How rankings work
Companies are ranked with the Bradley-Terry model, a statistical method that distills thousands of anonymous head-to-head votes into a single strength score for every company.
How voting works
Every matchup asks one question: "Would you rather work at X or Y?" You pick one. That's a vote. No accounts. No personal data. One anonymous preference at a time.
Voting runs in a gauntlet format. The winning company stays on screen as the reigning champion and faces a fresh challenger. It keeps its streak until it loses. A champion that wins 6 consecutive matchups retires and preventing any single company from dominating the data.
How votes become a ranking
Every matchup is a pairwise result: company A beat company B. The model takes all of those results and fits one latent strength number p per company such that the predicted win probability matches the observed vote share across every pair. Higher strength = more likely to win any given matchup.
That strength maps to the displayed Points:
The field is anchored so the geometric mean sits at 1500. In practice, companies span roughly 1230–1795 (unless you're Anthropic). The trend arrow tells you which direction the company moved in the last update.
Why the order of votes doesn't matter
A sequential system like ELO processes votes one at a time. The result depends on who played whom first. Shuffle the same votes and you get a different board. In testing, that churn reached ±6 ranks on average and up to 16 positions for individual companies on the same vote set.
Bradley-Terry is a global fit. It doesn't replay the log chronologically but rather solves for the single set of strength values that best explains all votes simultaneously. The same votes, in any order, always produce the same Points for every company.
When rankings update
Rankings update hourly. Every vote is logged immediately, but the solver runs as a single batch — one full recompute of the Points for all companies from all votes. The displayed index reflects the last completed recompute, not an instant per-vote tally.
This is intentional: batch recomputes are what give the board its order-independence guarantee. A live per-vote update would require sequential processing and reintroduce the ordering problem.