Set the comparison plan first
Choose the questions, products, visible settings and planned repetitions before reviewing results. Record the observation period and the reason for the schedule. Avoid claiming there is one universally correct number of tests. The scope should fit the decision you want to make and the effort available to inspect answers properly. A fictional service business might compare the same documented supplier questions before and after revising its service information. That plan should not change halfway through simply because another prompt produces a better-looking response.
Keep session conditions documented
Use a consistent approach to conversation history and relevant settings. If the method calls for a fresh conversation, do not add leading follow-ups to some runs and compare them with untouched answers. Record the visible product configuration each time and mark unknown details honestly. If a setting or product option changes, note it as a comparison limit. The aim is a repeatable procedure within what you can observe, not a claim that the underlying service remained identical between dates.
Preserve every planned outcome
Keep answers that omit the brand, refuse the question or fail to complete, alongside favourable responses. If you retry a technical failure, preserve both the failure and the retry with their relationship clear. Do not continue asking until the brand appears and then count only that final answer. Record which runs belong to the planned comparison and which were exploratory. This protects the denominator and lets another reviewer see whether an apparent improvement came from different answers or from a changed selection process.
Compare the same result definitions
Use the same categories for mention, recommendation, citation, relevance and factual accuracy across periods. Inspect the wording behind any change in counts. A brand appearing more often in negative or irrelevant answers is not the same result as being recommended for the intended service. Keep source changes visible and distinguish a link to your site from a third-party mention. If you revise the category definitions, reclassify earlier evidence where possible or start a clearly labelled new series rather than joining incompatible totals.
Separate change from its possible cause
Record website edits and other relevant business changes alongside the observation dates, but do not assume a later answer proves the edit caused it. Note competing explanations and limits in the evidence. A small set of answers supports a statement about those tests, not a universal ranking or future recommendation guarantee. Keep enquiry and referral measurements separate from answer observations. If the business needs to understand commercial impact, collect the appropriate evidence rather than translating a citation count directly into expected customers.
Report a bounded result and next action
State the prompt-set version, completed runs, missing evidence and actual counts before interpreting the pattern. Include examples that explain material improvements or problems. Avoid a single percentage without its underlying numbers. Decide whether the evidence supports another content correction, a longer observation period or no change. Keep the original answers available for review. A useful repeat-check report helps the team choose its next investigation while remaining honest about what was observed and what the exercise cannot establish.