How we measure whether AI names you
A ChatGPT screenshot proves nothing: ask the same question twice and the answer changes in one of every three cases. So this is how we measure.
1. The 25 questions
- They are questions one of your buyers would ask without naming your brand: "which booking tool would you recommend for a small yoga studio", "what's the cheapest alternative to X", "which invoicing software is compliant in Canada". Category and use-case questions name no brand at all; comparison and alternatives questions name competitors, and when we work out who holds a question we don't count the brand that appears in the question itself (section 6).
- We propose them from your site, your Search Console and the questions your competitors win today; you approve them before the admission measurement and sign them on day 0; they are the same questions. From then on they don't change for 90 days. If they changed, any improvement would be cheating.
- Four types, in fixed proportion: 10 category questions ("which tool for..."), 6 comparisons ("X or Y"), 5 alternatives ("alternatives to X") and 4 use cases ("how do I... with which tool").
- Questions are asked in your buyer's language and written for your buyer's market.
2. The 5 engines
ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews. If Google shows no AI block for a question, that answer counts as "not named" and appears as "no block" in the Ledger. Each engine is measured separately and has its own baseline: it is normal to show up in one and not in another.
3. Seven runs per question and engine
- Every week we ask each question 7 times in each engine: 25 × 5 × 7 = 875 answers.
- Why 7: an academic study from April 2026 ("Don't Measure Once", arXiv:2604.07585) found that the sources AI cites overlap only 34-42% from one day to the next, and that at least 7 runs per question are needed for the error margin of a brand's rate to fall below 10 points. They recommend 7 a day; we run 7 a week and accumulate weeks, which is what the same study advises for reading trends.
- No personalization: each run is a new conversation, with no history and no account of yours or ours.
- Market and language: we measure through the API (each engine's programming interface, not its consumer app), with web search on and with your buyer's country or location in the engines that allow it. This is what the measurer does, engine by engine (version of 2026-10-28):
| Engine | How we ask | Buyer's country or location | Language | Model (2026-10-28) |
|---|---|---|---|---|
| ChatGPT | API (Responses) with web search | Buyer's country | In the question text | gpt-6-luna |
| Claude | API (Messages) with web search | Buyer's country | In the question text | claude-sonnet-5-5, medium effort |
| Perplexity | API (Agent) with web search | Buyer's country | In the question text | perplexity/sonar |
| Gemini | API (generateContent) with Google search | Approximate location (latitude and longitude of your market); the API accepts no country code | Market language code | gemini-3.8-flash |
| Google AI Overviews | One capture a day of Google's results page by country (OpenSEO on DataForSEO); the week's seven captures are that question's seven repetitions, spread over seven days instead of one hour | Buyer's country | Market language code | Google's AI block; no model to choose |
The API model is not necessarily the one the same provider's consumer app uses; that is why a screenshot of ChatGPT, Claude, Gemini or Perplexity doesn't replace the Ledger and doesn't count toward the tally. If a provider changes or retires a model, we note it in this table with a date (section 12), because it changes comparability between weeks. That is why this page, which the contract incorporates, says "with your buyer's country in the engines that allow it", not "from your market".
4. What counts as "named" and what as "cited"
- Named: your brand or domain appears written in the answer text, including the visible text of a link. The URL of a link or of a source doesn't count: a link without the name is not a mention.
- Cited: your domain appears among the sources or links the engine returns with the answer, even if the text doesn't name you.
- We measure both, separately, and also record your position and who shows up in your place. The guarantee is calculated on named.
- We measure through the API. The API shows no ads, so nothing bought enters the Ledger or the count.
5. What a won question is
[PROVISIONAL, pending calibration with our own case] A question counts as won in a given week when:
- you are named in at least 4 of the 7 runs in an engine, and
- that happens in at least 2 of the 5 engines, and
- it repeats two weeks in a row.
That way a lucky streak doesn't count, and one bad week doesn't take it away.
6. What a winnable question is (the admission filter)
Before accepting a client we run the full measurement (875 answers). A question is winnable when no competitor is named in more than 3 of 7 runs in at least two engines: there is room. In comparison and alternatives questions, the brand named in the question itself doesn't count as holding it: the answer repeats it by construction. If your category doesn't have at least 8 winnable questions out of 25, we tell you and you don't join. The guarantee only makes sense if we can keep it.
7. The calendar
- Before signing: admission measurement, 875 answers with your 25 questions already approved. It is the baseline and you get all of it.
- Day 0: signing and first payment; the 25 questions are fixed and are the same as in the admission. The first weekly measurement of the sprint runs within the following 7 days.
- Every week, Monday morning (Munich time): 875 new answers; the exact time of each answer is in the Ledger. If an engine fails and a measurement is incomplete, we say so in the report and repeat the failed answers that same week, with their real time in the Ledger, or the whole measurement if needed; we never fill gaps. An incomplete measurement neither counts nor breaks the two-week streak.
- Between measurements (pulse): a single run per question and engine, so you can see the trend in the diary. It is informative only: it doesn't count toward the guarantee.
- Every Monday: one-page report with the raw answers attached.
- Week 13 (day 90): final measurement. It is the one that counts for the guarantee: a question is won if it meets the rule in section 5 in the week-12 and week-13 measurements.
8. The guarantee, in numbers
If at the day-90 measurement you have fewer than 5 won questions out of 25 (under the rule in section 5), we refund the three sprint payments ($790 × 3), and the second-language payments if you contracted it, within 10 days, through the same payment rail you used. What was published on your site stays.
Two things void the guarantee: your site being offline for more than 3 days in total during the sprint, or you removing AI crawler access or reverting changes you had approved. The guarantee also requires the three payments made and the sprint completed; full conditions are in Terms and guarantee.
9. The Answer Ledger: what you receive
Every Monday, a file with the week's 875 answers. Per row: date and time, engine, question, run number (1 to 7), named (yes/no), cited (yes/no), position, brands that appear, URLs cited and the full answer text. It is yours from day one and stays yours if you leave. If you have an SEO or an agency, they can check everything.
10. What we don't measure
- Traffic. AI sends very few visits today (in the largest 2026 study, ChatGPT sent 0.62% of the sessions Google sent). If someone sells you "traffic from ChatGPT", ask them what that is in your Analytics. What can be measured is whether you are in the answer when someone asks, and that visitor converts several times better than Google's.
- Google rankings. There are tools for that.
- Branded questions. If someone asks about you by name, AI names you almost always; that is not visibility, it is brand search.
11. We apply the same method to ourselves
From 2 November 2026 we measure namedwhenasked.com with these same 25 questions, 5 engines and 7 runs, and publish the result every week on We measure ourselves, good or bad. If we can't win 5 questions in 90 days, you'll see it there before anyone else.
12. Changes to this method
If we change any rule on this page, we note it here with a date, and it doesn't affect clients already in a sprint: their contract carries the version they signed.
- 29 Oct 2026 · Google AI Overviews: one capture a day of the Google results page (OpenSEO on DataForSEO); the seven captures of the week are the seven repetitions of that question. Before: seven consecutive queries via DataForSEO. Until a week has all seven captures per question, that column is shown as partial and does not count for the rule. The other four engines are unchanged.