Question sets come from real conversations
Question set quality determines whether the whole exercise is worth anything. The usual failure is a team inventing questions that sound reasonable and bear little resemblance to how customers actually ask, producing data that cannot be acted on.
Three reliable sources, in priority order: recorded sales and support conversations, the queries your own site search actually receives, and the questions buyers ask in competitor review sections.
- Thirty to fifty questions: too few lacks statistical meaning, too many cannot be maintained
- Cover informational, comparison and decision intent
- Once fixed, do not silently add or remove — that breaks comparability
Fields every record needs
Completeness now determines whether you can still reconstruct the past in three months. Err on more fields rather than storing a single summary number.
- The question as asked, and when
- Platform and, where visible, model version
- Whether the brand appeared, where, and whether it was recommended
- Full URLs of any cited pages
- Which competitors appeared and what they cited
- The complete AI response
Why the raw response is not optional
Model output is probabilistic. The same question can return different sources at different times. Without the original text you cannot tell afterwards whether a change came from your edits or from sampling noise.
This is also why storing only a mention rate is not enough. When that number moves you have no way to trace why.
The query tool stores full responses by default, which is what makes this kind of tracing possible.
Defining what counts as movement
A baseline exists to be compared against, so agree in advance what counts as change. Without that agreement, 'a bit of improvement' never becomes a decision.
- With small samples, prefer direction over precise figures
- Attribute movement to specific actions rather than to 'optimisation in general'
- Keep the parts that did not move; the absence of change often explains more than the change itself
A baseline is not valuable for its own numbers. It is valuable because it gives every later decision something to be measured against. Sources reviewed on 2026-10-17. This article records a method with verifiable boundaries; it does not claim that any search or AI platform will index, cite or recommend anything.