An AI Product Research Evidence Sheet
Turn reviews and demand signals into a traceable evidence sheet with counterexamples, sample counts, and a next test.
The most useful output from AI product research is a table you can audit. Each finding should point to an observation, show which product and situation it applies to, and give you a test that might change your mind. A supplier sample is a decision to buy information. Write down what you need to learn before you order it.
What belongs in one row?
Give each observation a source link, date, market, product model, use case, short paraphrase, counterexample, and next step. Keep any direct quote short and leave out personal identifiers. When several pages repeat the same review, give them one evidence ID.
| Field | Illustrative laundry basket example |
|---|---|
| Use case | A weekly trip to a laundromat with one hand free |
| Observation | A review describes a handle seam opening under a full load |
| Product scope | One model in one listing and one set of visible reviews |
| Counterexample | A separate group of reviews describes light loads without the failure |
| Next test | Ask for reinforced and standard samples, then run a defined load test |
How do you keep the model from overstating demand?
Ask it to cluster use cases and problems, then return a source link for each example. If 9 of 40 sampled reviews mention the handle, label the denominator, time period, and listing. This is an illustrative count from one sample. Keep the same complaint from being counted again in separate categories, and retain evidence that contradicts the emerging pattern.
Amazon's Product Opportunity Explorer can provide additional category and review insights where your seller account has access. Google Trends scores reflect relative interest within the selected time and region. Use them to investigate direction or seasonality and record your filters alongside the evidence. Amazon Product Opportunity Explorer · Google Trends data FAQ
Trace a finding back to a public sample
SCTCamp's pet hair remover demo contains 40 fictional statements. Its keyword rules assign at least one problem label to 36 statements, with 9 in the use-case mismatch group. Search for demo-201 in the sample JSON to see why it landed in the car seat group. In your evidence sheet, record it as a question about cleaning a car seat. A reported product failure would need a separate observation. A follow-up test could compare two brush heads using the same fabric, amount of pet hair, and number of passes.
Follow the public instructions to run uv run mining demo and inspect the label counts and evidence IDs in analysis.json. One statement can receive several labels. Use 40 fictional statements as the denominator when reporting group counts from this exercise. SCTCamp product research demo · Published sample demand map, in Chinese · Synthetic sample JSON · Open-source run instructions
When is a sample worth buying?
Pick a problem that appears in independent sources and can be reproduced cheaply. With the laundry basket, compare a standard handle and a reinforced one. Mark the units with neutral IDs, apply the same load cycles, and write down the failure threshold and maximum acceptable cost increase before looking at the results. A pass earns another test. A fail ends that particular design path.
What should the next operator receive?
Send the evidence sheet, one comparison case, an unresolved question, and a next action with an owner and due date. Use the same IDs when you add new evidence, so a later decision can cite the original observation. SCTCamp's open project repository shows how the product research tool fits into the broader workflow. SCTCamp open-source projects