What to demand from an AI visibility tool (so you don't buy smoke)

Earlier this year Natzir Turrado published a thorough critique of the tools that measure brand visibility in AI: Herramientas para medir visibilidad en IA: lo que no te cuentan los Prompt Trackers (in Spanish). It's worth reading in full.

His thesis, in short: AI models never answer the same way twice, so many tools sell a certainty they don't have. Rankings that don't exist, scores with no margin of error, and conclusions drawn from one question asked once.

We agree on the essentials. That's why we think the useful question isn't whether AI visibility can be measured, but how to measure it so the result is worth something. These are the seven things we'd demand from any tool, ours included.

1. No rankings

There is no "position number 3" in AI. Every answer is different, and the order of brands changes almost every time. If a tool shows you a position, be wary. What does make sense is measuring how often you are named when a customer asks about what you sell.

2. Questions that don't say your name

"What is [my company]?" measures whether AI knows you, not whether it recommends you. What matters is what happens when the customer describes their problem without knowing you exist. At Yovara we always keep the two apart: questions without your name give the main number; questions that name you are measured separately.

3. Every question repeated, on more than one assistant

One isolated answer says almost nothing. We ask each question three times on ChatGPT and three times on Perplexity.

What you learn from doing it is the interesting part. The SparkToro study Natzir cites found that getting the same list of brands twice is extremely rare. We measure something simpler, whether you are named or not, and that turns out to be far more stable: in our first measurements, when a company appeared for a question, it appeared in all three repetitions 93 % of the time.

What does change a lot isn't repetition: it's the assistants themselves. Of the questions where at least one named the company, both did in only 45 %. For Futboleras, a women's football platform: 8 of 21 questions. Measuring a single assistant tells you about that assistant, not about your visibility.

4. How much you can trust the number

A bare "64 %" is prefabricated certainty. Questions on one topic resemble each other, so counting answers isn't enough: each question has to be treated as one unit and the margin calculated. For Futboleras, that 64 % has a margin of about ±12 points. That's why Yovara shows the main number with its margin.

5. Always the same questions

To know whether your improvements work, the second measurement must ask exactly the same questions as the first. And subtracting two percentages isn't enough: each question has to be compared with itself, checking whether the change exceeds the margin.

We saw it with Futboleras: after fixing technical problems on its website, visibility went from 61 % to 64 %. Compared question by question, that +3 has a margin of ±8 points. It isn't a clear change, and Yovara says so. We tell the story in what GEO actually means in 2026.

6. Mentioning you is not citing you

AI using your website as a source doesn't mean it recommends you. Very different things can happen: it doesn't read your site, it cites it without naming you, it names you based on other sources… Each case calls for a different action, and a good measurement has to tell them apart.

7. Why, and what to do

Natzir stresses something we agree with: the value isn't a score but knowing which topics your competitors appear in and you don't, and why. At Yovara every question tells you who appears instead of you, which sources AI relies on and the concrete recommendation that follows, always linked to that evidence.

In short

Measuring AI visibility isn't impossible. What's impossible is measuring it with the precision some tools pretend to have. The answer isn't to stop measuring but to measure properly: questions without your name, repetitions, several assistants, a visible margin and honest comparisons.

If you want to see how we do it, the preview is free and our methodology is public.

Frequently asked questions

Can you measure a brand's visibility in ChatGPT?

Yes, but not with one question asked once. You need to repeat each question, use several questions without the brand's name, compare several assistants and show the margin of the result.

Why does each assistant give different results?

Because each one searches for and picks its sources in its own way. In our measurements, ChatGPT and Perplexity agree on naming a company in less than half of the questions where it appears.

How do I know if my changes improved my AI visibility?

By measuring again with exactly the same questions and comparing each one with itself. If the change doesn't exceed the margin, you can't claim it improved.