The short answer
Track questions your prospective customers need answered before buying. Build them from sales conversations, support questions and search evidence; cover different needs and constraints. Separate unbranded discovery from branded evaluation, then freeze a version of the set so later comparisons mean something.
It is easy to ask an AI twenty questions. It is surprisingly hard to choose twenty questions that deserve to stand in for your customers.
Start with “What are the best tools in our category?” and you will learn something about broad category recall. Start with a lovingly engineered prompt that describes your product to the letter and you will probably learn something about your own optimism. Neither is a sufficient measurement plan.
Borrow the language buyers already use
Collect a small pile of actual questions before writing any prompts. Ask sales which alternatives come up on calls. Read support tickets about limitations. Look at on-site search and relevant Search Console queries. Remove personal and confidential details before using any customer material.
Treat each source as partial evidence. Search queries describe search behavior, not a complete log of what people ask chatbots. Sales calls overrepresent people who already found you. Reddit can reveal uncertainty and vocabulary, but a popular thread is not proof of search volume.
Label each candidate with the buying decision it supports. “What is project management?” is an educational question. “Which project management tool works for a ten-person studio that bills by the hour?” is much closer to a shortlist. Both may matter; they should not silently carry the same business interpretation.
Build a small matrix, then challenge it
| Buyer need | Constraint | Natural question |
|---|---|---|
| Shortlist | Small team | Which project tools suit a ten-person design studio? |
| Fit | Client approvals | What tools let clients approve work without a paid seat? |
| Switching | Existing files | What should we check before moving our studio off spreadsheets? |
| Risk | Data location | Which project tools document where customer data is stored? |
| Value | Budget | What does a small studio actually need in a paid project tool? |
Have someone outside marketing read the set. Ask whether a buyer would actually say these things, whether the constraints are plausible together, and whether any wording smuggles in your positioning. A question that names your unique feature bundle can be a useful product-fit test. Label it that way.
Use a couple of natural phrasings for an important question rather than dozens of near-identical keyword permutations. For example: “What should our studio use to manage client approvals?” and “Which project tools make client sign-off straightforward?” Store the wording exactly.
Give branded questions their own lane
“Which tools should I consider?” tests discovery. “Is Acme better than Rival B?” tests evaluation after the buyer already knows the names. If you mix them, the branded questions can make your mention rate look excellent by construction.
- Discovery panel: questions without your brand name; report spontaneous inclusion.
- Evaluation panel: comparisons, drawbacks and suitability involving named products; report accuracy and recommendation context.
- Fact-check panel: specific claims about your business; report errors and missing evidence.
A national market belongs in the brief when it affects availability, language, regulation or buying preferences. Writing “in Switzerland” into a prompt does not reproduce a Swiss user’s location, account history or interface. Keep that distinction visible in the report.
A tracking brief you can reuse
Once approved, freeze the panel for the comparison period. Keep a separate exploration list for interesting new questions. When that list becomes important enough to change the main panel, publish a new version and rerun a baseline. Otherwise a “visibility improvement” may simply mean you changed the exam.
Our reports use a structured set of buyer needs and criteria that customers review before collection. That is a practical way to make a bounded study; it is not a claim to know every prompt a market uses. Whatever tool you choose, insist on seeing the questions first.
Sources & context
This guide responds to a question raised in Anyone here actually improving Share of Voice in LLMs? on Reddit. The discussion informed the topic; it is not evidence of the effectiveness of any product mentioned there. Platform documentation is linked alongside the claims it supports.
Examples and figures are illustrative unless explicitly stated otherwise. Written and published by Share of Voice, a Superstellar LLC product. We sell AI visibility reports; our commercial perspective is worth keeping in mind.
Spotted something we should correct? Let us know.