How to build a buyer prompt set for your AI visibility score
A step-by-step method for building the fixed list of buyer prompts behind a visibility score: scope, raw buyer language, writing rules, tagging, sizing, sign-off and the change log.
A visibility score is a fraction. The top half is how often AI engines name you. The bottom half is the prompt set, and it decides what the top half means. Get the prompt set wrong and you can report a rising number every month while buyers never see your name.
We spend the first week of every engagement on this list. Clients are sometimes surprised by that, because writing questions sounds like an afternoon's work. It is an afternoon to write them badly. Below is the method we use to write them well, in the order we do it.
Step 1: write the scope before any prompts
Before anyone drafts a single question, agree what the set is meant to measure. That means five decisions, written on one page:
- Which markets and languages are in scope
- Which buyer roles you sell to
- Which products or product lines count
- Which competitors you want to be compared with
- Which brand names and aliases count as being named
Skipping this is how prompt sets grow to 400 lines. Every stakeholder adds the questions they care about, nobody can say what is out of scope, and the score ends up measuring a blend of five different businesses.
Step 2: collect sentences, not keywords
The raw material is the language buyers use before they know you. Good sources, roughly in order of usefulness:
- Sales call recordings and notes, especially the first call
- Lost-deal notes, where the buyer explained what they picked instead and why
- Support tickets from customers in their first month
- On-site search logs
- Long, question-shaped queries in Search Console
- RFP and tender documents, which spell out requirements in the buyer's own words
- Public community threads where people in your category ask for recommendations
Copy out whole sentences. "We need something that handles approvals for a team spread over three countries" is useful. "Approval workflow software" is a keyword, and keywords tend to produce prompts nobody would actually type into ChatGPT.
Aim for a raw list several times longer than the set you want. Most of it will be cut.
Step 3: turn sentences into prompts
Now rewrite the raw sentences into prompts. A few rules do most of the work.
One need per prompt. If a sentence asks for pricing and integrations and a recommendation, split it or pick the part that matters most.
Write it the way a person asks an assistant. People give context. They mention team size, budget, region, the tool they are replacing. Keep that context, because it changes the answer.
No brand name in scoring prompts. A prompt that contains your name will name you. It belongs elsewhere (see step 7).
Keep constraints that buyers actually have. If your buyers care about a specific integration or regulation, put it in the prompts. If nobody has ever mentioned it on a sales call, leave it out, however proud the product team is of it.
Here is an illustrative before and after for a made-up expense software company:
| Raw buyer sentence | Prompt |
|---|---|
| "Our finance team is drowning in receipts from field staff" | What is the best way for a 60-person field services company to collect employee receipts and expenses? |
| "We tried a spreadsheet, it broke when we hit two currencies" | Which expense tools handle multi-currency reimbursements for a small team in the UK and EU? |
| "Does anything plug into Xero without a consultant?" | Which expense management apps integrate directly with Xero for a company without an IT team? |
Step 4: tag every prompt
Each prompt gets a short row of metadata: market, persona, buying stage, product line, and where the wording came from. The tags are what let you read the score later. Without them, a drop from 30% to 24% is just a drop. With them, you can see that it came from one persona in one market.
The source tag is worth the extra minute. When a prompt keeps scoring zero, you want to know whether it came from twenty sales calls or from one person's hunch.
Step 5: deduplicate by answer, not by wording
Two prompts can look different and test the same thing. Run your near-duplicates a few times each. If they keep producing the same shortlist from the same sources, keep the more natural wording and drop the other.
Query fan-outs help here. These are the narrower searches an engine runs on the way to an answer. When two prompts trigger nearly the same fan-outs, the engine is treating them as one question, and so should your set.
Step 6: size it by segment
Decide how many segments you will report on, then size each one. The reason is arithmetic. Suppose, purely as an example, one persona in one market has 5 prompts. One prompt flipping from named to not named moves that segment by 20 points. With 20 prompts in the segment, the same flip moves it by 5. Thin segments produce dramatic charts that mean nothing.
Our sets usually land between 40 and 100 prompts in total. Fewer segments with more prompts each beats many segments with a handful each.
Step 7: keep three lists, not one
The scoring set. The fixed list behind the headline score. No brand names.
The brand check set. Questions like "what does [Brand] do" and "is [Brand] good for small teams". These test whether engines describe you correctly, and they are reported separately so they cannot prop up the score.
The watch list. Candidate prompts you are observing before promoting them. Prompts where no engine names any brand at all also live here. If every answer is a generic explanation with no vendors in it, nobody can win that prompt, and leaving it in the scoring set just lowers the ceiling.
Step 8: sign off and freeze
Get the scoring set signed off by someone in sales and someone in marketing. Sales will catch prompts no buyer would ask. Marketing will catch missing product lines. Then freeze it for the reporting period.
Step 9: keep a change log
Prompts will need to change. Products launch, markets open, buyer language shifts. Make changes at agreed points, usually once a quarter, and log the date, the prompt, and the reason. For the month a change lands, report the score on both the old set and the new one, so the trend line has a bridge instead of a cliff.
Common mistakes we see in inherited sets
The prompts are written in the company's category jargon, which buyers do not use. Half the set is about one feature. Competitor comparison prompts only include competitors the brand already beats. And brand prompts sit inside the headline score, inflating it. Each of these makes the number look better and tell you less.
How we build the set in Promptwatch
We build and run every prompt set in Promptwatch, because the steps above map onto its prompt tracking directly. Each prompt carries a search volume and a difficulty score, which helps when deciding what earns a place in a thin segment. Query fan-outs are shown per prompt, which is how we do the deduplication in step 5 and find phrasings we missed. Topics and tags hold the segment metadata from step 4, personas let us measure the answer a specific buyer role sees, and country-level tracking covers each market. State and city tracking is there too, from the Business plan and on every agency plan. Google Search Console is one of its integrations, which helps when the raw language in step 2 starts in your own query data.
If you want to test draft prompts before committing, the free Explore plan covers 10 prompts on ChatGPT only. Run a handful of your near-duplicates on it for a couple of weeks. You will see quickly which ones are really the same question.