How do you measure it step by step?
- List the questions. Write the questions buyers ask, in their words. Do not name your company in most of them.
- Choose assistants. Pick the ones your buyers use.
- Run each question several times. Use separate days.
- Record each result. Cited, mentioned or absent, plus the date, assistant, mode and sources shown.
- Trace the answers. Find the pages each answer drew from.
- Report ranges. For example, named in two of three runs.
How sure are you of a finding?
| State | Meaning |
|---|---|
| Observed | Seen once. A note, not a measurement. |
| Repeated | The same result across several runs on separate days. |
| Traced | The source page of the answer has been found. |
| Retested | The source was changed and the same prompts were run again. |
What should each record contain?
- The exact prompt text.
- The assistant, its mode and the date.
- The location and whether an account was signed in.
- The result and the sources shown.
What goes wrong most often?
- Judging from one screenshot.
- Prompts that name your company, which flatters the result.
- Testing while signed in to an account with history, which personalises answers.
- Treating a change in one run as a trend.
Related answers
General information, checked 27 October 2026.
Assistants change often. Where a point is inference rather than a published rule, the page says so. Results vary and are never guaranteed.