AI visibility prompt research

A monitoring prompt set decides what you can even see. Anchor it on the brand's own marketing copy and you measure prompts nobody searches; anchor it on real market demand and you measure whether the brand wins the questions buyers actually ask an assistant. The whole discipline is holding the brand back until the end so the research stays grounded in the market.

The play

1. Frame — no brand yet. Collect the category (plain description, no brand name), the named competitor set, the target regions and their languages, and which assistants you'll monitor. Say up front that you're withholding the brand on purpose.
2. Harvest category language, brand last. Pull real buyer phrasing from three wells: keyword and category terminology (ask for any existing keyword or targeting file, then expand from the open web); real questions and pain points mined from communities, forums, Q&A, and "alternatives / best-of" comparison content; and — asked after the open-web pass — first-party sales-call and pre-sales transcripts, the highest-signal source. Tag every item with honest provenance. (See references/open-web-sourcing.md.)
3. Synthesize personas and the non-branded prompt set. Merge all signals into 4–7 personas with their pain points and the solutions they seek, in category terms. Map each to an awareness stage, weighting consideration and decision over pure curiosity. Cluster into a handful of topics. Write each prompt the way a real buyer asks an assistant — a genuine question, not a keyword string — in the target region's language (translate, don't just localize).
4. Brand pass — last, and isolated. Only now introduce the brand. Add competitive comparisons, "alternatives to", pricing, and fit questions as their own dedicated branded topic, flagged branded, kept separate from the category topics.
5. Assemble, validate, hand off. Build one file to your platform's importer contract, validate it (valid enums, one language per row matching its regions, no duplicates, sane per-topic and per-engine spread), show the distributions, and hand it to the user to upload. Never auto-import. (See references/prompt-file-contract.md.)

What good looks like

• What the best operator does first: refuses to look at the brand. The instant failure mode is anchoring on the brand's homepage and generating self-referential prompts nobody types. Real visibility only shows in non-branded category questions, where the brand is competing to be recommended at all.
• The common mistake: shipping keyword strings dressed up as prompts ("best cloud cost tool 2026") instead of how a person actually asks ("what do teams use to see cloud cost per customer?"), and skipping the sales transcripts — the one place the buyer's exact objections and phrasing already live.
• How you know it's good: every prompt traces to a labeled source (social / keyword / sales / competitor), reads like something a human would ask, sits in the right language for its regions, and the set skews to consideration and decision intent. Quality and authenticity beat raw count.

Rules