Module 2 — Measuring · Lesson 3/6
Building a prompt set that measures something
Your measurement is only as good as your prompt set. Don't ask questions containing your brand name — ask what your customer actually asks.
8 min read
The most common mistake in AI visibility measurement is putting your brand name in the prompt. Ask "what is AEOTime?" and the model will of course describe AEOTime — that is a mirror, not a measurement. What you want to know is whether you appear in a question asked by someone who has never heard of you.
Three properties of a good prompt
- No brand name. Frame it around the category, problem or need: "Can you recommend a project tracking tool for small teams?"
- Actually asked. Not invented — pulled from customer calls, support tickets or Search Console queries.
- Answered with a list of options. "What is project management?" asks for a definition, not a recommendation — useless for measurement.
Balance the set across intents
Ask one kind of question and you learn one kind of thing. A healthy set covers several intents:
- Recommendation / "best of" lists: "What's the best tool for …?" — the heart of GEO. Produces direct advice.
- Comparison: "What's the difference between X and Y?" — shows how you're positioned against rivals.
- Alternatives: "What are alternatives to X?" — measures whether a competitor's customer finds you.
- Problem-first: "How do I solve …?" — whether you appear inside a solution recommendation.
- Pricing / cost: "How much does … cost?" — closest to purchase intent.
- Local / niche: "What do teams in Türkiye use for …?" — geographic and sector specificity.
How many prompts is enough?
Ten to twenty prompts is plenty to start; what matters is the spread of intents, not the count. A hundred variations of "best X" tell you less than one prompt from each of six intents. Aim for at least two prompts per intent group, and make sure they genuinely ask different things.
Make the measurement repeatable
Models give different answers to the same prompt at different times. That is by design, not a bug. So a single scan is not "the truth"; what is meaningful is the rate that emerges from asking the same set repeatedly over time. Resist rewriting the set often — change it and your trend line stops being comparable.
Key takeaways
- Keep your brand name out of the prompts; you are measuring what a stranger asks.
- A good prompt has no brand, is genuinely asked, and is answered with a list of options.
- Spread the set across six intent groups; 10–20 prompts is enough to start.
- Keep the set stable — change it and the trend becomes incomparable.
Start measuring what you just learned
AEOTime is free; scans run on your own AI provider key. No card, no plans.