Name the decision before reviewing the report.
Suppose an owner asks AI where the business could grow. The answer recommends Saturday appointments, cites apparent customer interest and describes a gap in the competition. It sounds specific enough to act on. The first job is to check the reasoning behind the recommendation.
Write down the actual decision: “Should we investigate offering Saturday visits?” That is different from “Should we announce Saturday availability?” A promising question does not establish demand, staffing or an appointment you can deliver.
Keep the report beside its supporting sources. Split a sentence containing several claims into separate checks. Customer preference, competitor availability and your ability to provide the service need different evidence. Do not turn them into a single reassuring conclusion.
Find the source that can answer your question.
Follow an important claim to the original page, document or record. Check that it exists, says what the report attributes to it and applies to the relevant service, place and period. A citation can be real while supporting a narrower statement.
Record the source date and when you checked it. An old published schedule may not describe current availability. A national industry discussion may supply context without answering a question about your own customers.
SBA’s market-research guidance distinguishes using existing sources for broader questions from direct research about a particular business or its customers. Choose the evidence for the question you need answered. A few conversations and a broad report are not interchangeable.
Sort the evidence in a mixed recommendation.
Fictional exercise: A window-cleaning owner receives the four items below in an invented AI research note. The sources, requests and index value are authored examples, not findings about any actual business or local market.
On a small screen, swipe or scroll sideways to read the full table.
| Item in the note | Evidence supplied | What is established? |
|---|---|---|
| “Most local customers want Saturday visits.” | No survey, source, population or date is supplied. | The prevalence claim is unresolved. Keep it as a question; do not repeat it as a finding. |
| “None of your competitors offers Saturdays.” | A short competitor list and one service-page extract that says nothing about operating days. | Neither a complete competitor list nor Saturday unavailability is established. |
| “Interest of 80 means 80 prospective customers.” | A fictional relative search-interest index, with the selected term, region and period identified. | The conclusion uses the wrong unit. An index value is not a buyer count. |
| “Three customers asked about Saturdays, so demand is proven.” | Three distinct fictional requests in a selected inquiry set. Their eventual decisions are unknown. | Three recorded questions exist in this set. They do not establish market prevalence or purchase commitments. |
Keep the last item: it may identify a useful question for the business. Removing the unsupported conclusion does not make the underlying observation worthless. Likewise, the competitor extract can establish what that particular page says; it cannot establish everything the business offers.
Check the measure before interpreting the number.
Ask what a number counts, which group it describes and what period it covers. A count of requests, a survey percentage and an index of search interest answer different questions.
Google’s Trends data FAQ explains that its 0–100 scale represents normalized relative search interest. Equal values in different regions need not represent equal search volumes. Low-volume terms can appear as zero, and Trends is not a scientific poll.
In the fictional example, 80 therefore cannot become 80 customers or 80 percent of local households. Nor would zero establish that nobody wants the service. No actual Trends data was retrieved for this example.
The three requests need their own limit: they come from a selected set, with no basis supplied for representing a wider population. Preserve the questions without inventing a denominator. An inquiry also remains separate from a booking or payment; the measurement guide explains those differences.
Choose the next check, not the biggest conclusion.
The fictional owner has enough to investigate a question, not enough to announce a new availability commitment. A useful next step is to examine the permitted inquiry records: what did each person ask, what answer did the team give, and what happened afterward?
Choose a named reviewer and an agreed review point. If the records do not contain an answer, leave it missing. A later customer discussion or trial needs its own appropriate plan; this worksheet does not authorize contacting people or changing an offer.
Keep operating feasibility separate. Staffing, costs and actual appointments require their own review. If the observed issue is an unanswered request rather than unavailable Saturdays, the inquiry-path check may be the more useful next action.
Save the evidence beside the decision.
Download the blank AI market-research check worksheet (CSV). Use one row per important claim. Keep the source, relevant scope, measure, limitation and next question together. Store completed notes in the workspace appropriate for that information; references are usually enough without copying private customer conversations.
Mark each claim as supported within its stated scope, unresolved, contradicted or an interpretation to examine. Write the reason. Do not average these labels into a market-confidence score: one unsupported claim may be the premise the entire recommendation depends on.
The final decision can be to investigate further, narrow the idea, defer it or retain the current offer. Record why, who owns the next check and what would change the decision. A polished report is useful when it helps you ask a better question, without making the answer up.