Writing a prompt set

Every AI-visibility number is a share of the prompts you chose to track. Choose prompts your client already wins and the score is flattering and worthless; choose prompts nobody asks and the score is honest and worthless. This is the category's central credibility problem, and no tool solves it for you — but a recipe makes it defensible.

The recipe

1. Start from the journey, not the keywords. Write prompts for four stages: problem-unaware (“why does X keep happening”), solution-aware (“how do I fix X”), vendor-aware (“best X providers in South Africa”), and decision (“is Brand A or Brand B better for X”).
2. Cover the intent buckets. Informational, comparison, local, transactional. They behave differently: comparison and informational prompts trigger AI Overviews far more often than transactional ones, and local intent often resolves to a map pack rather than a generated answer.
3. Source the phrasing from real language. Search Console queries, sales-call transcripts, support tickets, the client's own FAQ, and the forums their buyers read. AI-generated prompt suggestions go last, not first — they regress to the phrasing an AI would use, which is exactly the population you are trying to measure rather than imitate.
4. Include prompts you expect to lose. A prompt set with no losses is a prompt set nobody will believe. Losses are where the retainer lives.
5. Write the whole set down and show the client. The full prompt list is visible in-product and printed with the report. A hidden denominator is an unauditable one.

How many

Fifty prompts per brand is the design point: enough for the aggregate bands to be usable, small enough to stay curated. Twenty is workable for a narrow category. Two hundred is usually a sign the set has stopped being a set and become a keyword dump.

Paraphrases are not duplicates

Two or three phrasings of the same underlying question are a prompt group, not three prompts. Grouping them is not tidiness — past roughly five samples of one phrasing, spending the next call on a *different phrasing* reduces measurement error four to fifteen times faster than spending it on another repeat. See Prompt groups and paraphrases.