Data Sources & Limitations
Exactly where the data behind this site comes from, what we compute, and the limits you should keep in mind when you read a score or a finding.
This page is the honest companion to our results. Read it before treating any number on the site as a measurement.
Sources
| Layer | Source | Vintage |
|---|---|---|
| Reviews | UCSD Amazon Reviews 2023 (Home & Kitchen subset), aggregated offline | 2023 |
| Product metadata | Amazon product metadata snapshot (titles, brand, price, materials) | 2023 |
| Labels & scores | Our offline pipeline (cookware.db), computed in Python |
2023 |
| Publication | Static JSON bundles + REST/MCP API; the website only reads and displays | current build |
We do not scrape Amazon in real time, and we do not redistribute raw reviews. Everything public is an aggregate, a count, a rate, or a short representative quote.
What we compute
- Six-dimension product scores (complaint rate, durability, review consistency, long-term satisfaction, maintenance ease, warranty signals). Method and weights: scoring standards .
- Problem mentions β how often reviewers report specific defects (coating peeling, warping, rust, and 20 more). A problem’s percentage on the pain points pages is its share of all labelled problem mentions, not the share of negative reviews.
- Brand rollups β the same six dimensions aggregated per brand, with brand-name normalisation so
CAROTEandCaroteare one entity. - Findings β statistically tested comparisons (price-band inversions, brand divergence, material effects). Each finding carries its metric, sample size, confidence and limitations; see Findings .
Coverage
| Layer | Value |
|---|---|
| Reviews in the corpus | 18,106,825 |
| Products in the catalog | 1,186,539 |
| Products with problem-level evidence | 28,556 |
| Products with a six-dimension score | 43,093 (21,482 above the sample threshold) |
| Brands scored with sufficient evidence | 669 |
A product is only scored when it has review evidence; an absence of evidence is not evidence of a defect, so no-evidence products are down-weighted, never silently labelled bad.
Limitations β read these
- Vintage. The corpus is a 2023 snapshot. We do not claim current pricing, current availability or current model quality.
- Self-selection. Reviews are written by people who chose to write them. Extreme experiences (both praise and anger) are over-represented.
- It is not a failure rate. A “mention rate” is the share of sampled reviews that talk about a problem. It is not a measured probability that a product fails, and it is not adjusted for how much a product was used.
- Labelling error. Problems are detected from review text with keyword/NLP rules. Some mentions are missed and some are false positives.
- Association, not causation. A difference between two brands or materials is a difference in what reviewers report, not proof that one causes the other. Cheaper products may be used harder; popular products have more reviews to complain in.
- No physical testing. We do not buy, handle or lab-test products. This is a data project, not a test kitchen. Where hands-on experience is required to be certain, we say what the data can and cannot settle.
- Price is a snapshot. Prices are from the 2023 metadata snapshot; the live price on Amazon will differ.
- Language and market bias. Reviews are predominantly US English. Non-US markets, non-English reviews and niche brands are under-represented.
- Review integrity. We cannot verify that every review is authentic; platform fraud exists and would bias any review-derived metric.
- Brand normalization is imperfect. Corporate parents, sub-brands and spelling variants can be merged or split imperfectly.
Where AI is used
Pages are assembled from the database by code. AI is used as a writing assistant for some explanatory text, with a numeric guard that rejects any number that is not present in the underlying evidence. AI never invents a statistic, and no page ships without passing the evidence gate. See the editorial policy .
How to cite
The aggregates behind every number here are downloadable as plain CSV/JSON on the Download the Data page (CC BY 4.0). You may cite a finding with its page URL and the metric shown on the page, for example:
“Ceramic reports food sticking at 3.0Γ the rate of PTFE on griddle pans (0.80% vs 0.27% of reviews, n=5,020 vs 31,301).” β AMZ Reviews Data, review-derived, 2023 vintage.
Please keep the sample size and the “2023 review-derived” qualifier. For data licensing, see Contact .
Corrections
If a number here does not match the data, tell us via Contact . Corrections that change a claim are logged on the changelog .