Handing Product Research to an AI Agent: One Conversation From Seed Term to ASIN Shortlist

Product research is not hard at any single step; it is just long. Expand terms, size the category, filter products, estimate sales, then go back and cut. Doing a round by hand means several interfaces, and afterwards it is hard to reconstruct what you filtered on.

The benefit of handing the chain to an agent is not that it is smarter. It is that the process is written down: which endpoint, which parameters, and what got cut are all in the transcript.

This assumes the Skill or MCP is already connected. If not, start with Querying Amazon data in Claude Code without writing code.

The shape of it

Four steps, each narrowing the field:

1. Expand — turn one seed term into related queries with demand figures (Keyword Miner)
2. Size — judge the niche's scale and concentration (Market Research)
3. Filter — pull candidates by price, rating, and seller count (Product Research)
4. Estimate — rank candidates by order of magnitude (BSR Sales Estimate)

The copyable script

Hand over the whole task at once, but require a cost report between steps:

Run a round of product research on Amazon US with the seed term "bath rug". Do it in four steps. Before each step, tell me which endpoint you picked and the expected number of calls, and wait for my confirmation.

Step 1: Expand the seed term into related search terms. Take the top 10 by demand and list the term, its demand metric, and its competition metric.

Step 2: For the two most promising terms, look up the niche category: market size, product concentration, and brand concentration. Tell me whether a handful of brands already dominate it.

Step 3: In the categories that passed step 2, filter products priced $20-80, rated 4.0 or above, with fewer than 500 reviews. Take the top 20 ASINs.

Step 4: Estimate the sales magnitude for those 20 ASINs and sort by it.

Finally output a table: ASIN, title, price, rating, review count, seller count, estimated monthly units, and a one-line reason you kept or cut it.

Constraints:
- Read-only queries. Do not retry repeatedly just to fill the table.
- Mark clearly which fields are estimates.
- If a parameter such as a category ID is missing, ask me. Do not guess.
- If any step fails, stop and give me the error code and request_id.

Why confirm step by step

Asking an agent to "just do the research" tends to fail in two ways: it calls the same endpoint repeatedly to fill out a table, or it continues past a failed step using invented values. Reporting cost per step controls call volume and gives you a chance to catch a wrong direction at step 2 rather than step 4.

Reading the output

Three kinds of field in that table deserve different treatment:

• Platform-visible — price, rating, review count, seller count. A current snapshot; it moves, but you can verify it.
• Estimated — monthly units, revenue, search demand. Model estimates. Useful as orders of magnitude, not as absolute numbers.
• Concentration metrics — product and brand concentration describe how much the top of the category holds. They answer whether you can get in, not how much you would make.

Limits

• Estimates get revised. The same ASIN a week later may read differently.
• Category trees and node IDs differ per marketplace; confirm the definitions line up before comparing across them. See getting started.
• This flow answers "is it worth doing". It says nothing about supply chain, landed cost, or compliance.
• Each successful billable request consumes one call by default; the real total across four steps depends on your page sizes.

The usual next step is building a keyword set for the survivors — see Building your own keyword library — or turning this into a scheduled script, in Bulk product research with three endpoints.