- Fix a set of real buyer prompts and keep it stable; change it only on a schedule.
- Track four numbers per engine: mention rate, first-position rate, citation rate and accuracy.
- AI answers vary between runs, so sample more than once and read trends, not single screenshots.
- Report the raw answers alongside the numbers so anyone can check them.
Step 1 — Fix the prompt set
Write 20–40 prompts the way your buyers actually ask: discovery ("recommend low-MOQ acne patch manufacturers"), comparison ("compare airless packaging suppliers by MOQ") and verification ("is [company] a real manufacturer?"). Include your category, market and constraints. Keep the set frozen for at least a quarter so month-to-month changes mean something.
Step 2 — Choose engines and capture method
| Platform or surface | Capture method | Keep separate |
|---|---|---|
| ChatGPT | App or API with web search, recorded as distinct cohorts | Consumer app results are not API results; record model and search state |
| Perplexity | Consumer interface or Sonar API, recorded separately | Save the full answer and its cited URLs; record missing citations |
| Gemini | Gemini app or Gemini API with grounding, recorded separately | App and API results remain separate from Google Search AI |
| Google AI Overviews / AI Mode (additional surfaces) | Separate captures of Google Search results | Record which Search AI surface appeared; do not pool with the six-platform matrix |
| Claude | Consumer interface or API with web search, recorded separately | Record search activation and any access failure |
| Microsoft Copilot | Recorded answers from the specified Copilot interface | Bing search results do not substitute for a Copilot answer |
| Grok | Consumer interface or xAI API, recorded separately | Record web search and X search use when available; missing access is not zero visibility |
Freeze the platform surface, model, search state, account state, location, language and prompt set. Full cohort fields and evidence labels are in the measurement methodology.
Step 3 — Track four numbers
| Metric | Definition |
|---|---|
| Mention rate | Share of answers that name your company at all |
| First-position rate | Share of answers where you are named first |
| Citation rate | Share of answers that link to or cite one of your pages |
| Accuracy | Share of mentions where the facts about you (location, products, MOQ) are correct |
Step 4 — Sample, then read trends
The same prompt can name you in one run and not the next. Plan independent repeated runs, keep the cohort fixed and record failures separately. Report each rate as a numerator over valid runs, with sample size and uncertainty. Our September pilot contains different acne-patch prompts on different dates; it does not isolate the effect of wording. A single run remains a snapshot, not evidence of a trend.
Step 5 — Keep answer evidence and publication scope clear
An auditable client report should retain full answer text, cited URLs and capture metadata under the agreed reporting and confidentiality scope. Our public AI visibility scoreboard currently publishes extracted summaries, not full transcripts. Public datasets and client deliverables have different disclosure scopes; see the measurement methodology and editorial policy.