Identify the measured event
Find the provider’s definition of a mention, recommendation or visibility result. A brand name in a warning is different from a product recommendation, even if both count as mentions. Check how the product handles aliases and similarly named businesses. Read examples that sit near the boundary. You should know what event causes a number to increase before comparing that number across brands or treating it as an indication of commercial performance.
From a number back to the evidence
| You see | Ask | Look at |
|---|---|---|
| Visibility 40% (example) | 40% of what: prompts, answers, weighted? | The tool’s definition and sample count |
| Down 10 points (example) | Bigger than the normal wobble? | Two runs in a row and the sample size |
| Rival ahead | Same rival list, same platforms? | The saved answers for both brands |
| New source cited | Which page, and what does it say? | The cited page itself |
Keep the sample attached
A percentage describes a set of observations. Look for the questions, platforms, locations and dates behind it. If those details differ, two percentages may not be comparable. Keep the sample visible when sharing results with colleagues. Explain what was measured and what was outside the exercise. This does not make the result less useful; it gives people the context they need to use it for the right decision rather than extending it to every possible AI conversation.
Distinguish a source from a mention
An answer can mention a brand without linking to its website. It can also link to a third-party page discussing the brand. Inspect both the captured text and its sources. A useful next action depends on the distinction: improving your own page, correcting a third-party description or investigating why a source appears. Avoid interpreting every source link as a visit or endorsement. Follow the link and understand what the answer appears to draw from before deciding what to change.
Treat small movements carefully
AI answers can vary across repeated observations. Look for a pattern across relevant questions and time rather than reacting to one result. Check whether your question set, model selection or collection method changed. Ask the provider how it presents repeat variation. A sudden movement can be worth investigating without being proof that a campaign worked. Keep the original answers available so a discussion can return to evidence instead of relying entirely on a chart’s shape.
Separate visibility from business outcomes
A tool may show how a brand appears in sampled answers. That is not automatically a measure of site visits, leads or sales. Bring other business evidence into the conversation where available, but do not imply a causal link simply because two lines move together. Decide which changes would warrant a follow-up investigation. Your reporting should connect findings to a practical task while making clear which outcomes the tool measured and which remain a separate question.
Write the decision in plain language
End a reporting cycle with a short explanation: what appeared, what changed, what evidence supports that view and what the team will investigate next. Include one or two representative answers. Name the limits that could change the interpretation. A good report gives someone a sensible next step without asking them to learn an unfamiliar scoring system. If you cannot explain the finding without relying on the product’s headline number, return to the underlying answers before presenting it.