Choose AI visibility prompts from decisions your customers actually face. Start with permitted enquiries, sales conversations and support records, then label any questions your team constructs to cover a missing situation. Record the wording's origin, buyer role, market and the decision the answer should help with.
Check that the question asks about the kind of provider you are. Map it to the evidence a buyer needs before deciding which page to improve. Keep a stable question set for later comparisons and a separate set for exploration.
Save the answers and collection conditions, including failures. These prompts test representation under stated conditions; they do not reveal private ChatGPT question volumes or establish customer demand by themselves.
Begin with conversations you can inspect
Ask sales for the questions that slow a deal, disqualify a request or change a shortlist. Where permission allows, use enquiry text and call notes. Ask delivery which missing facts create confusion after the initial contact.
A Salesforce consultancy might hear that a buyer needs migration help but is uncertain about the data scope. An equipment business might receive a model request without the operating conditions needed to confirm fit. These situations give you a reason to collect a question, even before a keyword tool returns a number.
Keep private details out of the public question set. Record a source reference in the internal working file and paraphrase carefully where needed. Do not turn one interview into a percentage of all buyers.
Label how each question entered the set
Use simple origin labels: an actual enquiry, a permitted interview, an observed public question or a researcher-written prompt. A Google People Also Ask question is observed Google evidence. It is not an observed ChatGPT conversation.
Retain the date, buyer role, market, buying stage and the decision the answer would support. If you cannot establish the origin, mark it unknown rather than adding a credible-sounding source after the fact.
| Question origin | What it establishes |
|---|---|
| A permitted customer enquiry | This person asked this question in that context |
| An owner interview | This owner described the problem or wording |
| A Google question feature | Google displayed the question in a dated result snapshot |
| A constructed prompt | Your team chose it as a test of a defined buying situation |
That last category is useful. It just needs an honest label.
A constructed prompt remains a test question. Google search estimates are not private ChatGPT question counts.
Check that the answer could help choose your kind of business
Broad category wording can produce the wrong comparison. In the saved report for the New York–based Salesforce consultancy, several CRM category answers placed software platforms first. The report was collected from India on 5 September 2026, with each question asked once. It provides a prompt-review warning, not evidence of New York buyer demand.
Before collecting more results, decide which provider types qualify. If you sell consulting, the question should concern consulting work. If you distribute components, do not count a complete-machine manufacturer as interchangeable unless the buyer's task genuinely makes them alternatives.
A narrower question should reflect a real constraint. Adding many flattering qualifications solely to make your company appear creates a weak test.
Build a small question-source matrix
The following questions are constructed examples. They have not been run as an experiment for this guide.
| Audience and decision | Constructed test question | Evidence needed to make it relevant |
|---|---|---|
| Salesforce project sponsor comparing migrations | What should we verify before choosing a consultancy for a CRM migration to Salesforce? | Actual migration scope and project-buyer wording |
| Equipment buyer replacing a component | What information is needed to confirm a replacement for this cylinder model? | Manufacturer identifier and operating context |
| Consultancy owner improving acquisition | How can a Salesforce consultancy attract suitable implementation enquiries? | Owner problem; relevant to TCC's audience |
| Equipment marketer improving enquiries | What should an equipment product page explain before a buyer requests a quote? | Sales and technical-team requirements |
The first two are your client's buyer questions. They can inform a demonstration or the client's pages. The last two concern owners who could buy TCC's services. Do not combine their estimated demand into a single acquisition opportunity.
Cover the decision rather than every wording variation
A useful initial set can include discovery, fit comparison, verification and implementation concerns. The proportions should come from your purpose, not an invented distribution of AI usage.
Several questions can point to one page. A migration service page can explain scope, dependencies and handoff without creating three thin articles. Map each question to the useful evidence a buyer should find, then decide whether an existing page can supply it.
Mark gaps that require a specialist. Missing technical facts do not become a writing task until someone can verify them.
Keep stable measurement and exploration separate
Give the baseline questions stable IDs and preserve their wording. When you add a new market or service, record a new version and report it separately before comparing totals.
Suppose a team replaces difficult questions with easier branded ones. Its mention rate can rise without any improvement in supplier discovery. Keep branded and unbranded results apart so the change is visible.
Exploratory questions are valuable for finding a new issue. They should not silently enter the denominator of last month's report. Show which questions were added, removed or edited and why.
A newly added question should not silently change the denominator of an older visibility result.
Review answers with the same care as questions
Record product, date, collection route and market. Capture failures as well as completed answers. Separate mention, recommendation, citation and factual accuracy.
Do not count a model's request for more information as an automatic failure to recommend you. It may be the appropriate answer to an underspecified technical problem. Review whether the question gave enough information for a responsible shortlist.
The audit checklist gives you a worksheet for that review. Surface Diet explains why source patterns need product and sample context.
Give the question set an owner
Ask the commercial team to review whether the questions still reflect current sales situations. Ask a specialist to review whether technical constraints are accurate. Record changes when the business adds or stops a service.
A carefully maintained set makes the report easier to interpret. Hundreds of plausible prompts with no origin can make it harder to choose the next action.
Request a tailored audit if you want help turning actual buyer questions into an inspectable test. Bring a few recent enquiries or objections you can share safely; that is more useful than a list built only to make a score look good.
Frequently asked questions
Where should buying questions come from?
Start with customer conversations, sales enquiries, support records and dated search research. Record the origin of each question and the buying decision it represents.
Are Google keyword estimates ChatGPT demand?
No. A search estimate describes the provider’s stated market and method. It does not measure private ChatGPT question counts.
How do we choose the questions to test?
Prioritise questions relevant to your verified offer and buyer market. Keep the wording and collection conditions so later observations can be compared.