Amazon Review Classifier — complaints, returns, praise & worst variants

Part of LeoWorks review analysis tools (same labels, any language): Korean · Japanese · Spanish · Amazon · AliExpress · Google Maps · social comments.

For Amazon sellers, brands, product teams, agencies and AI agents that already scrape Amazon reviews — and need every review labelled with what went wrong, whether the buyer returned it, what buyers love and which variant causes the trouble, for $0.50 per 1,000 reviews, with no prompt writing or LLM key.

Collect and classify in one run — new. Enter Amazon ASINs or product URLs in Products and this Actor collects the reviews and classifies them in the same run. The collection is done by a separate Actor, Amazon Reviews Scraper by junglee, which Apify bills to your account at its own price — about $6 per 1,000 reviews (it can change); this Actor adds $0.50 per 1,000 classified reviews, so about $6.50 per 1,000 reviews all-in. The Store price card shows only this Actor's part. No Amazon login or cookies are used and reviewer names and profiles are switched off. On the Free Apify plan, junglee returns at most 10 reviews per run. Already have reviews from any scraper? Use Reviews dataset — no collection cost.

• Complaint types — defect, not as described, size/fit, performance, usability, shipping, value, safety, seller service (each with a probability)
• Return intent — the buyer returned it, asked for a refund or replacement, or says they will — and return reasons per product
• What buyers praise — build quality, matches the listing, fit, performance, ease of use, looks, value, shipping, seller service
• Action queue — returns, safety reports and "not as described" reviews as a CSV, lowest stars first
• Star vs. text mismatch — 5-star reviews that complain, 1-star reviews that are actually happy
• Summary by product — complaint and praise mix, return reasons, worst variants, verified-purchase share, star distribution, monthly trend and last 90 days vs before, as JSON and CSV
• Your own labels — up to 10 yes/no criteria in plain language (e.g. "mentions the pump or bottle", "compares with another brand")

Use it to: find why a listing gets 1–3 star reviews (defects vs. sizing vs. "not as described") · catch safety reports the day they appear · see which size or colour gets returned · learn what to keep in the listing copy from what buyers praise · compare complaint mix against competitors' listings · feed labelled reviews to a dashboard or an AI agent.

How this differs from our other classifiers: this one is built for product reviews on Amazon: its complaint types separate a broken item from one that works poorly, from one that is not as described, and it reads the purchased variant (size, colour), the review headline, the date and the verified-purchase flag. For local businesses use our Google Maps Review Classifier; for AliExpress, AliExpress Reviews; for social media comments, the Social Comment Classifier.

How to use

Option A — enter the products (the Actor collects the reviews):

1. In Products, paste up to 20 ASINs (10 characters, e.g. B0CWXNS552) or Amazon product URLs (a URL carries its own marketplace; for ASINs choose Marketplace, default amazon.com — an ASIN belongs to one marketplace, the same product can have another ASIN elsewhere).
2. Set Reviews to collect per product (default 50, up to 100) and Collect: order (newest first by default).
3. Click Start. The run log shows the estimated collection and classification cost before anything is collected; with a maximum cost per run the Actor lowers the number of reviews to stay within it. If the collector finds nothing, the run stops with the reason and this Actor charges nothing.
4. Open Output as below. The summary (SUMMARY) records which collector run produced the reviews.

Option B — you already have reviews:

1. Run an Amazon reviews scraper for your products or competitors' — for example Amazon Reviews Scraper (junglee) (see Works with for others).
2. Open this Actor and pick that run's dataset in Reviews dataset (or paste review texts to try it).
3. Optional: add Custom labels, or a filter such as Only reviews newer than: 90 days, Only reviews with at most 3 stars or Verified purchases only. Skipped reviews are not charged.
4. Click Start. 36 reviews take about 3 seconds.
5. In Output, open Classified reviews, Returns, safety & not as described or Labels with probabilities, or download ACTION_QUEUE.csv and REPORT_PRODUCTS.csv. Export the table as CSV, Excel or JSON.

Run it automatically after every scrape

In your Amazon reviews scraper (or its saved task), open Integrations → Add integration → Apify Actor, choose this Actor and set its input to:

{ "datasetId": "{{resource.defaultDatasetId}}" }

Every finished scrape is then classified with no extra step. Schedule the scraper weekly and set Only reviews newer than: 7 days here for a weekly action queue. The same review scraped twice is classified and charged once.

