Practice
Our Methodology: How We Measure AI Visibility with a Prompt Catalogue
1The short answer
We measure AI visibility using a fixed catalogue of questions that real customers ask. Each question is asked several times in the agreed AI systems under documented conditions; each answer is analysed against fixed criteria: mention, position, context, citation and accuracy. We disclose the catalogue and the rules to our clients in full, so that every result can be traced.
This article describes exactly how we proceed. The general principles of measurement are explained in the article Measuring AI visibility.
1. Selecting questions
The catalogue is built together with the client from sources that reflect real demand: sales and support conversations, search queries from Search Console, questions from forums and reviews. We organise it by intent:
| Type | Example | What it shows |
|---|---|---|
| Category question | Which providers are there for …? | Is the brand mentioned at all? |
| Comparison question | X or Y – which is better for …? | How is the brand positioned against competitors? |
| Problem question | How do I solve …? | Is the brand recognised as a solution? |
| Brand question | What does [brand] do? Is [brand] trustworthy? | Is the brand portrayed correctly? |
| Regional question | … in [region] | Is the location assigned correctly? |
The number of questions depends on topics, markets and products and is set out in the proposal. Once defined, the core of the catalogue stays stable so that developments remain comparable over months; new questions are added on top.
2. Querying under documented conditions
- Systems: the agreed platforms, e.g. ChatGPT, Gemini, Perplexity and Google AI Overviews.
- Repetition: every question is asked several times, because answers vary.
- Neutral state: new sessions without conversation history and without personalised context.
- Log: date, system, mode (e.g. with or without web search) and location are recorded for each query.
3. Analysing answers
| Criterion | What we record |
|---|---|
| Mention | Does the brand appear in the answer – yes or no? |
| Position | Is it named first, in the middle or at the end? |
| Context | Recommendation, neutral mention or critical portrayal? |
| Citation | Is the brand's own website linked as a source – with which page? |
| Accuracy | Are services, location, prices and positioning correct? |
| Competitors | Which other providers are named? |
| Third-party sources | Which websites shape the answer? |
4. Building metrics
From the individual analyses we calculate the mention rate (share of answers with a mention), share of voice (the brand's mentions relative to all providers named) and citation rate. We report metrics per platform and per question type, not just as an overall figure – otherwise a good score for brand questions hides weak visibility for category questions.
5. Dealing with fluctuations
AI answers are not deterministic, and providers change their models continuously. That is why we assess trends across several measurements rather than individual outliers, and note known model changes in the reporting. For us, a single test question has no evidential value.
What we deliberately don't do
- No guarantees of mentions or positions – AI answers cannot be guaranteed.
- No manipulation through hidden text, fake reviews or artificial mass mentions.
- No metrics without disclosed questions: every figure can be traced back to the underlying answers.
This methodology is the basis for our GEO Audit and ongoing AI Visibility Monitoring.
Author
Jan Hugo · Managing Director, Search, Tracking & AI
Technical foundations, data models, tracking architecture and AI-assisted analysis.
