An AI answer is an observation made under particular conditions. A screenshot can preserve the visible result, but it may not show the full prompt, the settings, the date, or the question group being tested. Without that context, a later reviewer cannot tell whether two observations are comparable.
The purpose of an evidence log is to preserve those conditions. It is not to turn a small set of prompts into a claim about the entire market. Keep the method attached to the result, especially when the number appears in an executive report.
The observation record
| Field | Example or instruction |
|---|---|
| Observation ID | A stable reference for this run |
| Question group | Implementation, comparison, category education, or another defined group |
| Exact prompt | Copy the complete submitted wording |
| Environment | Engine, visible model label, sign-in state, relevant settings |
| Time | Date, time, and timezone |
| Result | Mention, recommendation, citation, or no observed appearance |
| Context | How the brand or page was described |
| Evidence | Permitted record of the answer and exact cited URL |
| Review note | Ambiguities, errors, or conditions worth repeating |
Keep the interpretation separate
A brand mention can be favorable, neutral, or critical. A recommendation can have conditions. A citation might point to a page discussing several companies. Preserve the exact context before assigning a label such as “positive visibility.” If the answer is ambiguous, record that ambiguity instead of forcing a clean category.
For an illustrative panel of 30 observations, five citations means five observed citations in that panel. The number does not establish the percentage of all customer queries that cite the business. A changing mix of branded and unbranded questions can change the total without showing a comparable improvement.
Use the log to choose a next investigation
Group records by the underlying buyer question and inspect patterns. If the same outdated page is repeatedly cited, investigate its content and the newer resource intended to replace it. If a recommendation consistently includes a condition, check whether the website explains that condition accurately.
Keep a stable core of questions when reviewing changes over time, and record additions separately. The log should help the team identify specific pages or claims to inspect. It should not encourage optimizing for a handful of convenient prompts while ignoring the real audience.
For the next implementation step, use What to investigate when AI citations point to old pages.
For the next implementation step, use Store an AEO observation so another reviewer can inspect it.