Output sample

Real rows from run F2HHqgj5XFYcTssxy (2026-10-11, build 0.2.1) on public Amazon reviews of 8 products.

review (headline / text)	stars	variant	complaints	praise	return	act on it
Raw edge on grate / I just received this and was trying to remove the grate and it completely cut open my finger …	3	10-QT DualZone	safety, usability	—	no	yes
Disappointed / Returning and buying a different brand. This was way too cheaply made. And not enough large blocks … seems used.	2	1 pack (Deluxe Bag)	defect, not as described	—	yes	yes
Not as described / Not fast charging	3	6FT×2	not as described, performance	—	no	yes
Came without the strap! / This is a very nice mat but I specifically ordered it because it had the strap …	3	Teal	not as described	—	no	yes
Meh / The good: They're soft … The bad: … The area around the ankles … digs into the underside of my ankles. … I have to return these.	3	Size 7.5	usability, performance	usability	yes	yes
manufacturing defects in the fabric / I originally bought these leggings in black, and they are an absolute perfect dupe for Lululemon …	2	Size L	defect	size_fit, usability, performance	no	no
Best out there / Great and true lightning fast charging cables. Perfect length too. Awesome deal for two cords.	5	6FT×2	none	performance, size_fit, value	no	no

In a second run (1hcSoiX78ZDmPBcfY, 36 reviews of 6 products straight from the junglee scraper) the per-product summary flagged, for example, the grey backpack as the worst variant (2 of 3 reviews with a complaint) and a shoe review that was 5 stars but returned for size.

Input example

The form default — two pasted reviews, no dataset needed (about $0.001, 2 seconds):

{
  "texts": [
    "Stopped charging after two weeks. Returning it.",
    "Fits great but the fabric is see-through when I squat."
  ]
}

To classify an Amazon reviews scraper's output, pass its dataset instead — the text, headline, stars, variant and IDs are detected automatically:

{
  "datasetId": "YOUR_AMAZON_REVIEWS_DATASET_ID",
  "reviewsNewerThan": "90 days",
  "verifiedOnly": true,
  "customLabels": ["mentions the pump or bottle"]
}

Pricing

Pay only for classified reviews — no subscription.

Event	Price	When
review-judged	$0.0005	One review classified (complaint types, sentiment, return intent and any custom labels).

That is $0.50 per 1,000 reviews. First run with the form defaults: about $0.001 (2 reviews, 2 seconds).

Cost examples

Reviews	Cost
100	$0.05
1,000	$0.50
10,000	$5.00
100,000	$50.00

With the free $5 monthly Apify credit you can classify about 10,000 reviews.

When the Actor collects the reviews (Products): the collector bills you separately — about $0.006 per review for the default collector, junglee (about $0.001 for the optional web_wanderer collector; the price you pay depends on your Apify plan and can change). All-in with the default collector that is about $6.50 per 1,000 reviews (50 reviews of one product: about $0.33). The run log shows the live estimate before collecting, and SUMMARY.collection records the run.

Items without review text are skipped and not charged. Reviews that fail after retries are reported with an error field and not charged. If you set a maximum cost per run, the Actor stops cleanly when it is reached.

Works with

Field names are matched ignoring case, _ and -, star ratings like "4.0 out of 5 stars" are read as numbers, and variants given as lists are joined. Checked against each scraper's published output schema (2026-10-11); junglee also against real output.

Scraper on Apify Store	Text	Stars	Variant	Date	Verified
Amazon Reviews Scraper (junglee)	reviewDescription + reviewTitle	ratingScore	variant	date	isVerified
Amazon Reviews Extractor (web_wanderer)	reviewText + reviewTitle	rating	variantSpecs (list)	reviewDate	verifiedPurchase
Amazon Reviews Scraper (axesso)	text + title	rating ("5.0 out of 5 stars")	variationList (list)	date	verified
delicious_zebu Amazon reviews	ReviewContent + ReviewTitle	ReviewScore	Variant (list)	ReviewDate	Verified
Scrapers with snake_case fields (review_body, review_title, is_verified_purchase)	✓	✓	✓	✓	✓

Rows without review text (product summary rows, error rows) are skipped and not charged. Anything still not detected can be set under Field mapping. Also: the Apify API and JavaScript/Python clients · Apify Schedules · Claude, Cursor and Claude Code through the Apify MCP server (below).

