Method
How we measure
AI answers change from one run to the next. A single answer per prompt is a snapshot, not a measurement. This page explains exactly what we count, so the numbers in your reports can be checked.
What we ask, where and when
- Engines: ChatGPT, Gemini, Google AI Overviews and Google AI Mode on every plan, and Claude as an add-on. For ChatGPT, Gemini and Google we capture the answer a signed-out user sees, with the sources it cites. Claude can't be used signed out, so its answers come from Anthropic's API: Claude Sonnet 5.5 with web search turned on and no instructions of ours, the closest public equivalent of the Claude app. Perplexity, Microsoft Copilot and Grok are coming soon.
- Location: the country you choose for each brand (default United States, English). ChatGPT and Claude can only be located by country.
- Frequency: every prompt runs once a day on every engine, at a random time during US daytime.
- Every answer is stored word for word for at least 12 months. Each number in a report links to the answers behind it.
What counts
- Visibility is the share of answers that name the brand explicitly: by its name, an alias you set, or its domain. Indirect references don't count towards it.
- When a search shows no AI answer (common with AI Overviews), it is reported as no answer, not as an answer that left you out.
- Position is the brand's rank among every brand the answer recommends, not only the competitors you track.
- Share of voice is mentions of the brand divided by mentions of the brand plus its tracked competitors.
- Won without you counts answers that recommend at least one tracked competitor and not the brand.
Ranges, not single numbers
Every rate comes with a 95% range, like 38% (31–45%). Runs of the same prompt are correlated, so we resample prompts, not answers (a bootstrap over prompts, 1,000 resamples). With fewer than 10 prompts or 30 answers we use the Wilson interval instead. When this week's range overlaps last week's, the report says "no clear change". The headline number pools the last 4 weeks, which narrows the range.
How wide the range is depends on how many answers it rests on. Approximate margin (±, percentage points) for a true rate of 50%, the widest case, and 20%:
| Answers | Typical sample | 50% · 90% | 50% · 95% | 20% · 95% |
|---|---|---|---|---|
| 28 | 1 prompt on 1 engine, 4 weeks | ±14.8 | ±17.4 | ±14.4 |
| 175 | 25 prompts on 1 engine, 1 week | ±6.2 | ±7.3 | ±5.9 |
| 700 | 25 prompts on 4 engines, 1 week | ±3.1 | ±3.7 | ±3.0 |
| 2,800 | 25 prompts on 4 engines, 4 weeks | ±1.6 | ±1.9 | ±1.5 |
| 11,200 | 100 prompts on 4 engines, 4 weeks | ±0.8 | ±0.9 | ±0.7 |
Wilson interval with independent answers. Real ranges are a little wider because runs of one prompt move together; the bootstrap accounts for that.
Known limits
- Answers are captured without a signed-in user, so they don't reflect any one person's history or memory.
- Prompts are written to match how buyers ask. Nobody has real prompt volumes yet, so treat prompt choice as a sample, not a census.
- Engines change models and layouts without notice. We record the model name with every answer when the engine reports it.
- Visibility is not traffic or revenue. It tells you whether the AI recommends you, not how many people clicked.