AI Search Prompt Research: Build a Panel from Real Buyer Questions
Published: · Reviewed by the geo-rank.ai editorial team · Our methodology
An AI visibility panel is a set of questions you repeatedly ask selected AI systems. Its results describe that set. If the panel contains questions your buyers would never ask, a precise-looking visibility score can still tell you very little about your business.
Build the panel from decisions people need to make: whether a solution fits, what it costs under their conditions, which limitations matter and what switching would involve. Keep the original evidence for those questions. Then define the wording and sampling rules before you start comparing results over time.
Start with the question before turning it into a prompt
Collect recurring questions from sales conversations, support tickets, site search and customer interviews that you are authorized to use. Remove names, account details and other unnecessary personal information. Search Console queries can help identify website demand, but a Google query is not proof that someone typed the same sentence into ChatGPT.
Record the source and confidence. A question heard repeatedly in sales calls is different from a sentence an AI tool generated as a possible idea. Both may help discovery; only one is direct evidence of an observed customer question.
Consider this fictional example for a project-management product:
Scroll horizontally to see all columns.
| Observed or proposed question | Evidence status | Buyer decision | Candidate monitoring question |
|---|---|---|---|
| Can external contractors view only their assigned work? | Recurring question in anonymized sales notes | Access control | Which project-management tools let contractors see only assigned tasks? |
| What happens to our data when we leave? | Question in a procurement checklist | Exit risk | What should a small agency check about data exports before choosing project-management software? |
| Which tool is best for everyone? | Generated suggestion with no customer evidence | Too broad to guide work | Exclude until narrowed to a real situation. |
| Is this product available in French? | Support question | Language fit | Does the named product provide its interface and support in French? |
The table is a method example. It is not a claim that these questions are common in the wider software market.
Separate discovery questions from brand checks
An unbranded question tests whether a company appears while the user explores a need or category. A question that names the company tests what the system says once the company is already known. Combining the two can make a brand seem more discoverable simply because you supplied its name.
Use distinct groups for problem discovery, category selection, comparison, implementation and named-brand accuracy. Add constraints only when they affect a real choice: integration, service area, team size, delivery language or another supported condition.
Do not create dozens of nearly identical sentences to increase the apparent sample size. Three paraphrases of one decision add information about wording sensitivity, but they do not represent three unrelated buyer needs. Keep a shared task ID so you can report results by underlying decision as well as individual prompt.
Make language and market explicit
The language of a question and the market of the buyer are separate variables. A Spanish-language question may concern Spain, Mexico or a service available internationally. Translating an English prompt does not establish that the translation reflects local demand.
For each panel, document the language, relevant location, named products and any assumptions about availability. Have the localized question checked for natural phrasing and preserved intent. If a constraint changes between markets, record the change and avoid presenting the two prompts as perfectly equivalent.
The site's multilingual GEO guide covers maintaining language versions. For measurement, the additional requirement is to preserve exactly which question was asked in each version.
Keep a stable core and a separate discovery set
A stable core lets you examine changes over time. A discovery set lets you investigate new products, objections and emerging terminology. Label them separately so a changing mix of questions does not masquerade as an improvement.
There is no universal number of prompts that makes a panel representative. Choose coverage from the decisions and markets that matter, then check whether additional questions reveal genuinely different information. Start with a manageable set you can inspect answer by answer. Expand when important decisions remain unrepresented.
Before collection, write a compact specification:
Scroll horizontally to see all columns.
| Field | What to record |
|---|---|
| Task ID | The underlying buyer decision, shared across intentional paraphrases. |
| Prompt version | Exact text, language and any geographic condition. |
| Evidence | An anonymized customer source or a clear label that the question is a hypothesis. |
| Platform conditions | Product, mode, date, account/session conditions and available location controls. |
| Outcome definitions | Brand mention, explicit recommendation, cited domain, cited URL and factual accuracy. |
| Missing-data policy | What counts as a collection failure or unavailable feature. |
| Review rule | Who checks ambiguous recommendations and unsupported claims. |
Avoid turning a demand estimate into a census
Tool-generated demand indicators can help prioritize investigation, but inspect their definitions. For example, OtterlyAI's monitoring documentation describes Intent Volume as an estimate. Ahrefs explains that its broad index and custom questions have different collection and coverage characteristics.
Ask whether a number represents observed search queries, modeled prompt demand, a weighted index or counts within the vendor's own dataset. Do not label all four “monthly ChatGPT searches.” Record unknown values as unknown instead of substituting zero.
This also changes how you weight the panel. If you cannot defend a traffic weight, report task groups separately or use an explicitly unweighted view. An unweighted panel is useful as a diagnostic; it is not an estimate of your share of all AI searches.
Run a pilot before freezing the panel
Collect a few repeated answers under the intended conditions. Inspect whether each prompt produces the kind of decision you meant to study. A question may be interpreted as a definition, a shopping request or a troubleshooting problem even when you intended a vendor comparison.
Google describes query fan-out as a technique its AI search features may use to explore related subtopics. Your prompt is therefore an input, not a complete transcript of every retrieval step. Do not assume that a long prompt precisely controls every source the system considers.
Revise ambiguous questions during the pilot, document why, then freeze the core version for your baseline. If a material change becomes necessary later, give it a new version. Preserve old results rather than connecting two different series as if the wording never changed.
Turn panel gaps into content decisions
For each important task, ask whether your site already provides a clear, current answer. A missing article is one possible finding. An incomplete product page, ambiguous support condition or contradictory integration description may be the more direct problem.
Use an AEO content audit to map those questions to existing pages before commissioning new ones. GeoRank can include panel design and a documented review of answers in an agreed GEO audit. The useful output is a set of defensible decisions about your content, with enough evidence to repeat the check.