Executive Summary
Reviews are the richest first-party product feedback in commerce, given freely by the people who bought the product — and most review datasets waste them by capturing the star and discarding the signal.
A four-star average is the least informative number a review corpus contains. The value is in what reviewers say, which aspect they praise or criticise, whether that is changing, and whether the reviewer actually bought the product. A dataset that captures rating and text but does not structure them into aspects, verify purchase status, or track velocity is a pile of opinions, not an analysable asset.
This report covers product review data scraping done properly: what a review dataset must capture, the traps that make review data misleading, and what the data looks like when built for analysis across Amazon, Flipkart, Nykaa, and Meesho.
This report is published by Product Data Scrape. Sample figures are illustrative of structure, not a live census.
Why Raw Reviews Are Not a Dataset
Volume defeats reading. A single popular SKU can carry thousands of reviews across marketplaces. Nobody reads them at scale; teams read the few most recent or most extreme and mistake them for the whole.
The average hides the cause. A 3.9-star product has a specific problem dragging it down. The star average cannot tell you what. Only aspect-level structure — hydration, fit, battery, packaging, delivery — reveals which attribute is costing the rating.
One review contains several signals. A review praising the fabric and criticising the sizing is positive and negative at once. Whole-review sentiment scoring blurs the two; aspect-level scoring separates them.
Not all reviews are equal. Verified-purchase reviews, reviews with photos, and reviews weighted by helpfulness carry different reliability. A dataset that flattens them into one undifferentiated stream discards that structure.
The Traps
Trap one: capturing the star, not the aspects
A rating is an output. The aspects are the explanation. A review dataset that stops at the star supports counting, not understanding — and understanding is the reason to collect reviews at all.
Trap two: ignoring verified-purchase status
Unverified reviews behave differently from verified ones and are more susceptible to manipulation. A dataset that does not capture verified-purchase status cannot filter for reliability, and any sentiment computed across the mixed pool inherits the noise.
Trap three: treating sentiment as static
A complaint that is rising matters more than one that is stable and larger. Review velocity per aspect — is this criticism accelerating? — is where the actionable signal lives, and it requires reviews to be captured with dates and tracked over time, not scraped once.
Trap four: mixing product and non-product feedback
On marketplaces with delivery and seller layers, reviews conflate the product with the logistics. "Arrived late" is not a product signal. A usable review dataset separates product-related feedback from delivery and seller complaints — a distinction buyers explicitly ask for.
What a Usable Review Dataset Captures
| Field group |
Fields |
| Identity |
review_id, product_id, variant_id, brand, retailer |
| Review |
rating, review_text, review_title, review_date, language |
| Reliability |
verified_purchase, has_photo, helpful_votes |
| Aspects |
aspect, aspect_sentiment, aspect_confidence |
| Classification |
is_product_related, is_delivery_related, is_seller_related |
| Trend |
aspect_velocity, first_seen, sentiment_trend |
| Capture |
captured_at, source_marketplace |
The aspects and classification groups are what turn a review dump into a dataset a product team can act on.
Sample Data: Aspect-Structured Reviews
An illustrative aspect breakdown for one SKU.
| Aspect |
Mentions |
Positive |
Negative |
Trend |
Verified share |
| Product quality |
High |
79% |
21% |
Stable |
88% |
| Sizing / fit |
High |
38% |
62% |
Worsening |
91% |
| Packaging |
Medium |
71% |
29% |
Stable |
85% |
| Value |
Medium |
64% |
36% |
Stable |
87% |
| Delivery (non-product) |
Medium |
55% |
45% |
Stable |
— |
Illustrative figures.
The structured record for a single review:
{
"review_id": "RV-88213",
"product_id": "APP-KURTA-0455",
"variant_id": "M-navy",
"retailer": "marketplace_a",
"captured_at": "2026-07-15T11:00:00+05:30",
"rating": 3,
"review_date": "2026-07-09",
"verified_purchase": true,
"has_photo": true,
"helpful_votes": 12,
"aspects": [
{"aspect": "product_quality", "sentiment": "positive", "confidence": 0.9},
{"aspect": "sizing_fit", "sentiment": "negative", "confidence": 0.94}
],
"classification": {
"is_product_related": true,
"is_delivery_related": false,
"is_seller_related": false
},
"sentiment_summary": "liked_quality_disliked_fit"
}
The single review carries two opposite aspect signals, is verified, has a photo, and is classified as product-related — everything a raw star-and-text capture would collapse into "3 stars." Aggregated across thousands, the sizing/fit aspect emerges as the rising negative dragging the rating, distinct from the strong product-quality sentiment worth protecting.
What the Aggregate Data Reveals
One aspect usually explains a soft rating. A 3.9-star product is rarely mediocre across the board; it is strong on most aspects and dragged by one. Aspect structure finds it.
Rising beats large. A criticism accelerating over recent weeks is a more urgent signal than a larger but stable one. Velocity, not volume, flags the emerging problem.
Verified and unverified diverge. Filtering to verified-purchase reviews frequently shifts the sentiment picture, which is why the flag has to be captured, not assumed.
Product and delivery sentiment move independently. Separating them stops a logistics problem from being misread as a product problem, and vice versa.
Who Uses Review Data
Brands and sellers diagnose which specific attribute is costing a rating, track it after a fix, and benchmark aspect sentiment against competitors.
Product teams use aspect-level, verified feedback to prioritise reformulation and design changes on evidence rather than the loudest recent review.
Review-analytics and consumer-intelligence builders need structured, aspect-tagged review data with full untruncated text via API or bulk export — a requirement buyers state explicitly when evaluating vendors.
Researchers use review corpora for sentiment and linguistics work, where verified-purchase filtering and clean dating are essential to credibility.
Limitations
Aspect extraction and sentiment scoring are probabilistic and category-dependent. Verified-purchase and helpfulness signals vary in availability across marketplaces. Velocity requires continuous capture. We collect publicly available review content only; sample figures illustrate structure rather than audited statistics.
About the Data
This report was produced using product review data scraping methods from Product Data Scrape. We build analysis-ready review datasets across Amazon, Flipkart, Nykaa, Meesho, and other marketplaces — full untruncated review text, ratings, verified-purchase and photo flags, aspect-level sentiment, product-versus-delivery classification, and per-aspect velocity.
Delivered as JSON, CSV, via REST API, or pushed to your warehouse, with historical depth for trend and sentiment work.
Want an aspect-level review sample on your hero SKU? Product Data Scrape will build it across the marketplaces that matter, so you can see which single attribute is dragging your rating.
Product Data Scrape — turning marketplace complexity into decision-ready data.
Singapore
EMAIL
PHONE
.webp)
.webp "thumb")
.webp "thumb")
.webp "thumb")
.webp)
