Validate demand for tracked prompts

What makes a prompt worth tracking?

Connect the question to a decision a buyer actually needs to make. A feature limitation, an eligibility condition or a supplier comparison may matter even if it is asked infrequently. State the consequence of a wrong answer as well as the evidence that the question arises.

Gather candidates from authorized sales notes, support tickets, interviews, site search and ordinary Search Console query data. Record the source type and date. Keep sensitive customer details out of the public prompt text, using only the information needed to represent the task.

Generated topics and personas can add hypotheses, but their presence in a tool does not establish population demand. The public synthetic-persona paper describes constructed profiles and queries. Treat those outputs as hypotheses for validation.

What does Search Console tell you?

Search Console provides observations about queries associated with your site in Google Search. It does not measure all searches in the market or all prompts submitted to AI assistants. Its documented data limitations include omitted rare queries and truncated tables, so an absent row is not proof that nobody asks the question.

Match the underlying information need and record meaningful differences in wording. A short query such as “museum booking refund policy” and a conversational prompt about cancellation conditions may express related needs. Record the mapping and let a reviewer challenge it; avoid labeling the match as measured AI prompt volume.

If using Rankfor’s connection, open Google Search Console, choose Connect Google account, authorize the intended property and select Link to this project. Check sync status before using stored data. The current integration’s ordinary query data should not be confused with ingestion of Google’s separate generative AI report.

How should you record evidence strength?

Use a compact register:

An illustrative safety-critical implementation question may deserve monitoring after one credible customer incident. A broad “best company in the world” prompt may remain a weak candidate despite producing a stable answer. Relevance, frequency and answer consistency are separate considerations.

How do you test wording without corrupting the trend?

Keep the current core panel unchanged while trying paraphrases in an exploratory set. Record model, language, region and search mode. Use repeated observations to understand variation, following the design-specific approach in the Dice Roll Method preprint.

In Dice Roller Test, enter one prompt, record the optional brand and Memory/Search mode, then choose an allowed iteration count and inspect the saved responses. The current interface allows up to ten iterations in one run; that product limit is not a universal scientific sufficiency rule.

Promote or retire questions at a documented panel revision. Preserve the reason, date and comparable subset for historical reporting. Never remove an unstable or low-performing question simply to improve the chart.