A prompt set determines which questions are used to measure whether and how a brand appears in AI answers. A provider of time-tracking software, for example, might collect prompts like these:
- “How does digital time tracking work?”
- “Which time-tracking software is a good fit for a roofing company with 50 employees?”
- “How much does Company X’s time tracking cost per user?”
In prompt tracking, these prompts are put to AI systems such as ChatGPT, Gemini, Claude, and Perplexity again and again. The answers yield metrics such as mention rate.
Why the prompt set shapes the results
Every AI visibility metric applies only to the questions that were actually asked. A methodological survey from September 2026, a preprint released ahead of peer review, therefore describes the prompt set as part of the measuring instrument: it selects which situations are evaluated at all and need not reflect what people actually ask. A figure such as “named in 40% of answers” only becomes meaningful once you know which questions it refers to.
The mix matters, too. Suppose a brand is named almost every time in answers to questions that include its name, but only in one answer in ten when people ask about its product category. If half of the prompt set consists of questions with the brand name, its mention rate is above 50%. If the set contains only category questions, the rate is around 10%, even though the AI systems answer the same way in both cases. That is why results are also reported for each group of prompts separately, and why the composition of the set stays the same over time.
What goes into a prompt set
The best starting point is the questions customers actually ask: in sales conversations, in support requests, and in emails and inquiries. They show which problems prospects describe, which terms they use, and which details they mention, such as their industry, company size, or region. These questions are turned into prompts worded the way people would put them to an AI system. To make the results easy to analyze, the prompts are then grouped, for example by these characteristics:
- Topic: the products, services, and use cases for which a company wants to appear in AI answers.
- Buying stage: from early research through comparing providers to the decision, for example about prices or contract terms. This shows at which stage a brand is named and where it is missing.
- Market and language: prompts in the language of every market a company operates in.
- Persona: questions from the perspective of the different customer groups an offering addresses.
- With and without the brand name: prompts that include a company’s name show how an AI system describes the brand; prompts without it show whether the system names the brand on its own.
A persona is a description of a typical customer, such as “HR manager at a logistics company with 300 employees” or “owner of a roofing company whose crews work on job sites.” Both need time tracking, but they ask different questions: the first about shift schedules and a connection to payroll, the second about an app that works on the go. Prompts written from the perspective of several personas show for whom a brand comes up in AI answers and for whom it does not.
Prompts in the language of every market belong in the set because the language of a question can influence which brands are named. In a June 2026 preprint whose author also works for a vendor of AI answer analysis tools, three language models with web access, one each from OpenAI, Google, and Perplexity, were asked in April and May 2026 about the leading companies in an industry in eleven markets in Northern and Central Europe and the Baltics. Brands known mainly nationally or within their region were named far more often when the question was asked in their home market’s language rather than in English; for internationally known brands, the difference was much smaller. Measuring only in English therefore mainly understated the visibility of the nationally or regionally known brands.
Freezing and versioning the set
Results are comparable over weeks and months only if the prompt set stays the same: the same prompts, with the same wording, in the same mix. A fixed set does not prevent answers from changing for other reasons, such as new model versions or new content on the web. But it does make sure that the questions are not the reason for a change. AI systems do not always answer even the same question the same way; this is called answer variability.
Still, a prompt set sometimes has to change, for example when a new product or market is added or customers start asking different questions. In that case, a new version is created, with the date, the prompts added and removed, and the reason for the change. The September 2026 survey recommends disclosing the version of the prompt set and what changed since the previous round of measurement. A changed composition is one of the possible reasons a figure moves, alongside a company’s own measures, competitors’ measures, and changes to the AI systems. If a figure jumps right after a new version, the cause may be the changed composition rather than a change in how the AI systems answer.
This makes the prompt set the starting point for a company’s own measurements in GEO monitoring: only once it is clear which questions count can the measured AI visibility be tracked and reported reliably over time.