Keep an honest prompt and citation log

Start with a question you can answer

An observation log is a record of what happened, not a universal visibility score. A useful first question is “Does this recorded answer describe our product’s limitation correctly?” Another is “Which pages were linked for these five evaluation questions?” Avoid a question so broad that any result can be presented as success.

Create the record before running the prompts

Use one row per observed answer. Store the exact prompt, timestamp, product and mode, language, relevant context, answer reference, brand mentions, cited URLs, and review notes. Mark unavailable context as unknown. Keep the original observation separate from your interpretation so another reviewer can disagree without losing the evidence.

An illustrative record might read:

Field	Example
Prompt	How should a small team evaluate inventory tools?
Product and mode	Recorded by the observer
Brand mention	Present, with an ambiguous edition name
Official-site citation	Absent
Action	Clarify edition naming on the product page

This is an example of a logging format, not a captured product response.

Keep four measures separate

• A mention is an occurrence of a name under your counting rule.
• A citation is an identified source reference or link in the answer.
• A referral is a recorded visit under your analytics configuration.
• A business outcome is an event such as a qualified inquiry, with its own attribution limits.

You cannot infer a visit from a citation alone, or a sale from a mention. Microsoft’s AI Performance announcement describes citation telemetry for its supported surfaces and sampled grounding-query information. That coverage is not a complete census of all AI products.

Compare like with like

Keep the prompt set and counting rules stable while measuring a baseline. If you change them, create a new version and label the break. Report raw counts with denominators: for example, “three of eight recorded answers,” not an unexplained percentage. With a small convenience sample, do not turn that percentage into an estimate of market-wide exposure.

Use differences as leads

If a changed page appears in a later answer, check the cited URL and passage. Other variables may have changed too. Record the sequence as an observation; a causal conclusion requires a stronger design. For an early-stage team, resolving an inaccurate description can be a worthwhile result even when no traffic improvement can be attributed to it.