Open data
Every score on this site is community judgment, so the community should be able to check it. The aggregates behind the pages are downloadable, documented, and free to cite.
The export
Download the CSV. One row per listed tool and per tracked model: slug, name, the recommend score, the review count, six category star aggregates each paired with a source column, and endorsement totals for models.
The numbers are computed by the same helper the pages render from, not a second implementation, so the file cannot quietly disagree with the site. It refreshes on request and is cached for an hour.
What is deliberately not in it
No people. No user ids, no handles, no review text, no per-review rows. Members wrote their reviews to be read here on their own terms, not to be redistributed in someone else’s dataset. If you need to quote a review, link to it.
How the numbers are made
The recommend score is the share of members who would recommend a tool, weighted so that a vote from someone with a verified review counts 3 times as much as an unverified one. Demonstrated use is the only weighting in the system. The full method is on the ranking page.
Category stars blend a seeded community baseline with live per-aspect ratings, which is why a single new rating nudges a category rather than swinging it. A category switches to live data once 3 members have rated it; below that the site shows a dash rather than a confident-looking average.
Every {category}_stars column in the export ships beside a matching {category}_source column: live once enough members have rated it, seeded while the number is still the showcase baseline standing in for that, the same case a dimmed figure marks on a tool page. Check that column before citing a figure as a finished community verdict rather than a starting estimate.
Every row also carries a tier column: cold, emerging, provisional or established, by review count. A cold row has an empty score cell on purpose, because fewer than 3 reviews is not a community verdict yet. Do not read an empty score as a zero.
Imported data never touches a score
Pricing, context windows, release dates and deprecations are synced nightly from public catalogs (OpenRouter, models.dev, Hugging Face). That data fills spec panels and price histories. It never writes a score, an endorsement, or a category rating. When two sources disagree about a price by more than ten percent, the site publishes neither number and files it for a human instead of guessing.
The line matters: the moment an imported metric reaches a score, the score stops meaning “the community said so”.
Citing this
TrustRank community dataset, retrieved from https://trustrank.dev/api/v1/export. Scores reflect member votes and reviews at the time of download and change as people keep reviewing, so note your retrieval date.
There are no sponsored placements and no paid ranking anywhere in this data. If that ever changes, this page changes first.