A product page shows one rating and one list of reviews. For a product with variations, that list is shared by every variation in the family.
So "this product has a lot of negative reviews" is often the wrong sentence. The accurate one is "this family has negative reviews, and they are not evenly spread, and they are not all about the same thing". Two splits get you there: by variation, and by what the complaint is about.
Pulling the critical reviews
The reviews endpoint filters by star rating, so you can request only the reviews worth reading closely:
curl --request POST \
--url https://ecommercedataapi.com/v1/amazon/review \
--header "Content-Type: application/json" \
--header "X-API-Key: ${ECOMMERCE_DATA_API_KEY}" \
--data '{
"marketplace": "US",
"asin": "B08CK5Z5Q1",
"page": 1,
"stars": [1, 2, 3]
}'Each review carries star, title, content, date, and skus: the variation the reviewer bought. skus is the field that makes the first split possible.
A real sample, split two ways
On 12 October 2026 we pulled the critical reviews for a dog water fountain family on the US marketplace. The response reported 388 reviews of one to three stars. We read the 20 most recent. All 20 were verified purchases.
Split one: by variation.
Variation purchased (skus) | Reviews |
|---|---|
| The fountain with a standard filter set | 13 |
| The fountain with a larger filter set | 4 |
| A replacement pump, sold as a variation of the same listing | 3 |
Three of the twenty were not about the fountain at all. They were about a spare part listed in the same family. They count toward the rating shoppers see on the fountain's page, and a reader who skims the list takes them as complaints about the main product.
Split two: by what the complaint is about. We sorted each review by its main point:
| Theme | Reviews | 1 star | 2 stars | 3 stars |
|---|---|---|---|---|
| Stopped working or lost power within months | 12 | 9 | 3 | 0 |
| The pet would not use it, or the design did not suit the pet | 6 | 1 | 0 | 5 |
| Hard to keep clean | 2 | 1 | 1 | 0 |
Read across the rows and the two splits start to explain each other.
The one-star reviews and the three-star reviews describe different problems. Nine of the eleven one-star reviews were about the product failing. All five three-star reviews were about fit: the dog was put off by it, or the bowl was the wrong shape for a large dog. Several of those reviewers said the product itself seemed well made.
Those call for different responses. A durability problem is a product and supplier problem. A fit problem is often a listing problem: the page did not say who the product is not for.
Most failures were reported after months, not days. Of the reviewers who gave a time, most said between one and nine months. That is past a standard return window, which is why the complaint ends up in a review instead of a return.
The method in five steps
1. Pull only the critical reviews. Pass stars as [1, 2, 3]. Reading five-star reviews to find problems wastes calls.
2. Group by skus first. Before reading a word, count reviews per variation. Set aside any variation that is a different product: a spare part, a refill, a bundle.
3. Tag each review with one main theme. Keep the list short, five or six themes at most. A review that mentions two problems gets the one the reviewer led with.
4. Cross theme with star rating. A theme concentrated in one-star reviews is what makes buyers angry. A theme concentrated in three-star reviews is what leaves them disappointed. They are rarely the same.
5. Add the date. date is a millisecond timestamp. Group by month to see whether a theme is growing, or started after a particular point such as a new batch.
Tagging is the slow part by hand. It is also a job an AI agent does well when the themes are fixed in advance and it is told to treat review text as data, not instructions. An e-commerce data API for AI agents covers that setup.
The wider method, from a pile of reviews to conclusions you can act on, is in how to analyse Amazon reviews. What the API returns and what official APIs do not is in getting Amazon reviews through an API.
What to do with each result
| What you find | What it points to |
|---|---|
| Negatives concentrated on one variation | A problem with that size, color or batch, not the line |
| Negatives on a spare part or accessory in the family | Consider whether it belongs in the same listing |
| One theme dominating one-star reviews | The defect to fix first |
| A different theme dominating three-star reviews | An expectation the listing should set |
| Failures reported after the return window | A warranty and after-sales question, not only a quality one |
| The same theme on a competitor's product | A weakness of the product type, and a way to differentiate |
The last row is where this turns into product research. Run the same split on the two or three best-selling competitors. A complaint every product in the category shares is a requirement the category has not met.
What this cannot tell you
How common a problem really is. Reviews are written by people who chose to write, and unhappy buyers write more. Twelve failure reviews out of twenty does not mean sixty per cent of units fail.
The cause. "Stopped working" names a symptom. Which component failed needs returned units, not reviews.
Anything about all 388 from the latest 20. The newest reviews show what is being said now. They are not a sample of the whole history.
Whether a review is fair. Some describe misuse. Read them as reports, not verdicts.
Three practical notes
- Check
pagesbefore planning a full pull. In our test each page returned 20 reviews, so 388 reviews meant 20 pages, and each successful call uses one billing unit. For most questions the most recent few pages are enough. - Keep review sources apart. A Vine review and a verified purchase are different evidence. The four sources of an Amazon review explains why.
- Know the family before you start. Look up the product and read its variation list, so you recognise an accessory when it appears in
skus. Parent and child ASINs covers the structure.
Questions
Can I get reviews for only one variation?
The reviews of a family are returned together. Filter on skus after you receive them.
How many reviews do I need to read? Enough for the themes to stop changing. For a first pass, 40 to 60 critical reviews usually show the main ones.
Should I count three-star reviews as negative? Read them separately. In the sample above they described a different problem from the one-star reviews.
Can I do this for a competitor's product? Yes. The endpoint takes any ASIN in the supported marketplaces.