Call it from code

curl -X POST "https://api.apify.com/v2/acts/leoworks~amazon-review-classifier/run-sync-get-dataset-items?token=YOUR_APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"datasetId": "YOUR_DATASET_ID", "maxStars": 3}'

JavaScript (npm install apify-client):

import { ApifyClient } from 'apify-client';
const client = new ApifyClient({ token: 'YOUR_APIFY_TOKEN' });
const run = await client.actor('leoworks/amazon-review-classifier').call({ datasetId: 'YOUR_DATASET_ID' });
const { items } = await client.dataset(run.defaultDatasetId).listItems();
const products = await client.keyValueStore(run.defaultKeyValueStoreId).getRecord('REPORT');

Python (pip install apify-client):

from apify_client import ApifyClient
client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("leoworks/amazon-review-classifier").call(run_input={"datasetId": "YOUR_DATASET_ID"})
items = client.dataset(run["defaultDatasetId"]).list_items().items
to_fix = [r for r in items if r.get("action")]

The dataset also connects to Google Sheets, Make, Zapier, Slack and webhooks from the run's Integrations tab.

Use with Claude, Cursor or Claude Code (MCP)

Ask your AI agent. Add the Apify MCP server with this Actor as a tool, then say, for example:

• "Why are buyers returning this product? Classify these Amazon reviews: “Stopped charging after two weeks. Returning it.” / “Fits great but the fabric is see-through when I squat.” / “Arrived late and the box was crushed, but the product itself is fine.”"
• "Classify dataset YOUR_DATASET_ID from my Amazon reviews scraper run, only 1–3 star reviews, and list the top 3 return reasons with one example review each."
• "Do these Amazon reviews say the buyer returned the product or wants a refund? “Sent it back, the zipper broke.” / “Love it, bought a second one.” / “Asked for a replacement, it ran small.”"

The agent calls the tool leoworks--amazon-review-classifier and reads the labels with get-dataset-items. Replace YOUR_DATASET_ID with the ID of your own dataset.

Claude Desktop or Cursor (mcp.json):

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com?tools=leoworks/amazon-review-classifier",
      "headers": { "Authorization": "Bearer YOUR_APIFY_TOKEN" }
    }
  }
}

Claude Code: claude mcp add --transport http apify "https://mcp.apify.com?tools=leoworks/amazon-review-classifier" --header "Authorization: Bearer YOUR_APIFY_TOKEN". Leave out the header to sign in with OAuth in the browser instead. We verified this setup with the Apify MCP server (v0.17.4) on 2026-10-10: the agent classified a pasted review in 7 seconds end to end (run LKN9tXlzXAXKWI6yi).

Output

Where	What
Dataset — view Classified reviews	One row per review with flat, spreadsheet-friendly columns
Dataset — view Returns, safety & not as described	The rows to act on — filter action = true
Dataset — view Labels with probabilities	Every label with its probability, for your own thresholds
ACTION_QUEUE.csv	Returns, safety reports and "not as described" reviews, lowest stars first, then newest
REPORT_PRODUCTS.csv	One row per product: reviews, average stars, complaint rate, return rate and top return reasons, top complaints, top praise, worst variant, verified share, last 90 days vs before, star counts, low-sample flag
REPORT (JSON)	The same per product plus monthly trend, worst 5 variants, sentiment shares, custom label shares and example complaints
SUMMARY (JSON)	Run statistics: detected fields, filters, counts, skipped and duplicate reviews

Nothing extra is charged for the files.

One row

From run F2HHqgj5XFYcTssxy (complaintScores left out):

{
  "reviewId": "R2KTQWKW060AUZ",
  "productAsin": "B0B15Q3HCQ",
  "asin": "B0B15Q3HCQ",
  "variant": "Size: 10-QT & DualZone Smart XL (Black, w/ probe)",
  "rating": 3,
  "verified": true,
  "text": "Raw edge on grate\nI just received this and was trying to remove the grate and it completely cut open my finger trying to remove it.",
  "complaintTypes": "safety, usability",
  "praiseTypes": "",
  "sentiment": "negative",
  "returnIntent": false,
  "starTextMismatch": false,
  "action": true,
  "labels": {
    "complaint": [{ "label": "safety", "probability": 0.98 }, { "label": "usability", "probability": 0.89 }],
    "sentiment": { "label": "negative", "probability": 0.96, "confidence": 0.94 },
    "returnIntent": { "probability": 0.04, "matched": false },
    "praise": [],
    "starTextMismatch": { "matched": false, "rating": 3, "sentiment": "negative" }
  }
}

publishedAt, verified, vine, helpfulVotes, reviewUrl and country appear when the dataset has them. When no complaint type passes the threshold, complaintTypes is none. Minimal mode returns IDs and label keys only.

