How to choose prompts for AI visibility tracking
How to build a set of buyer questions worth tracking in AI search: where the questions come from, how to group them, and how to keep the comparison fair.
Steen Stones · Reviewed
Photo: Leeloo The First / Pexels
Track the questions your buyers actually ask before they choose a provider, taken from their own words, not from a keyword list. Keep questions that name your company separate from questions that do not. Group them by buying situation, keep the wording fixed, and ask each one more than once. A small, well-chosen set is worth more than a large random one.
Avouch’s free scan shows which companies AI recommends in your market.
Why the question set matters
- The buying situation changes the shortlist. In our research, a small change to the buyer’s brief often changed most of the companies AI suggested. Track each buying situation as its own group.
- Keyword-style questions miss what buyers ask. In our data protection study, 25% of real buyer questions were about cost or contract terms. None of the keyword phrases were.
- Answers change all the time. About half the websites AI cited for a question changed from one day to the next. SparkToro found less than a 1 in 100 chance that ChatGPT or Google’s AI would give the same list of brands twice.
| Buying brief that changed | ChatGPT | Gemini | Claude |
|---|---|---|---|
| IT support for a 30-person company | 43% | 67% | 83% |
| IT and phones together | 14% | 44% | 33% |
| App development, £200k charity brief | 0% | 12% | 71% |
| Penetration testing, software buyer on a budget | 0% | 22% | 20% |
See how AI shortlists differ between buying briefs for the full study.
1. Start from your buyers’ own words
Take the questions from sales calls, interviews, support requests, lost-deal notes and reviews, and check who is speaking before you trust a public post. Our guide to jobs to be done shows how to turn what buyers say into questions.
For each question, write down the buyer’s need, where the question came from and which of your services it relates to. Treat questions an AI tool suggests as ideas to check, not as evidence.
2. Separate discovery questions from questions that name you
| Test | Example question | What you are checking |
|---|---|---|
| Discovery | Which IT support providers could help a growing UK software company change supplier? | Whether you appear when nobody names you |
| Named company | What does Example MSP include in its support service? | What AI says about you once you are named |
Add the other kinds of question buyers ask. Semrush suggests four: buying questions (“best IT support for…”), reputation questions (“is [company] any good?”), competitor questions (“alternatives to [competitor]”) and gap questions, where competitors appear and you do not. Use them as ideas, not a fixed mix.
Never tell the AI tool to recommend you, and use words such as “affordable” or “enterprise” only when they describe the buyer you want.
3. Group by situation, and keep a stable core
Give each question a short code and save its exact wording, buyer need, source, service, country and whether it names you.
Group questions by buying situation. Keep useful wording variations, labelled, so near-duplicates do not dominate the report.
Keep a stable core that you compare over time, and test new questions in a separate group first.
When the core changes, start a new comparison period. Save the old wording and the reason.
We do this ourselves. When Avouch moved its focus to tech buyers on 19 October 2026, we replaced 11 of our 35 tracked questions and started a new comparison period, so old and new results are never mixed.
Semrush’s advice on size is a useful check: 25 well-chosen questions are worth more than 500 random ones.
4. Record the same things every time
- the exact question, the buyer need and the source;
- the complete answer and the addresses of the cited pages;
- the date, the AI tool and any visible model information;
- search, location, account and conversation settings;
- mentions, recommendations and citations, as separate fields;
- position and sentiment, with the sentence that supports your judgement;
- your share-of-voice rule and named competitor set;
- errors, kept separate from answers that left you out.
Start each check in a fresh conversation, with memory and personal instructions turned off. Our guide to AI visibility metrics explains each measure.
5. Ask each question more than once
One answer is a sample. Ask each question on several days and compare the weeks, not single answers. Keep every answer, including the ones that leave you out. SparkToro notes that nobody yet knows exactly how many runs give a reliable result, so treat early reports as a direction of travel. See how often AI answers change their sources.
6. Use the results to choose the work
Open the pages cited in the answers before you decide what to do. A missing service explanation, a wrong company description and a competitor that keeps appearing each need different work. Our guides to checking what AI says about your company and why AI recommends your competitors show the next steps.
How Avouch can help
In the Recommendation Index™ assessment, we build your question set from your buyers’ own words, group it by the situations you sell, and run it across the main AI tools. Where it is part of an agreed engagement, we ask the same questions every week and track your Recommendation Rate™, with scope and price agreed before we start.
Want to know where you stand?
Start with the free scan. For a fuller review and a written plan, choose the paid Recommendation Index™ assessment.