What makes a good prompt set
Unbranded. “best project management tool for agencies” measures competitive visibility. “Ryze pricing” only confirms you rank for your own name. Branded prompts inflate the numbers and teach you nothing. Buying intent. Questions a prospect asks when close to a decision, not general curiosity. Phrased as people speak. Full questions, not keyword fragments. People type sentences into assistants. Deliberately varied in framing. The same underlying need worded three ways can return three different competitive fields. Large enough to survive noise. Answer engines are volatile. A handful of prompts produces noise; a substantial set produces a trend.Prompt categories worth separating
Read these as separate populations, because they behave differently and mixing them produces a meaningless average.
Problem-led prompts are the tractable ones. They have correct answers, and a clearly written documentation page can supply them. Recommendation prompts lean on what the model already believes about brands, which content work moves slowly if at all.
Platform-worded prompts have a structural ceiling. When a question is phrased around a platform, that platform’s own domains take most of the citations. Track them, but do not expect documentation to win them — a marketplace or app-store listing is the lever there, not a docs page.
Keeping the baseline stable
- Do not edit prompts mid-measurement. Changing wording resets comparability. Start a new set instead and track it separately.
- Add, don’t replace. New prompts start their own baseline from zero.
- Record the engine and date with every run, since engine behaviour changes over time.