How to measure your AI visibility without fooling yourself

Short answer. Asking ChatGPT about your business once doesn’t measure your AI visibility: the answer changes almost every time, and if the question includes your name, it only checks that the model knows you. To measure it, you use a fixed set of customer questions, always the same ones, count the share of answers you appear in for each engine separately, report that figure with its margin of error and cross-check it with Search Console, Bing Webmaster Tools and Google Analytics 4.

We checked every fact on 25 October 2026.

Less than 1 in 100. The chance that ChatGPT or Google’s AI gives the same list of brands in two answers (SparkToro and Gumshoe).

99.4% vs 3.32%. Questions using its name in which a model recognised a startup, against open searches in which it named it (The Discovery Gap).

Up to ±13 points. The margin of error of a percentage measured from 50 answers (our calculation).

Why doesn’t asking ChatGPT once work?

Because an AI answer is a sample, not a ranking. In the SparkToro and Gumshoe study, around 600 volunteers put 12 prompts to ChatGPT, Claude and Google’s AI and collected 2,961 answers. The chance of ChatGPT or Google’s AI repeating the same list of brands was less than 1 in 100; in the same order, around 1 in 1,000.

What held steady was the percentage. The big headphone brands appeared in 55–77% of answers, even though the prompts were very different. The author concludes that “visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric”, and that a “ranking position in AI” is not.

There’s a second reason. Google explains that “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.” If you lump several engines into one figure, you’re mixing different things.

Which questions should you measure with?

The ones a customer who doesn’t know you yet would ask. If you ask “what do you know about [your business]?”, you only check that the model knows you exist.

The Discovery Gap (arXiv, January 2026) tested 112 startups with an OpenAI model queried through the API. Asked by name, it recognised them in 99.4% of questions; in open discovery questions, it named them in only 3.32%.

A useful panel mixes three types of question:

1. Discovery, without your brand: “Which accountant in Girona would you recommend for freelancers?”
2. Comparison: “Which is better for a small online shop, X or Y?”
3. Facts, with your brand: opening hours, prices, services. These measure accuracy, not visibility.

Write several phrasings for each intent: in the SparkToro study, 142 prompts written by different people with the same goal had an average semantic similarity of 0.081. Then freeze the panel; if you change it, compare only the questions both versions share.

How many answers do you need?

Enough for the margin to be smaller than the change you want to detect. “You appear in 30%” taken from 10 answers could mean almost anything. The answers can come from many different questions or from repeating the same ones; if you can choose, adding different questions usually narrows the margin more than repeating. And to accept a change between two measurements you need even more: with 50 answers and a presence of around 30%, only a rise or fall of about 18 points stands out from chance.

With 10 answers, 30% is consistent with any true value between 10.8% and 60.3%. With 50, between 19.1% and 43.8%; with 200, between 24.1% and 36.7%. These are our calculations, using the 95% Wilson interval.

If you don’t appear in any of 50 answers, don’t say “0%”: with 95% confidence, the true value is below 7.1%.

What should you count in each answer?

We separate four things, because each one is fixed in a different way:

• Mention: the AI names your business.
• Citation: it links to or credits one of your pages as a source.
• Recommendation: it puts you forward as an option, not just names you.
• Accuracy: the facts it gives about you (price, opening hours, services) are correct.

One of your pages can be cited in an answer that recommends someone else, and a mention can come with the wrong opening hours.

Also count the times the AI doesn’t answer. In a Seer Interactive study of its own brand (28,123 answers from six platforms), the AI tools didn’t answer 31.2% of questions about the brand, and 80.3% of comparison questions. When they did answer, they were 95.9% accurate on average.

What data do the search engines themselves give you?

Three free tools complete the picture, although none of them tells you what the AI answers.

• Google Search Console. Since 31 August 2026, every site has the Search Generative AI performance reports. They show impressions: how many times your URLs appeared in AI features in Search and Discover, by page, country, device (in Search) and date. Google’s announcement mentions no clicks.
• Bing Webmaster Tools. Its AI Performance report counts citations of your site in Microsoft Copilot, in Bing’s AI summaries and in some partner integrations. It also shows the cited pages and a sample of grounding queries, the phrases the AI used to retrieve the content it cited.
• Google Analytics 4. The default “AI Assistant” channel groups visits from assistants such as ChatGPT, Gemini, Deepseek, Copilot or Grok. Note: Google defines it without AI Overviews or AI Mode, whose traffic counts as organic search.

Why isn’t traffic enough to measure?

Because much AI visibility produces no click. Pew Research Center analysed 68,879 searches by 900 US adults in March 2025. With an AI summary on the page, users clicked a traditional result in 8% of visits, against 15% without one; they clicked a link inside the summary itself in only 1%.

If someone reads your name in an answer and then searches for you on Google, that visit isn’t credited to AI. That’s why we think Analytics undercounts AI traffic (this is our inference, not a measured figure).

Which measurement mistakes mislead the most?

Which measurement mistakes mislead the most?

Mistake	Why it misleads	What to do instead
Asking only once	The lists almost never repeat	Use a fixed panel of many questions and report a percentage
Asking with your name	It measures whether the AI knows you, not whether it recommends you	Use discovery questions without your brand
Talking about “position”	The order changes in almost every answer	Report the share of answers you appear in
Mixing engines	Each engine uses different models and sources	One series per engine, never an average
Giving figures without a margin	3 out of 10 is consistent with 11–60%	Publish “k out of n” and the interval
Measuring only traffic	Many answers produce no click	Cross-check panel, Search Console, Bing and Analytics
Crediting everything to your changes	A model update can move the figures without you doing anything	Log the date of every change, yours and the engine’s

Frequently asked questions

Next step

Nobody controls what an AI answers. Google says so about its own search engine: “No one can guarantee a #1 ranking on Google.” What you can do is measure with a method, change what’s in your hands and measure again with the same questions.

If you want to know where you’re starting from, our free diagnostic is the first step. If you already know, in services we explain how we implement the changes and track them month by month.