Every Amazon niche finder does the same thing underneath: it lists category nodes and lets you filter them by a few numbers. What differs between tools is how many filters they expose and how much they decide for you.
Calling the API directly removes the hidden decisions. There is no "opportunity score" to trust, only your own thresholds. That is more work, and it is also the only way to know why a niche was on the list.
This post covers what to filter on, how to read the row that comes back, and what the filters cannot see. Sizing a single category in depth is a different job, covered in where the market size number comes from; this post is the wide pass that decides which categories deserve that work.
What a niche is in this data
The Market Research endpoint returns one row per category node for one month. A row carries about a hundred aggregate fields, and the ones most people filter on fall into four groups:
| Question | Fields to filter on |
|---|---|
| Is there demand? | minAvgUnits, maxAvgUnits, minTopAvgUnits, minAvgRevenue |
| Is it dominated? | maxBrandCrn, maxSellerCrn, maxGoodsCrn, maxAmazonSelfProportion |
| Can a new listing get in? | minNewProportion, minNewCount, maxAvgRatings, newProduct (the definition of "new") |
| Does it pay? | minAvgPrice, minAvgProfit, maxWeight, maxVolume |
Each min has a matching max, so any field can be used as a floor, a ceiling or a band. Ratio fields such as maxBrandCrn and minAvgProfit use a 0–100 scale: the documented examples are 45 and 20, not 0.45 and 0.2.
nodeIdPath restricts the search to one part of the category tree. The documented request uses 172282:281407, and the documented response row belongs to 172282:281407:17854127011:10967801, so the path appears to act as a prefix and one request can return nodes at several depths. Look up paths for your marketplace with the Product Categories endpoint; node IDs differ between marketplaces.
A request that asks four questions at once
curl --request POST \
--url https://ecommercedataapi.com/v1/amazon/market/research \
--header "Content-Type: application/json" \
--header "X-API-Key: ${ECOMMERCE_DATA_API_KEY}" \
--data '{
"marketplace": "US",
"month": "202507",
"nodeIdPath": "172282:281407",
"minAvgUnits": 300,
"maxAvgUnits": 5000,
"maxBrandCrn": 40,
"maxAmazonSelfProportion": 10,
"minNewProportion": 10,
"maxAvgRatings": 3000,
"minAvgProfit": 25,
"page": 1,
"size": 50,
"order": { "field": "avgUnits", "desc": true }
}'The numbers are placeholders showing the shape of a screen, not thresholds to copy. What "enough demand" means depends on your capital, and what counts as a low review barrier depends on how fast you can earn reviews. Write your own numbers down before you run the request, so the result cannot talk you into them.
Two practical notes. size defaults to 50 and allows up to 200, so a wide category can be scanned in a few pages. And month takes yyyyMM; leave it out and the endpoint uses the most recent 30 days.
Reading one row
The documented response is a row for Power Strips. The fields worth reading first:
| Field | Value | What it says |
|---|---|---|
totalProducts | 2815 | Products under the node |
topProducts | 100 | The sample the averages are computed over |
avgUnits / avgRevenue | 14874 / 229732 | Monthly units and revenue per product, in the sample |
avgPrice | 17.83 | Average price |
avgRatings | 10756 | Average rating count: the review barrier |
brands / sellers | 45 / 45 | How many of each |
ebcProportion / fbaProportion | 91 / 96 | Percent with A+ content, percent fulfilled by Amazon |
amazonSelfProportion | 3 | Percent sold by Amazon itself |
sellerNation / sellerProportion | CN / 86 | The most common seller country and its share |
The row can be checked against itself. Multiplying avgUnits by the sample size gives 14,874 × 100 = 1,487,400, which is totalUnits (1,487,414) within rounding. Multiplying avgRevenue by 100 gives 22,973,200, which is totalRevenue (22,973,201.92). Every "avg" field here is an average over the 100-product sample, not over the 2,815 products in the node. It describes the head of the category.
Two more things come out of the same row.
Filters can contradict each other. This row has strong demand, but its avgRatings of 10,756 fails the maxAvgRatings: 3000 filter above. That is the screen working, not a mistake: high volume and a low review barrier rarely arrive together, and each filter you add tightens the trade-off. Loosen one at a time and note which niches appear.
avgPrice and revenue per unit are not the same measure. Dividing totalRevenue by totalUnits gives 15.45, while avgPrice is 17.83. The first weights every product by the units it sells, so it is the closer estimate of what most buyers actually pay; the second is the endpoint's own average and need not agree with it.
What a niche finder cannot see
The head is not the whole category. The sample is the top 100 of 2,815. A niche can look healthy at the top and be empty underneath, or the reverse. Compare the sample size with totalProducts before trusting any average.
Parent and child nodes describe overlapping products. A request under one path can return a node and its sub-nodes. Their rows can describe the same products, so do not add rows together to get a market size.
One month is a snapshot. A niche screened in July looks different in January. For anything seasonal, check the shape over time with seasonal product research before committing.
Profit is a figure for an average product, not for yours. avgProfit of 60.3 does not know your landed cost. Treat it as a way to drop categories that cannot pay, not as a margin forecast. The sales figures are estimates too; see what the API's sales data can and cannot tell you.
From a shortlist to a decision
Keep the screen short. Ten niches that survived your own filters are worth more than two hundred you never looked at.
- Run the wide pass above and keep 10–20 rows.
- Check each survivor's structure with the four distributions in category analysis.
- Pull the products with the product research workflow and read their reviews before you decide.
An AI agent can run the first pass for you as well: the same endpoint is available through the remote MCP server, and handing product research to an agent shows what that conversation looks like.
FAQ
Is this the same as a niche finder tool? The data is the same kind. What you gain is the filters and their thresholds in your own hands, and the ability to run the screen on a schedule instead of by clicking.
Which filter matters most? None on its own. Demand without a review-barrier filter finds saturated categories, and a review-barrier filter without demand finds empty ones. The value is in combining two or three and knowing what each one hides.
How often should I rerun the screen? Whenever the data month changes if you are actively looking, and before any purchase order. A niche list older than a season is history.
Does it work outside the US?
Yes, marketplace accepts the same 13 codes as the other endpoints. Category node IDs and prices are local, so do not carry thresholds from one marketplace to another; the same analysis does not transfer as-is.