Two answers to the same question can name different companies. This variation is one reason to repeat measurements and keep a documented configuration instead of treating one answer as a stable ranking.
What can change between scans?
A different model, search mode, prompt version or context changes the measurement. Even with the same inputs, generated answers and search results can vary. Record what you know changed rather than guessing why the outcome moved.
More repetitions reveal variation
Basic collects one answer per combination; Medium collects three and Thorough five. More repetitions give you more observations within the scan. They do not make a small sample statistically significant or remove every source of variation.
| Illustrative scan | Brand mentions | Read it as |
|---|---|---|
| First scan | 1 of 3 analysed answers | An initial observation |
| Next scan, same configuration | 2 of 3 analysed answers | A higher observed rate, with little history |
| Scan with a changed prompt | 3 of 3 analysed answers | A new configuration, not a clean continuation |
These fictional counts explain interpretation; they are not a performance claim.
Choose the right comparison
Current visibility can include this week’s available answers. A historical baseline should use clearly defined earlier periods, avoid later data and compare matching configurations. Propps run reviews use the preceding eight complete calendar weeks for their primary historical comparison, with actual available coverage shown.
Pooling repeated answers within a series and week, then weighting comparable weeks and series equally, prevents one unusually large batch from automatically dominating the comparison.
When to act
Investigate repeated changes across relevant prompts and supporting answers. A single poor scan can still expose an inaccurate description worth checking, but it is not proof of a lasting visibility decline. Check failures and missing analyses before interpreting the score.
Keep a log of website changes, campaign launches and prompt edits. Use it to generate questions, not to declare that one change caused another. Read the methodology or set up a useful baseline.