Summary by product (REPORT)

One product from run 1hcSoiX78ZDmPBcfY (some fields left out):

{
  "group": "B07QN9NPKN",
  "reviews": 6,
  "lowSample": true,
  "averageRating": 4.83,
  "complaintRate": 0.333,
  "complaints": [{ "label": "usability", "count": 2, "share": 0.333 }],
  "praise": [{ "label": "quality", "count": 6, "share": 1 }, { "label": "size_fit", "count": 5, "share": 0.833 }, { "label": "performance", "count": 4, "share": 0.667 }, { "label": "value", "count": 3, "share": 0.5 }, { "label": "usability", "count": 2, "share": 0.333 }],
  "returnIntentRate": 0,
  "returnReasons": [],
  "verifiedShare": 1,
  "worstVariants": [{ "variant": "Color: Grey", "reviews": 3, "averageRating": 4.67, "complaintRate": 0.667, "negativeShare": 0 }],
  "last90Days": { "reviews": 2, "averageRating": 5, "complaintRate": 0, "negativeShare": 0 }
}

lowSample is true below 20 classified reviews — shares from a handful of reviews swing a lot. Without login Amazon shows only part of a product's reviews, so scraper datasets are often small.

Photo evidence (beta)

A complaint is stronger with a photo. If your Amazon reviews dataset carries the photos buyers attached (the Amazon reviews scraper by junglee puts them in reviewImages), turn on Judge review photos and the Actor checks whether a photo really shows the damage the review describes. It is off by default and adds no extra charge; it adds a few seconds per photo review.

Setting	What it does
Judge review photos	Only reviews with a complaint or 3 stars or fewer (recommended) or All reviews that have photos
Photos judged per review	The first 1–3 photos of each review (2 by default; at most 2 with "All reviews that have photos")
Photo size	Longest side requested from the photo server: 512, 768 or 1024 pixels
Photo field	Only if your scraper names the photo list differently (auto-detected: reviewImages, reviewImageUrls, images, imageUrls, photos)

Each judged review gets photoEvidence: supports_complaint_product (a real product photo shows a defect on the product and the review complains about it), supports_complaint_packaging (damaged box or bag), damage_seen_text_silent (damage in the photo, not mentioned in the text), no_evidence, or not_product_photo (a catalog image, label, screenshot or something unrelated). A review whose photo supports the complaint is flagged Act on it, and ACTION_QUEUE.csv lists photo-backed product defects first, then packaging damage, then the rest; the queue gets the columns photoEvidence and photoUrl. The dataset view Photo evidence (beta) lists them, and photoStatus says why a review was not judged (no_photo, skipped_by_rule, too_small, download_failed, judge_failed).

• Beta, small check: We ran it on 200 reviews of the Amazon reviews scraper's output; 30 reviews had photos (52 photos). It flagged 11 photos from 7 reviews as supports_complaint_product, and a person checked them by hand: 10 were real damage, and 1 was too ambiguous to judge (a dent or a reflection on the surface) and was left out of the count. Counted without that photo it is 10 of 10 (95% lower bound about 72%); counted as a miss it is 10 of 11 (about 62%). It flagged no packaging damage, so packaging is not measured on Amazon. Of the five photos just under the cut-off for a product defect, three showed real damage, so expect it to miss some defects. The cut-offs were tuned on AliExpress photos (where a larger check was done), not on Amazon. Use it to find evidence quickly, not as proof for a claim.
• Photo size: Amazon lists review photos as 88-pixel thumbnails. The Actor asks Amazon's image server for a larger copy of the same photo (worked for 52 of 52 photos in our check; Amazon does not document this, so it can change). Photos that cannot be read are skipped and the review is still classified.
• Privacy: photos are downloaded only while they are judged and are sent to TypeSafe (the AI model provider) for that purpose. This Actor does not store them, and no question is about people. photoUrl links to the photo on Amazon's server, which can stop working later.

Accuracy

Measured on 179 hand-labelled English Amazon reviews of 12 products (leggings, running shoes, earbuds, cables, moisturizer, shampoo, air fryer, coffee maker, toy blocks, yoga mat, backpack, dog toy; 2026-10-10), mostly the 1–3 star reviews Amazon shows first plus some 4–5 star ones, first 520 characters of headline + text. The questions and the second threshold were adjusted on one part, then checked once on a separate part that was not used for adjusting.

