Quick answer
Ask a fixed set of 20 to 50 real buyer questions on the consumer versions of ChatGPT, Gemini and Google AI Mode (and Perplexity if your buyers use it), at least three times each. For every answer, record whether you are named, recommended and linked, which competitors appear and which sources are cited. Report share of voice against named competitors, keep a set of questions you do not work on as a control, and re-run the same set monthly.
Key takeaways
- One run is not a measurement. Only 30% of brands stay visible from one AI answer to the next, and 20% across five runs (AirOps, 2026).
- Measure what buyers see. The consumer apps search the web live; a model's API often answers differently.
- Filter out same-name companies. An unfiltered count put REVEX at 8 of 24 answers. Every one of them was about a different company called Revex.
- Being known is not being recommended. Across four Swiss B2B tech companies we checked, engines named each company in every answer about the company itself (36 of 36), but in only 24 of 60 answers to category shortlist questions.
- Share of voice is zero-sum. If everyone's mentions go up, the engines are naming more brands per answer, not favouring you.
The method in one table
| Step | What we do | Why |
|---|---|---|
| Questions | 20 to 50 buyer questions in the buyer's words, grouped by theme and buying stage | Engines answer the question asked; a keyword is not a question |
| Engines | ChatGPT, Gemini and Google AI Mode as a signed-out user; Perplexity where relevant | These are the answers buyers actually read |
| Runs | 3 per question per engine | Answers vary from one run to the next |
| Per answer | Named, recommended, linked, competitors named, sources cited | Each one points to a different fix |
| Control | A theme we deliberately do not work on | Separates our effect from engine-wide changes |
| Cadence | Same questions, monthly | Comparable over time |
Five mistakes that inflate the numbers
1. Measuring once
AI answers change from one run to the next. AirOps' 2026 State of AI Search found only 30% of brands stay visible from one answer to the next, and just 20% across five consecutive runs. A single run can show you in or out by chance, so a before-and-after built on single runs proves nothing.
2. Measuring the API instead of the app
A model's API answers from its training data unless web search is switched on, and even then it does not behave exactly like the consumer app. Buyers use the app. We capture the consumer surfaces signed out, and we label any API-based measurement as such.
3. Counting other companies with your name
When we measured REVEX itself, a simple name match counted 8 of 24 answers as mentions. Every one was about a different company: a UK payments firm, a supplement and a software tool. Filtered to answers that were actually about us, the score was 0 of 24. If your name is not unique, count a mention only when the answer also shows it is about you: your domain, your category or your location.
4. Treating a citation as a recommendation
A link in the sources panel is not the same as your name in the answer. Buyers read the answer. We track both, and explain the difference in cited vs. recommended.
5. Confusing training bots with search bots
Blocking GPTBot or ClaudeBot stops OpenAI and Anthropic from training on your content. It does not remove you from AI answers, which use separate search crawlers such as OAI-SearchBot, Claude-SearchBot and PerplexityBot. OpenAI's crawler documentation says so directly: its settings are independent, so you can allow OAI-SearchBot while disallowing GPTBot. (Google-Extended works differently and is worth checking against Google's own documentation.) Blocking the training bots is a legitimate choice; blocking the search bots makes you invisible.
Known is not the same as recommended
The most common pattern in our measurements is a company the engines know but do not shortlist. Asked “what is [company]?”, all three engines describe it. Asked “what are the best [category] tools for [buyer]?”, they name competitors instead.
| Question type | Answers naming the company | Share |
|---|---|---|
| About the company itself | 36 of 36 | 100% |
| Category shortlist questions | 24 of 60 | 40% |
Four Swiss B2B software companies, 8 questions each, ChatGPT, Gemini and Google AI Mode, one run each, 28 October 2026. A small sample: it shows the pattern, not a market rate.
That gap is fixable, and it is not fixed by more brand content. It is fixed by appearing in the pages engines read when they build a shortlist: comparison pages, roundups, review sites and trade coverage.
The metrics we report
- Mention rate: share of answers that name you, per engine and theme.
- Recommendation rate: share of answers that put you forward, not just mention you.
- Citation rate: share of answers that link to your site.
- Share of voice: your mentions divided by all brand mentions in the answers, against named competitors.
- Source map: the domains each engine cites for your questions, which tells you where to earn mentions.
What it costs to measure
Measurement is cheap; acting on it is not. Our eight-question snapshot on three engines costs about $0.10 in data fees. A full baseline of 20 questions on four engines, three runs each, costs a few dollars. The work that changes the numbers (answer pages, fixes and earned mentions) is where the budget goes, as covered in what AEO and GEO cost.
Questions people ask
How do I check if ChatGPT recommends my company?
Ask the questions your buyers ask, in their words, on chatgpt.com while signed out, at least three times each, and note whether you are named, recommended or linked, and who is named instead. Repeat on Gemini and Google AI Mode.
How many prompts do I need to track?
Between 20 and 50 well-chosen buyer questions is enough for most B2B companies, grouped by theme. More questions run once each are worth less than fewer questions run several times.
Does blocking GPTBot remove me from ChatGPT answers?
No. According to OpenAI, GPTBot collects content that may be used for training, while OAI-SearchBot decides whether your site can appear in ChatGPT's search answers. Each setting is independent: you can block GPTBot and allow OAI-SearchBot.
How often should AI visibility be measured?
Monthly on the same question set is enough to see trends. Weekly makes sense only when you are testing a specific change.
Sources