prettier reports.
Logged-in screenshotsThey show what you see, not what buyers see.
We run the questions your buyers ask in clean, non-personalized sessions: the audit and the day 90 re-run cover ChatGPT, Gemini, Claude and Perplexity, and the client portal tracks ChatGPT, Gemini, Google AI Overviews and Google AI Mode every night. We repeat each prompt and score seven metrics: share of AI answers, prompt mention rate, average position, citation share, competitor win rate, the source map and Google AI impressions from Search Console.
What is AI share of voice?
AI share of voice, which we call share of AI answers, is the percentage of tracked AI answers in your category that name your brand. If your brand appears in 310 of 1,000 tracked answers, your share is 31%. It is the closest thing to market share for AI search.
Why don't you give a single AI visibility score?
A single score hides what you need to act on. A brand can be named often but always fifth, or rank first on prompts no buyer asks. We report the separate metrics, per prompt and per engine, so every number points to a specific fix.
Why do you test in logged-out or temporary chat sessions?
Because personalization flatters your brand. In our own test on September 10, 2026, the same prompt named Arobis AI first in a logged-in ChatGPT account and did not name it at all in a temporary chat. Buyers do not share your history, so we only report clean sessions.
How many prompts do you track?
The AI Visibility Audit tests 40 buyer prompts. In a monthly engagement we agree on money prompts with you, the questions asked right before a demo: 25 on Foundation, 50 on AI Search Growth. That tracked set runs every night.
How often do you re-measure?
The tracked prompt set runs every night on ChatGPT, Google AI Overviews, Google AI Mode and Gemini. Money prompts are re-checked by hand every month and fully re-run on ChatGPT, Gemini, Claude and Perplexity at day 90. AI Search Growth clients also get weekly re-checks.
Can I reproduce your results myself?
Yes. Every result in our reports lists the engine, the date, the session type and the prompt, so you can re-run it in a temporary chat. Answers vary from run to run, which is why we report rates across repeated runs rather than a single screenshot.
What is a good AI share of voice?
There is no universal benchmark, because it depends on how many vendors compete for each prompt. In a category with 16 tracked brands, 39% can make you number one, as it did at day 90 for the client in our sample report. Compare yourself with the competitors you meet in deals, on the same prompts, and watch the trend month to month. If the gap is wide, closing it is the job of our AI visibility service for SaaS.
Is AI share of voice the same as share of model?
Close enough to be used the same way. Share of model is the term our State of AI Search Visibility report uses for how often an AI engine names your brand versus competitors on a set of buyer prompts. We call it share of AI answers because we count it per answer and per engine.
Which tools do you use?
Google Search Console including its Generative AI report, your web analytics, Bing Webmaster Tools for Copilot citation data, Ahrefs for authority, ZeroRank for nightly AI answer tracking, and manual clean-session checks on every major engine. Tool costs are never billed to clients.
Measured the right way
Find out what AI engines really say about your brand.
Your free AI Visibility Audit preview uses this exact method: real buyer prompts, clean sessions, dated results.