Set	Reviews (with a complaint)	Complaint types	Sentiment	Praise
Check set — not used for adjusting	100 (60)	88%	99%	91%
Adjusting set	79 (48)	92%	96%	—

Praise was added in version 0.2 (2026-10-11) and labelled by hand on the check set before the first run; the rule for extra praise labels (only from 0.8, like complaints) was set once afterwards.

"Complaint types" counts a review as correct when the labels found include a true complaint type and add none that is wrong (reviews without complaints must get none). Many 3-star reviews are mixed, so sentiment accepts either label where both fit. Probabilities are calibrated — use the two thresholds to trade coverage for certainty. Automated labels can be wrong; check samples before making big decisions.

Limits

Item	Limit
Review length	First 4,000 characters are used
Custom labels	Up to 10, each up to 200 characters
Dataset size	Any — datasets are read in pages of 1,000
Language	English (measured). Other languages are accepted and usually work, but accuracy has not been measured — treat them as beta
Speed	36 reviews in about 3 seconds; 10 reviews in parallel by default (up to 40)
Data	Only the review text and headline, its stars, the variant and your custom labels are sent for classification; reviewer names and profiles are not used or output
Filters	Reviews without a star rating or date are kept by the star and date filters
Collecting (Products)	Up to 20 products and 100 reviews per product per run; one Amazon marketplace per ASIN (amazon.com, .co.uk, .de, .fr, .it, .es, .ca, .com.au, .co.jp, .in and more). Amazon shows a limited set of reviews without a login, so a product can return fewer than you asked for
Collecting on the Free Apify plan	junglee returns at most 10 reviews per run (its own limit)
Collecting — tested	amazon.com with junglee, 2026-10-18: 10 reviews collected and classified, 0 errors. The optional web_wanderer collector returned no reviews for amazon.com that day (its notice: Amazon US restricted text reviews); other marketplaces with it were not verified

FAQ

Which AI makes the judgments? Jev, TypeSafe's decision model (version jev-1.13.0, pinned). Jev answers each label with a calibrated probability instead of generated text, so a 0.9 means the same thing for every item. Answers are stable but not fixed: re-running the same input can change a few borderline answers (under 1% of answers in our repeat tests) — filter on probability when an item must never flip.

Does it scrape Amazon? No. It classifies reviews you already have — run an Amazon reviews scraper first (see Works with) or paste texts.

Why both "defect" and "performance"? A broken item and an item that works poorly need different fixes (quality control vs. product design or listing claims), so they are kept apart.

How is this different from Amazon's "Customers say" summary? That summary is one paragraph per product, chosen by Amazon. This Actor labels every review you scrape — yours and competitors' — so you can count return reasons per variant, filter the reviews behind a number, track them month by month and export them.

Why not ask ChatGPT? You can, for a few reviews. For thousands, this gives a probability per label you can filter on, measured accuracy and a fixed price of $0.50 per 1,000, with no prompt to maintain.

Which reviews go into the action queue? Reviews where the buyer returned the item or asked for a refund or replacement, reviews with a safety complaint, and reviews saying the item is not as described (wrong, used, counterfeit, missing parts or claims that are not true).

What is a star vs. text mismatch? A 4–5 star review whose text is negative, or a 1–2 star review whose text is positive. They often hide a problem behind a good rating, or a misunderstanding behind a bad one.

Can I classify only new reviews every week? Yes. Schedule the scraper, connect this Actor through Integrations (see How to use) and set Only reviews newer than: 7 days. Older and duplicate reviews are skipped and not charged.

Are reviews in other languages supported? They are accepted and usually labelled well, but only English is measured — treat other marketplaces' languages as beta.

Compliance and privacy

• Public data, no login. With Products, the reviews are the publicly visible Amazon reviews collected by a separate collector; no Amazon account, cookies or login of yours are used.
• No reviewer profiles. The collector is run with reviewer personal data switched off; only the review text, headline, stars, date, variant and verified-purchase flag are used and output.
• You stay responsible for the use. Amazon's terms apply to collecting the data. This is not legal advice.

Disclaimer

Independent tool — not affiliated with, endorsed by or sponsored by Amazon, or by the authors of the scraper listed above. Names are used only to describe compatible data sources. The reviews shown are public reviews used as examples.

Reviews and support

If this Actor saved you time, a short review on Apify Store helps others find it. Questions or a dataset whose fields are not detected? Open an issue in the Issues tab — we answer within a day.

Changelog

See the Changelog tab.