How the ranking works.
Everyone has a “best restaurants” list. Ours is different in one way: it's the agreement of independent judges, and we show you the receipts. Here's the exact method — no black box.
1 · The sources
Each source is a different kind of judge. Its weight reflects how much signal it carries; its category is what makes it independent from the others.
| Source | Kind of judge | Weight |
|---|---|---|
| Eater SF | Food-media guide | 0.9 |
| The Infatuation | Food-media guide | 0.8 |
| SF Standard | Food-media guide | 0.8 |
| MICHELIN Guide | Anonymous inspectors | 1.0 |
| r/sf locals | Locals | 0.6 |
| 7x7 | Food-media guide | 0.7 |
| Bayrrito Watch | Locals | 0.6 |
We dropped the SF Chronicle: its lists are paywalled and can't be verified or linked, so it would be a receipt you couldn't check. Every source above links back to the original.
2 · Position within a source
An ordered list (a critic's #1–#10, or a locals' ranking) gives more credit near the top: credit = 1 − (rank − 1) / length. An unordered “these are the good ones” guide gives every place full inclusion credit. MICHELIN is tiered: a Bib Gourmand or star counts more than a plain listing — and, crucially, can't be bought with fake reviews.
3 · Consensus, not volume
Here's the important part. Being on three food-media guides isn't three independent endorsements — those outlets read each other. So we group sources by kind of judge and apply diminishing returns within a group: the second source in a category counts half, the third a quarter. Across different categories, everything counts full.
The result: to reach the top, critics and MICHELIN and locals all have to agree.That's the signal a single restaurant can't manufacture — and it's why the ranking resists gaming.
4 · What it does in practice
On SF pizza, this pushes Tony's Pizza Napoletana and Del Popolo to a tie for #1 — each backed by guides and MICHELIN and locals. It also drops Damnfine from a naïve #5 to #9: it's a critic panel's literal #1, but all its support sits in one category, so the echo gets discounted.
5 · Coverage, stated honestly
Every ranking shows a confidence badge from how many independent categories back it: Strong (3+), Moderate (2), Thin(1). A cuisine with too few sources isn't shipped at all rather than faked. Scores are normalized 0–100 against the leader.
Machine-readable
Every ranking is available as JSON at /api/rank/{geo}/{cuisine} and indexed in /llms.txt. Anything a person can see here, an agent can read directly.