Inputs to the product score. Amazon's own customer ratings are the only source; the Bayesian prior and the percentile conversion are what correct for rating inflation and small-sample noise.
Amazon customer ratings	The star average and rating count Amazon publishes on the listing — the sole input
Bayesian prior	Pulls thin-sample products toward the prior mean until volume earns the score
Global percentile	The headline: where the adjusted average ranks across the whole product catalog
Cohort percentile	The same adjusted average re-ranked inside the product's own category
Reviewer normalization	Not used in this category. Our films, hotels, wines, cruises and books re-center every review against the reviewer who wrote it; the product scrape reads only page one of each listing's reviews, so the same reviewer almost never appears twice and there is no baseline to subtract.
Excluded listings	Products whose scraped rating average or rating count failed the integrity checks are dropped before ranking rather than scored — as are ISBN-numbered listings, which are published media and belong to the Books catalog. They never reach the catalog or the structured data.

Inputs to the product score. Amazon's own customer ratings are the only source; the Bayesian prior and the percentile conversion are what correct for rating inflation and small-sample noise.

How the score works

score1 — 0–100 Bayesian percentile across the whole catalog, the headline (top 5% = 95)
Category cohort — the same score re-ranked within its category
Average rating — the familiar 1–5 Amazon star average, shown alongside
Ratings count — the reviewer base behind the score

A cordless drill and a phone charger can share a 4.6-star average and still sit at very different cohort percentiles — 88th among power tools versus 61st among chargers — because chargers, as a category, run hotter. Every product page shows both numbers side by side: the global percentile and the cohort one. That gap is the information the raw star average throws away.

One difference from the rest of the site: the methodology page describes a per-reviewer normalization — every review re-centered against the baseline of the person who wrote it — and that step is notapplied here. It can't be: each listing is read one review page deep, so a given reviewer turns up on one product and then never again, and a baseline needs repeat appearances to exist. What a product score is instead: Amazon's own star average, shrunk toward a prior so a handful of five-star ratings can't reach the top, then converted to a percentile globally and inside the category. That is a weaker correction than the reviewer graph and a much stronger one than the star average everyone else quotes — the compression and the category effects both come out, which is where nearly all of the misreading happens.

Frequently asked questions

Why not just trust the Amazon star average?+

Because it's compressed and category-blind. Most products with real volume sit between 4.3 and 4.7 stars, and a 4.6 in an easy category (phone cases, cables) is not the same as a 4.6 in a demanding one (laptops, mattresses). The percentile inside a category peer set restores the resolution the raw average loses.

How do you handle fake and incentivized reviews?+

Imperfectly, and we would rather say so. The Bayesian prior dilutes small-sample manipulation: a product flooded with a handful of five-star ratings is pulled toward the prior mean until it accumulates real volume. Separately, listings whose scraped figures fail our integrity checks are dropped rather than ranked — an implausible cluster of perfect star averages, or a rating count repeated verbatim across unrelated products, both of which are collection artifacts rather than real signal, and both of which would otherwise land at the top because the score is built from exactly those two fields. What we cannot do is audit an individual review — we see the aggregate Amazon publishes, so a large, patiently-built fake rating base would still get through.

Are the buy links affiliate links?+

Yes. Each product links to its Amazon listing as an Amazon Associates link, and as an Amazon Associate we earn from qualifying purchases at no extra cost to you. The ranking is computed from the review data first and the link is attached afterwards — nothing about the score depends on where it points, or on whether you click it.

Do products have photos?+

A product shows its photo whenever we actually hold the image file, and a category-colored tile when we don't — the exporter only attaches a photo to a listing after checking the file is on disk, so a page never points at an image that isn't there. We're filling that library in now, and coverage climbs with each export. It will not reach every product — a portion of the stored image URLs no longer resolve — and it makes no difference to the ranking either way, which is computed from the review data alone.

Status: 54,590 products with ≥3 ratings are live in the catalog across 14 categories, each with its own page showing the Bayesian percentile, the cohort percentile and the raw Amazon star average. Listings whose scraped rating average or rating count failed our integrity checks were excluded before ranking, along with ISBN-numbered listings that belong to our Books catalog, so the catalog is smaller than the raw scrape and cleaner at the top. See full methodology for the math.