AI can help you think faster, but a convincing answer is not the same as a validated opportunity. Use this repeatable process before making a sourcing decision.
1. Start with a decision, not a prompt
Replace broad requests like “find a winning product” with a falsifiable question: “Could a new brand profitably enter this subcategory at a $28–$34 selling price?” Name the marketplace, search terms, customer segment, target unit economics, and what would make you reject the opportunity.
AI-generated product ideas are useful leads, but product claims, sales estimates, review patterns, and competitor metrics must be traced back to observable evidence. Ask the assistant to label each output as verified, estimated, assumed, or unknown.
2. Build an evidence matrix
| Question | Evidence to inspect | Failure signal |
|---|---|---|
| Is demand real? | Consistent search and sales trends | One-week spike only |
| Is the market accessible? | Seller mix, ratings, review counts, brand share | Few brands dominate most sales |
| Can it earn money? | Land cost, marketplace fees, ads, returns | Margin disappears under plausible costs |
| Is there a buyer problem? | Recurring review complaints | No concrete differentiation possible |
3. Evaluate confidence, not just the conclusion
A claim supported by several independent periods and metrics deserves more confidence than one based on a single screenshot. Read negative customer reviews, compare products at similar price points, and watch for inconsistent estimates. Never let a language model invent historical sales or imply access to private Seller Central data.
Record the date, market, sample, time period, and confidence of each metric. Keep a separate list of unknowns to verify before ordering inventory.
4. Interactive: evidence confidence check
Run product and niche analysis, ask the AI Agent to summarize advantages and risks, then cross-check the underlying demand, competition, and financial inputs before deciding.
Explore Mettra5. What to write in your research note
Include your hypothesis, relevant search phrases, observed leaders, a demand trend summary, a competitor summary, expected price range, adverse scenario, and an explicit decision: investigate further, test in a small batch, or reject. A rejection supported by evidence is a successful research outcome.
Common mistakes
- Treating AI confidence as proof.
- Comparing products from different categories or time periods as if they were equivalent.
- Optimizing a prompt instead of validating unit economics.
- Relying on review sentiment without checking representative sample sizes.
Key takeaway
Use AI to accelerate structured analysis, not to replace real-world validation. Strong decisions survive contradictory evidence and conservative cost assumptions.