Run this test before reading on. Open ChatGPT, ask it for companies in your industry in your country, write down the names. Open a new window and ask exactly the same thing. Compare.
Almost certainly you didn't get the same list. That isn't a bug, and it is the most important thing to understand before hiring anybody in this category.
Why Google has a position and this doesn't
A classic search engine has a ranking. For the same query, from the same place on the same day, the result is practically the same for everyone. That is why it makes sense to say "we're third" and track that number over time.
An assistant has no ranking. It generates a fresh answer each time, and that generation has randomness in it by design. It also depends on which pages it managed to read at that moment, on how the question was phrased, and on the language it was asked in.
A recent paper from the University of St. Gallen puts it better than we could. It is called "Don't Measure Once", and its argument is that AI search is probabilistic: answers vary across runs, prompts and time, which makes any one-off observation unreliable. Its practical conclusion is that brand visibility has to be measured repeatedly and treated as a distribution, not a single data point.
Language weighs more than you'd think
There is a second factor almost nobody measures, and it matters a lot here, because many Costa Rican companies sell into two markets in two languages.
A July 2026 preprint decomposed where the variation in what models say about a brand comes from, across 20 brands, 8 languages and 3 models. The language of the query explained 26.5% of the variance, against 1.5% for brand identity.
Take it with the right weight: it is a preprint, the brands were Central and Eastern European, and what it measures is the tone of the answer rather than exactly who gets named. But the direction is clear and it matches what we see: measuring only in Spanish, or only in English, measures a part.
So what can be measured
Plenty, and well. What changes is the unit.
You don't measure a position. You measure how often a company shows up when the same question is asked many times. With enough repetitions, that frequency is stable and comparable from one month to the next.
Which gives you three questions worth asking anyone selling monitoring in this field:
- How many times do you run each question? If the answer is once, the monthly number they hand over is noise presented as a time series.
- In which language and from which country? Both change the result.
- Do you hand over a range or a round number? A number with no variation band, in a phenomenon that varies by definition, is hiding exactly what should be shown.
What this does not mean
It doesn't mean the work can't be done. It means separating two things that usually get sold together.
On one side is what is verifiable without argument: whether your site can be read, whether you let the right robots in, whether your pages answer something concrete, and whether what the AI says about your company is true. None of that depends on chance. A wrong opening time or a service you don't offer is a binary error: you fix it and it stays fixed.
On the other side is how often you get named, which does vary, and which is therefore measured with repetitions, with a date and with a range.
Anyone mixing the two into a single dashboard with one big number in the middle is selling reassurance, not measurement. The difference shows up the day the number drops and nobody can explain why.
That is how we work, and it is laid out in Get recommended by AI.
