Keep the question and its conditions together
The prompt tracker records one observation per row. Preserve the exact question, platform or API, date, market, language, returned answer and citations. Without those fields, a change in the result may simply reflect a different test.
Download the prompt tracker (CSV)The file includes two explicitly illustrative prompts and blank observation fields. Replace the placeholders with your service and market. No observation has been run for these example rows.
Choose questions by intent
- Discovery: a customer describes a problem without naming a provider.
- Evaluation: the customer compares methods, services or tools.
- Price: the customer asks what a service costs or includes.
- Local fit: the question includes a real service area or delivery constraint.
- Brand accuracy: the customer asks about a named business. Report this group separately.
Map each question to the page that should answer it. Use keyword research to inform topics, but do not label the complete prompt as having a measured search volume unless the source actually measures that question.
Record a result without overstating it
Store the raw answer or a permitted evidence reference. Mark whether the business was mentioned and whether an owned page was cited. Record competitor names that actually appeared, rather than adding a preferred comparison list to the result.
A failed API call is not “no mention.” Keep an explicit run status and exclude failed observations from the successful-run denominator. When a platform does not expose a model version or region setting, record “not exposed” instead of guessing.
Repeat a comparable sample
Keep prompt wording stable for a reporting period. Version the set when questions change. Compare the same intent group, platform and conditions; keep old observations rather than overwriting them. Review mention rates, owned citations and answer accuracy separately.
The prompt-testing guide explains our historical sample. The monitoring service page sets out the responsibilities to agree when someone else runs the work.