Which prompts to track for AI visibility: building a client's prompt set

AI visibility numbers are only as good as the questions behind them. Track "what is Acme?" and the client will look famous. Track "best project management tool" for a niche construction-software vendor and they'll look invisible. Neither tells you anything useful.

A good prompt set is the questions a real buyer asks before they've chosen a supplier, written the way they'd type them, in their market and language. About 20 per client is a good default.

The rules

1. Unbranded. Don't name the client. "Best e-bike shop in Hamburg", not "is Northwind Bikes good?" Branded prompts measure reputation, not discovery — useful later, not as the core set.
2. Buyer stage, not research stage. "How do e-bike motors work?" brings encyclopaedic answers with no brands. "Best e-bike for a 20 km commute" brings recommendations.
3. Real wording. Write the way customers write: full questions, the occasional constraint ("under €3,000", "for a team of 10"), no keyword-stuffed fragments.
4. One market at a time. Answers depend on country and language. Track a German client's German market in German, and keep it consistent.
5. Specific enough to matter, broad enough to recur. "Best CRM" is too broad for a niche vendor; "best CRM for property managers in the UK" is the question they need to win.

Five intent groups

Spread the 20 prompts across the ways buyers actually search. These five groups cover most businesses:

A typical split is 6–8 "best options", 3–4 each of alternatives, comparisons and problems, and a few local prompts if the business serves a place.

Where to find the real questions

• Sales and support. Ask the client what prospects ask on first calls. These are the best prompts you'll get.
• Search Console. Long, question-like queries that already bring impressions.
• People Also Ask and related searches for the client's main terms.
• Reviews — of the client and competitors. Customers describe their need in their own words.
• Communities. Reddit threads and forums in the client's niche, especially "what should I use for…" posts.

Examples for three kinds of client

Local service business (dental clinic, Munich)

• best dentist in Munich for anxious patients
• emergency dentist Munich weekend
• invisible braces cost Munich
• alternatives to [large chain clinic] Munich
• dentist Schwabing accepting new patients

B2B software (scheduling tool for clinics)

• best appointment scheduling software for physiotherapy clinics
• alternatives to [market leader] for small clinics
• [competitor A] vs [competitor B] for multi-location practices
• how to reduce no-shows at a physio clinic
• GDPR-compliant patient booking software

E-commerce (outdoor apparel)

• best waterproof hiking jacket for Scotland
• lightweight down jacket under £200
• [brand] vs [brand] rain jacket
• sustainable outdoor clothing brands UK
• what to wear for winter hiking in the Lake District

Notice the pattern: each prompt implies a purchase, and each could reasonably produce a list of named brands.

Mistakes that skew the numbers

• Leading prompts. "Is Northwind the best e-bike shop?" invites a yes.
• Too many branded prompts. They inflate mention rate without saying anything about discovery.
• Only head terms. "Best e-bikes" is dominated by national brands and publishers. A local shop will be absent, which is true and not actionable.
• Changing the set every month. Then month-over-month change is meaningless. Fix the set; when you must change it, note the date and treat it as a new baseline.
• One prompt per intent. One prompt is one data point. Several prompts per intent let you see patterns.

How many prompts?

Enough to cover the main intents, few enough that each one is a deliberate choice. Around 20 per client works well: it's a realistic list of buyer questions, and asked weekly on three engines it produces about 240 answers a month — enough to see real movement. How often to track AI visibility goes into the sample-size maths.

Let the tool draft, then edit

Writing 20 good prompts per client by hand is the slowest part of onboarding. In Ranxii, you enter the client's website and it drafts about 20 buyer questions across these five groups — plus 3–5 likely competitors — for the client's country and language. You edit, delete or add (up to 25 per client), and from then on every prompt is asked on ChatGPT, Gemini and Perplexity each week. The draft is a starting point; the agency's knowledge of the client is what makes it good.