Prompt research collects questions customers ask and the requirements that shape their choices. The resulting list supports content planning and repeated visibility tests. Record where each question came from instead of presenting generated prompts as observed customer demand.
What people actually ask
The cited ChatGPT usage study grouped much of the measured activity into practical guidance, information seeking, and writing. It provides context for common uses, but it does not estimate demand for your individual product category.How People Use ChatGPT (Chatterji, Cunningham, Deming, Hitzig, Ong, Shan, Wadman; NBER Working Paper 34255), the first large-scale study of real ChatGPT conversation data.
Customers may use short phrases or detailed questions. For example, “best CRM” and “What CRM fits a six-person agency moving from spreadsheets?” describe different tests. Keep both when there is a reason to measure them.
From keyword list to prompt inventory
Use existing keyword research as one input. Add questions from customer interviews, support, sales, and on-site search:
- Start with relevant customer questions and existing keyword research.
- Include requirements such as team size, price range, location, and supported integrations.
- Use customer interviews, support questions, and permitted research records. Remove private information before sending prompts to an external service.
- Record the question’s source, intent, buyer requirements, and version.
- Review the inventory when the offer or customer questions change.
Start with a set small enough to review and run repeatedly. A list of 50–150 prompts is a planning option, not a required minimum. The useful size depends on category breadth, model coverage, and request cost.
The intent taxonomy
Useful groups include information, comparison, recommendation, validation, and transaction. Examples are “how does term life insurance work,” “Webflow versus WordPress,” and “what CRM fits a ten-person firm.” A prompt may serve more than one purpose.
Prioritize questions connected to actual customer decisions. Recommendation prompts show which options appear. Validation prompts show what an answer says about a business already under consideration.
Researching related queries
An AI search product may issue related queries to answer a prompt. Google calls this query fan-out. Where executed queries are available in logs, save them separately from the original user question.Google documents both retrieval-augmented generation and query fan-out, with this example, in its guide to AI features on Google Search.
When the queries are unavailable, list plausible follow-up questions about price, requirements, alternatives, and limitations. Label these as proposed queries. Use them to identify missing information, not as a reconstruction of a private search process.
Group related questions on a page when they concern the same task. Create separate pages when the subjects deserve separate explanations. Avoid producing near-duplicate pages mainly to influence search rankings.The warning against creating separate content for every possible search variation, including fan-out queries, is in the same Google AI features guide.
Testing buyer requirements
Test buyer differences relevant to the category, such as budget, location, company size, or intended use. The companion guide’s hotel experiment provides one recorded example of changes across prompt wording. Test your own requirements before generalizing from that example.Recorded hotel experiment. The July 25, 2026 dataset compares six vehicle wordings. Its sample and exclusions are documented on the page.
Choose several representative prompts and add requirements from real customer groups. Repeat each variant under the same conditions. Compare both the brand lists and the reasons supplied for recommendations before deciding how much additional segmentation is useful.
Turn a customer question into a test
Start with a question from a sales call, support request, or search query. Remove personal details. Record where the question came from. If you generate an additional question yourself, label it as a proposed test rather than observed customer demand.
For the fictional Example Cooling audit, the useful requirements are service area, weekend availability, and diagnostic fee. A generic “best HVAC company” prompt leaves those requirements unspecified. Both prompts can be tested, but they answer different research questions.
| Prompt group | Example question | Purpose |
|---|---|---|
| Discovery | Which HVAC companies serve central Phoenix on weekends? | Unprompted inclusion when the service fits |
| Comparison | Compare the weekend service and published fees of Example Cooling and Company B. | Accuracy and differentiation when both brands are supplied |
| Verification | Does Example Cooling list a diagnostic fee for central Phoenix? | Retrieval and reporting of a specific fact |
Keep branded questions out of the unprompted-discovery total. A model repeating the name already supplied in a question is weak evidence of independent recommendation.
Assign each prompt a stable ID and version. Preserve the exact wording. Group related variants under the same customer need, then examine the variant results before reporting a combined rate.