You measure AI visibility yourself by asking every real customer question at least ten times per AI engine and tallying how often you are mentioned. That percentage says far more than a position, especially when you set it against your branded searches.
You can measure AI visibility yourself, without an expensive tool. Ask every important customer question at least ten times per AI engine, tally whether you are mentioned and turn that into a percentage. That percentage says far more than a position.
The rule comes from advice by SparkToro, the company of Rand Fishkin. In the Office Hours of March 2026 it was put like this:
Please, dear god, ask ten times before you trust the list.
In other words: for heaven's sake ask ten times before you trust the list. Below you can read how to apply that rule in practice.
Why asking once is not enough
If you ask ChatGPT or Google AI the same question 100 times, the chance of getting the same list of brands twice is less than 1 in 100 (SparkToro, January 2026). One answer is therefore a snapshot, not a measurement.
At the same time, some names keep coming back. Fishkin therefore concluded in January 2026: "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric". Visibility as a percentage across many repeated prompts is thus a reasonable yardstick.
A position is not, read more about that in the article on the position that does not exist.
First: take your prompts from real customer questions
The most important choice is which questions you ask. When hundreds of people wrote a prompt for the same intent, the average semantic similarity was only 0.081 (SparkToro, February 2026). So everyone asks it differently.
AI visibility is an audience research problem before it's a measurement problem.
Put differently: AI visibility is first a question of audience research and only then a question of measurement. In March 2026 SparkToro also advised deriving prompts from real audience behaviour. SparkToro recommended running them "a few dozen times", in other words several dozen times.
If your AI visibility strategy assumes people ask questions the same way, it's already misaligned with reality.
In short: if your strategy assumes everyone asks the same question, it is already out of step with reality. Still, you do not need to cover every phrasing. Prompts written by people, 142 variants for headphones, produced a comparable core set of brands (SparkToro, January 2026).
So AI mainly picks up the underlying intent. Make sure your set of questions covers the real intents of your customers, not every possible choice of words.
Good sources for prompts are:
- questions customers ask on the phone or by email;
- questions from quote requests and intake conversations;
- searches from Google Search Console;
- questions you see on forums and in reviews about your field.
Write each question in the customer's words, not in jargon. A good question often names a place or situation, such as "which installation company in Amersfoort fits heat pumps". For place-based questions, read more in local AI visibility per town and region.
Then: tally per engine and work out a percentage
- Choose a fixed, small set of customer questions.
- Ask ChatGPT each question at least ten times, each time in a new chat.
- Repeat that in every engine you want to follow, such as Google AI Overviews and Gemini.
- For each answer, tally whether you are mentioned, and whether as a recommendation or neutrally.
- Divide the number of mentions by the number of runs: that is your visibility as a percentage.
Use our tally sheet for AI visibility for this. Keep the engines separate, because an average across everything hides differences. Even the ChatGPT app and ChatGPT via the API shared only 2 of their top 3 in the Gold Lemon benchmark of September 2026.
Also note the date and whether you were logged in. That way you can compare later under the same conditions. Preferably measure a few competitors as well, so that your percentage gets context.
Also check whether the names mentioned are correct. In the SparkToro research of January 2026 AI also named influencers, TikTok accounts and estate agencies that no longer existed. If you see outdated details about your own business, clean them up wherever they appear online.
Why twenty repeats are more accurate
Asking ten times gives a first impression. If you want a measurement with a margin of roughly plus or minus one percentage point, run a set of prompts about twenty times (SparkToro, March 2026). The research itself went further, with 60 to 100 repeats per prompt (SparkToro, January 2026).
The advice from that research was:
if you really want to know an AI's set of recommendations, you need to ask over and over again; usually at least 60-100X, then average these out
If you really want to know which names an AI recommends, ask again and again and take the average. For a small business, twenty times per engine is a workable compromise.
In small markets the picture stabilises faster. The smaller the pool of possible names, the greater the overlap between answers (SparkToro, January 2026). A local specialist therefore sees a clear pattern sooner than a national category does.
Link your measurement to branded searches
A percentage on its own does not tell you whether it pays off. Similarweb research that Fishkin discussed in June 2026 showed that after AI influence people search slightly less and go directly to a site more often. When AI recommended American Express, the chance of a site visit was 7.2% higher.
Fishkin himself adds caveats: it concerned large brands in three sectors (SparkToro, June 2026). Still, his advice is useful for everyone: check in your own analytics whether direct traffic and branded searches are rising. You will find an explanation in searches for your brand name as evidence.
What you can do now
- Collect real customer questions from your email, your phone calls and Search Console.
- Ask each question at least ten times per engine and tally the outcome.
- Work out your visibility as a percentage per engine and note the date.
- Repeat the measurement every month under the same conditions.
- Set your percentages against your direct traffic and then decide whether you need an AI tracking tool.
Frequently asked questions
How do I measure my AI visibility myself?
Ask your most important customer questions repeatedly in each AI engine. Tally whether you are mentioned and divide the number of mentions by the number of runs. See also: 10 questions for your agency.
What is the ten-times rule?
The advice to ask an AI a question at least ten times before you trust the list. One answer is too random to draw any conclusion from. See also: Follow-up client calls.
Why not just ask once?
AI answers change in names and order from one time to the next. Only across many repeats do you see which names really come back often. See also: Best GEO agency: how to recognise a fair ranking.
How do I find good prompts?
Take them from real customer questions: email, phone, quote requests, Search Console and forums. Write them in the customer's words. See also: Getting mentioned in ChatGPT: the evidence lies outside your website.
Should I measure each AI engine separately?
Yes. Each engine gives its own answers, and even the app and the API of the same engine differ. An average across everything hides those differences. See also: AI visibility tally sheet.
Should I start a new chat for every question?
Yes, that is sensible. It prevents earlier questions in the same chat from influencing the answer.
How accurate is asking ten times?
It gives a first impression, not a precise measurement. For a smaller margin of error you need more repeats.
How often should I measure again?
A fixed monthly measurement under the same conditions works well. That way you see trends instead of flukes.
What counts as a mention?
Every time your business name appears in the answer. Note separately whether you are mentioned as a recommendation or neutrally.
How do I know whether AI visibility pays off?
Check whether direct traffic and searches for your brand name rise along with it. You will find those signals in your own analytics.
Sources
- Office Hours: Can You Actually Track AI Visibility? (SparkToro) (March 2026)
- NEW Research: AIs are highly inconsistent when recommending brands or products (Rand Fishkin, SparkToro) (January 2026)
- Why Everyone Prompts AI Differently (Amanda Natividad, SparkToro) (February 2026)
- New in SparkToro: uncover the AI prompt topics your audience is using (SparkToro) (March 2026)
- GEO-benchmark oktober 2026 (Gold Lemon) (September 2026)
- New Research from Similarweb: How AI Brand Mentions Influence Direct Visits (Rand Fishkin, SparkToro) (June 2026)