Building a prompt set

A keyword is a fragment typed into a box. A prompt is a whole question, and it usually carries context — budget, team size, industry, a constraint that rules things out. That context is not decoration; it changes which brands the engine names.

Strip it and you are measuring a different question. 'CRM' and 'cheapest CRM for a two-person agency that integrates with Xero' return different shortlists, and only one of them is what your buyer actually asked.

This is why prompt sets are tracked as written. Normalising them into keywords throws away the very thing that decides the answer.

Generated prompt sets are a reasonable starting point and a poor finishing one. They tend toward the obvious questions, which are also the ones every competitor is already visible for.

The questions worth the most come from places where real buyers already speak.

• Sales call recordings — the objections that come up in the last third of a call sit closest to a decision.
• Support tickets — the confusions that recur are questions buyers also ask assistants before purchase.
• Your own search console — long-tail queries phrased as questions are prompts in everything but name.
• Community threads in your category — how people describe the problem in their own words, before they know the vendor vocabulary.
• Lost-deal notes — the comparison that killed the deal is a prompt somebody is asking right now.

Topic grouping tells you what subjects you cover. Stage grouping tells you where in the journey you are losing, which is far more actionable.

Discovery questions describe a problem without naming a solution. Comparison questions weigh named options. Intent questions are close to acting. Objection questions probe for reasons not to buy, and they are the ones most teams never track.

Strong discovery numbers with weak intent numbers is a specific, fixable diagnosis. One blended score hides it completely.

Unbranded prompts are where new brands enter consideration — being named in one puts you in front of someone who was not thinking of you. They are the growth surface.

Branded prompts are where accuracy matters most. Someone asking whether your product is any good is checking specifics, and a stale fact does real damage there.

Comparison prompts sit closest to a decision and deserve disproportionate attention, because the answer frequently carries an explicit recommendation.

Real buyers do not ask once. They ask, get a shortlist, then push back — on price, on implementation, on a specific requirement. The brand named in the final turn is often not the one named first.

Tracking multi-turn journeys is more work and considerably more honest. The turn where someone states their real constraint is the turn that decides the purchase.

Thirty to fifty prompts covers most B2B categories properly. Below twenty you are measuring noise: answers vary run to run, and a small set makes that variance look like a trend.

Weekly scanning suits most teams — frequent enough to catch a real change, slow enough that model variance does not read as movement. Resist the urge to expand the set constantly; a stable set is what makes the trend comparable over time.