Methodology — how the review data is counted
Dataset scope, unit of analysis, sample thresholds and the claims we deliberately do not make.
Dataset
The figures come from the UCSD Amazon Reviews 2023 dataset (McAuley Lab), a snapshot covering May 1996 to September 2023. We process one category at a time from the review and metadata files and compute aggregates offline. We never republish review text; only counts, shares and short representative quotes are allowed.
Unit of analysis
The unit is the review, not the product and not the person. A category average is therefore review-weighted: a product with 10,000 reviews carries more weight than a product with 30. That answers “what do buyers say”, not “how does the typical product perform”. Where we report a product-level score (in the cookware section), the sample threshold is applied per product and the sample size is shown next to the score.
Thresholds
- A category note requires at least 300 reviews in the processed sample.
- A brand score requires at least 30 reviews across at least 3 products.
- A product score requires at least 30 reviews for that ASIN.
Below those thresholds we show raw counts with an “insufficient data” label, or we do not publish the page at all. The research desk no longer publishes per-ASIN, brand, store or manufacturer pages: at 3–29 reviews per product in some categories, those pages restated a number without adding evidence.
What we do not claim
- No causal claims. A rating gap between two years or two materials is a description of the corpus, not proof that one thing caused the other.
- No real-time prices. Prices in the dataset are a 2023 snapshot. Current prices must be checked on Amazon.
- No safety verdict. Review counts cannot establish whether a product is safe; they can only show what reviewers reported.
- No representative sample of “all buyers”. People who leave reviews are a self-selected subset, and the share of verified purchases differs by category.
Corrections
If a figure on this site is wrong, email us with the page URL and the number you believe is correct. We will check it against the pipeline and publish a correction note on the page. The full scope and limitations are on data sources & limitations .
Dataset & scope
- Source: UCSD “Amazon Reviews 2023” (McAuley Lab), the public academic dump of Amazon product reviews and metadata.
- Vintage: reviews span 1996–2023; the metadata snapshot is 2023.
- Unit of counting: one review record = one row. Reviews without rating are excluded from rating statistics; ratings without text still count toward distributions.
- Product identity: reviews are rolled up to the parent product (ASIN), so colour/size variants are counted together.
- Category taxonomy: taken verbatim from the official Amazon breadcrumb stored in the metadata (
categories), normalised to slugs. No proprietary re-categorisation.
Definitions
| Statistic | Definition |
|---|---|
| Reviews (sample) | Number of review records in the dump for that entity. Not the number of ratings on Amazon today. |
| Average rating | Arithmetic mean of 1–5 star ratings, rounded to 2–3 decimals. |
| Std deviation (σ) | Population standard deviation of the star ratings — a large σ means opinions are split. |
| Verified purchase | Share of reviews flagged verified_purchase. |
| Reviews with images | Share of reviews that attach at least one image or video. |
| Review-length buckets | Short < 100 chars, medium 100–500, long > 500. |
| Word frequency | Mechanical token counts after lower-casing and stopword removal. No stemming, no sentiment, no AI. |
| Store | Amazon’s store field on the product metadata — the seller / storefront, not always the brand. |
| Manufacturer | Metadata details.Manufacturer, present for a subset of products; blank values are ignored. |
Store and manufacturer totals are derived from the per-product aggregates, so adding them never requires re-reading the reviews.
Sample-size thresholds for standalone pages
Entities below these thresholds still appear inside their parent's tables — they simply do not get a page of their own. This is deliberate: no thin pages.
| Page type | Requirement |
|---|---|
| Category | ≥ 300 reviews (top-level categories always qualify) |
| Brand | ≥ 30 reviews and ≥ 3 products |
| Product | ≥ 30 reviews |
| Word-frequency table | Words must reach ≥ |
Known limits — read before citing
- Historical ≠ current. A 2019 statistic says nothing about a product sold today. Availability, formulation and ownership change.
- Coverage is uneven. Review counts differ by year, category and product. Always compare alongside the sample size.
- Review volume ≠ sales. Reviewing behaviour changes over time and varies by category; do not read sales from review counts.
- Selection effects. People who write reviews are not a random sample of buyers.
- Metadata gaps. Where the official breadcrumb is missing it falls back to the top-level category.
Reuse & citation
Aggregate counts and shares are factual data. You may cite them with a link back to the relevant page. We never publish full review text, reviewer identities or images. The underlying dataset remains subject to the UCSD licence and Amazon’s terms.
Sample size. This page is built from
Data vintage. Source: UCSD Amazon Reviews 2023 (data through September 2023). Historical statistics are not a statement about a product's current quality, availability or price. Review volume is not sales volume.
Scope & deduplication. One review per reviewer per product is counted; ratings without review text still count toward distributions. Coverage is uneven across years and categories, which is why the sample size is shown with every table. See the methodology for the exact definitions.
No raw reviews. We publish aggregated counts and word frequencies only — never full review text, reviewer identities or images.
Affiliate disclosure. Outbound links to Amazon are Amazon Associates links; if you buy through them we may earn a commission at no extra cost to you. Links point to a product page or to an Amazon search for the brand / store / category. We never display live prices or stock because those change constantly.