Are you named in AI answers? How to measure it

The client tried it himself. He types “best SEO agency in Vienna” into ChatGPT, is not named, and calls. You type the same thing, are named, and afterwards nobody knows more than before. It only gets expensive at the point where that coin toss turns into a budget: texts get rewritten, author boxes built in, schema retrofitted — and afterwards the measurement from before is missing, the one the effort could have been held against.

How to go about it in JMX

1. The GEO tab, “AI crawler access” overview. Read the Purpose column first. Blocking model training costs no referral, blocking answer search takes you out of the answers. The most common find is a robots.txt that blocks everything with “AI” in the name across the board and catches OAI-SearchBot and PerplexityBot with it.
2. The same view, four areas. 33 GEO rules assess answerability, evidence, authority & attribution and delivery. That is the precondition, not the measurement.
3. Settings → Google → DataForSEO for the four fixed engines (ChatGPT, Gemini, Claude, Perplexity); alternatively Settings → AI → OpenRouter for freely chosen models from the catalogue. The first route needs no keys of your own at the AI providers.
4. The AI Visibility tab. One prompt per line, plus your own brand including the domain, plus the competitors. Set Samples per prompt to at least three, otherwise you are measuring an accident; the cost line works out requests times the fixed share before the start, and with OpenRouter the upper bound from the price list ($ per 1M tokens) as well. Then Measure visibility.
5. Read the model comparison and share of voice. Every engine and every model sits in one row with its mention and citation rate, with a direction arrow against the previous measurement day. Open the Answer, verbatim at every surprise — it is in the project file and explains the why.
6. The Rankings tab. The AI Overview columns fall out of the same position measurement, and the Citation register evaluates the source lists that come with it anyway, without another paid request. The GEO tab then shows under Cited own pages whether a cited address still answers 200 at all or sits on noindex.

What to watch out for

The most common misreading is the round percentage. JMX writes the sample size next to every rate, because “2 of 3” claims something different from “67 %”. At three samples a single outlier is a deviation of 33 points, and anyone who changes the content strategy off the back of it has taken noise for a signal.

The limit of the method: what is measured is the model’s API answer, not your client’s screen. His history, his location and his account settings feed into that and cannot be reproduced from outside — a deviation between his sample and yours is therefore not an error of the measurement but a property of it. Only what has searched can cite. If a model runs without web search, the citation rate stays empty, and the mention rate is the only number that carries there. JMX does not scrape Google result pages at all; the AI Overview comes in exclusively via the paid DataForSEO request.

Two special cases regularly cost an hour of searching. Mentions count with word boundaries — “example.com” does not count inside “new-example.com” — and changing the brand list afterwards re-evaluates the entire history retroactively, because the rates are local computations and not stored numbers. In web analytics, on top of that, visits from AI apps without a referrer count as “Direct” by design.

Take the first measurement only after the rebuild and you no longer have a before value, and you answer the same question at the next phone call with another coin toss.