Research Customers and Markets Without Inventing Demand

Use AI to organize evidence while keeping observations, interpretations, and hypotheses separate.

A founder asks an AI assistant to describe the ideal customer for a new scheduling service. The assistant supplies a name, age, income, favorite brands, daily frustrations, and a quotation about desperately needing the product.

The persona feels specific. None of its details came from a customer.

An invented persona can help you generate questions, but it cannot demonstrate demand. Customer research becomes useful when it changes a decision: whom to serve, which problem to solve, what to test, or whether to stop. AI can accelerate organization and comparison without replacing contact with the people whose behavior matters.

Define the decision and its boundary

Replace “Research the market” with a question tied to a choice. For example: “Should a fictional repair business test appointment windows for customers who currently request evening visits?”

Specify the service area, customer group, offering, and time horizon. A national trend may provide context while saying little about a neighborhood business. A household count may describe a population without revealing purchasing authority, budget, or interest.

The U.S. Small Business Administration distinguishes existing-source research from direct research and encourages examining demand, location, alternatives, and pricing. Those categories help structure questions; they do not answer them for your organization. SBA: Plan your business—market research.

For the worked example, consider Evening Bench, an entirely fictional service concept for booking small household-item repairs. The decision is whether to conduct a limited scheduling trial, not whether to launch a national business.

Work from an evidence packet

Here are five invented interview summaries for practice. They are not actual quotations or field research.

The packet supports several observations about these fictional individuals. It does not show what proportion of a real market holds any preference. Five conveniently selected interviews would not automatically establish that proportion even if they were real.

Preserve record IDs. If the assistant says “Customers need evening service,” ask which records support that statement. I1 suggests a scheduling conflict. I2 prefers mornings. I3 prioritizes price. I4 raises reliability. I5 offers evidence against switching.

Keep three levels of statement distinct

An observation reports what the evidence contains. An interpretation proposes what that observation might mean. A hypothesis predicts something you can investigate.

For I1, the observation is that repairs were postponed because of a conflict with work. One interpretation is that existing appointment windows create a meaningful obstacle. A testable hypothesis is that offering a specific evening window would increase completed bookings among similar customers at the stated price.

The hypothesis adds assumptions. The customer may still reject the price, distrust a new provider, or prefer replacement. Do not let the assistant shorten the chain into “There is proven demand for evening repair.”

Use a prompt that preserves these distinctions:

Analyze the supplied interview packet. Return separate columns for direct observation, supporting record ID, interpretation, competing explanation, and testable hypothesis. Preserve disagreements. Do not invent quotations, demographic details, prevalence estimates, willingness to pay, or market size. Describe who is missing from this packet.

Review whether the AI has silently converted a paraphrase into quotation marks. A real quotation needs a verified source passage and permission for the intended use. A reconstructed sentence is a summary, even when it captures the general idea accurately.

Compare needs and existing solutions

Segment by the problem relevant to your decision before creating a decorative persona. In this packet, useful dimensions include scheduling conflict, sensitivity to price, dependence on reliable arrival, and satisfaction with an existing provider.

One person may fit more than one dimension. Do not force everyone into mutually exclusive categories merely to make a slide look tidy.

Include doing nothing as an alternative. People may tolerate an inconvenience because solving it costs more than it is worth to them. Competition also includes replacement, help from relatives, self-service, and providers outside the category you initially named.

A practical brief might conclude: “The packet suggests several different obstacles. Only I1 directly supports prioritizing an evening-window hypothesis. I3 and I5 caution against treating convenient scheduling as sufficient to win a customer.” That is less dramatic than an ideal-customer story, but more actionable.

Use public data for the question it can answer

For a real U.S. service-area study, Census Business Builder can supply selected demographic and economic information. Record the geography, data year, measure, and source metadata when using it. The tool’s existence is not evidence that a particular local market is attractive. U.S. Census Bureau: Census Business Builder.

Build a research plan that identifies each missing fact and an appropriate source. Population and business counts may come from official statistics. Current competitor hours may require checking the businesses’ own pages. Actual prices may require a dated quote for a defined service. Customer switching reasons require direct evidence from customers.

Do not mix incompatible units. Households are not people, establishments are not firms, and a service-area population is not a customer count. If you create a market-size model, show each assumption separately from observed inputs. A made-up conversion rate does not become evidence when multiplied by an official population figure.

For Evening Bench, no real market size is calculated here. The appropriate deliverable is a plan for obtaining the relevant local evidence.

Ask questions that can challenge your idea

A weak interview asks, “Would a convenient evening service make your life easier?” It invites agreement and supplies the answer’s framing.

A more useful guide begins with recent behavior:

1. Tell me about the last time you needed this kind of repair.
2. What did you do first, and what happened next?
3. Which alternatives did you consider?
4. What made the process difficult, if anything?
5. What did you pay or decide not to pay?
6. What would make your current solution good enough to keep?

After exploring the existing process, you can show a clearly described concept and ask what is confusing or unsuitable. Keep concept reactions separate from actual purchase behavior. A positive response is not a paid booking.

Use authorized notes and minimize unnecessary personal details before giving material to an AI tool. Preserve consent and intended use. Removing a name may not anonymize a distinctive story, so consider the whole record rather than only obvious identifiers.

Turn the findings into a decision brief

The brief should fit the decision: question, evidence available, findings, limitations, competing explanations, next test, and decision owner.

For the fictional packet, a defensible recommendation is to investigate the scheduling-conflict segment further, especially customers who recently postponed repairs. The next step might compare clearly described appointment options in a small authorized trial after confirming staffing capacity. It should also test whether price and reliability dominate timing.

Define what would weaken the hypothesis. If eligible customers repeatedly choose morning windows, reject the service at the quoted price, or report that their current provider is adequate, revise the idea. Do not keep changing the success definition until every response appears supportive.

For students: research a need without fabricating a survey

A class team might investigate equipment booking, tutoring access, or club communications. Start with a fictional packet if the assignment is about analysis rather than collecting personal data. If you conduct research, follow school requirements for participants, consent, and AI use.

Never ask AI to generate survey responses and describe them as classmates’ opinions. Synthetic responses are useful for testing a spreadsheet or interview guide only when clearly labeled. Include disconfirming evidence in your report and explain how it changed your recommendation.

Practice: build a customer brief

Use I1–I5 to prepare an evidence table and a one-page brief. Ask AI to identify interpretations that go beyond the records. Write six neutral interview questions and a public-source research plan for a real or hypothetical location.

Completion check: Every observation has a record, every interpretation is labeled, and every demand or market-size claim has appropriate evidence. Your next research step could change the decision rather than merely decorate it.

For a stretch exercise, add a sixth fictional record contradicting your preferred hypothesis. Revise the brief so the disagreement remains visible.