# Using the observation worksheet

Use one row per collected observation. Keep the exact question and known settings, then link to the complete saved answer. Give each row a stable record ID. Record unavailable account, model, or device information as unknown. Do not infer it from a different interface.

Suggested collection statuses:

- `formed_answer`: an AI answer was returned. Assess mentions, citations, and factual accuracy separately.
- `no_ai_answer`: the search completed, but the relevant AI answer was absent. A brand-mention score is not applicable.
- `collection_error`: the attempt failed. Record the reason and leave answer-based scores unassessed.

For an answer-bearing observation, a missing brand can be recorded as `no` only after a target brand was selected and the answer was read. When no target was selected, use `not assessed`. Use `unknown` where the evidence does not resolve the question.

The filled MSP record links to `ai-answer-snapshot.en.md` in this same directory. It records a desktop request while preserving the provider’s unreported device field. No site change was performed. Future rows should record the exact published change, URL and date, rather than just the hoped-for effect.

The answer is preserved as source material. Its individual IT, security, or contract recommendations have not all been independently verified.