How we rank

• 50Sourcing evidence — what a venue says it buys, weighted by how specific it is. A named farm is worth far more than “farm-fresh”, and terms that describe the default state of the supply chain — wet-aged beef, a bare breed name — are worth nothing at all.
• 20Farm-partner quality — how good the specific producers it names actually are, researched independently. Naming a commodity processor is weaker evidence than naming the farm that raised the animal.
• 12Registry credentials — verification against closed, authoritative registries — the Kobe beef authorised-merchant list, organic certifiers — rather than a venue’s own claim about itself. On its own a licence is capped at 4.7 of those 12 points, because it proves a venue is authorised to sell one product, not how its kitchen buys; past that, every point a venue earns in the two categories above unlocks a further 0.6, so the full 12 is only reachable by a venue whose own evidence stands up.
• 18Press — added on top, never subtracted, and structured so one respected review or award cannot be outweighed by ten listicles. Lifestyle coverage tracks marketing budget more than sourcing quality, so its contribution is capped relative to a venue’s own evidence: it can amplify a sourcing story, not substitute for one. Anonymous inspection is the deliberate exception — a MICHELIN star or a James Beard award is a judgement of the food by a body that pays its own way and has no advertising relationship with the kitchen, so one such award counts on its own, even for a restaurant that publishes nothing about its sourcing. Anything beyond that first award rejoins the cap.

We measure repetition, not ownership. A restaurant group running several distinct concepts is not affected — Bazaar Meat, Jaleo and China Poblano are three different restaurants that happen to share a back office, and each is judged on its own sourcing. What we rank lower is one concept stamped out dozens or hundreds of times, because a menu designed once for four hundred kitchens is not the kind of buying this guide exists to find.

It is a curve, not a switch. One to three locations carries no adjustment at all — opening a second restaurant is a success, not a demerit. Past that the reduction grows with the number of locations: gentle while a group is still a group, steeper as the same concept is stamped out at scale, and levelling off once a concept passes 150 outlets, where it keeps a fifth of what its evidence earned. It levels off; it never excludes.

Evidence still decides. A replicated concept that genuinely names its producers still outranks an independent that can show us nothing. The model enforces that as a hard rule, not a tendency.

What it actually changed. Almost nothing, and that is the honest headline. The evidence mechanism was already ranking cookie-cutter outlets last, because a menu designed once for four hundred kitchens publishes nothing about where it buys — so the overwhelming majority of the venues this adjustment touches score zero either way, and a share of zero is zero. The preference makes an outcome the evidence had already produced explicit and auditable rather than creating a new one — which is also why we are comfortable publishing it.

No venue pays for placement. There is no advertising, no sponsored ranking and no commission. When a venue page links to a booking platform, that link is there because it is useful to you; we are not paid for it.

We read what is publicly published. Terroir’s crawler reads pages that are served to any visitor. We do not use credentials we were not given, we do not defeat paywalls or login walls, and we do not hammer a site — requests are paced and identify themselves. Where a site refuses us at the network level, we record the refusal rather than route around it.

Provenance is always labelled. A menu we transcribed from a photograph, a claim a journalist reported, and a certification a certifier published are three different kinds of evidence, and we never present one as another. A restaurant’s own claim about its sourcing stays marked as its own claim until an independent source confirms it.

“Evidence” does not always mean sourcing evidence. A venue is in our evidence class when we found something we could score. For most of them that is what the kitchen itself published about what it buys. For some it is press: a MICHELIN star or a James Beard award counts on its own, because it is a judgement of the food by a body that pays its own way — so a restaurant can carry a real score built entirely on coverage. Those pages say “Press coverage found” rather than “Sourcing evidence found”, and the score breakdown shows which of the four categories earned the points. We do not add the other half of that sentence: a kitchen whose sourcing we have not seen published may still buy beautifully and never say so online, and some of these venues do have sourcing claims we are holding back while we re-read the page they came from.

We don’t guess, and we say when we haven’t looked. “We analysed this venue and found no sourcing evidence” and “we haven’t analysed this venue” are different statements and always render differently. A venue we have never reached is never described as having no sourcing.

Scores are low by design. Most venues score zero because we found nothing to score, and even the strongest dossiers rarely pass the low seventies out of 100 — the scale measures published evidence, not how good dinner is. So we show a qualitative band rather than a bare number, and a position only where the evidence earns one. There are exactly three things a venue can be:

• Ranked. We measured this venue and the measurement separates it from every other restaurant on the page, so it gets a number — for example “Exceptional evidence · #5 of 538 scored in Las Vegas, NV”. A position is counted against the restaurants in that city we could actually score, never against every restaurant we hold — most of those published nothing we could measure, and it would be meaningless to say a kitchen beat them. Both numbers appear on a venue page, in words: how many we hold in that city, and how many of them we could score. The city page quotes the same “could score” number, so the two pages agree.
• Tied. We measured it, and it came out exactly equal to one or more others. Then we print a shared position (“Tied #7”) and never a bare number, because we have no evidence that any of them is better than the rest. Equally scored is not the same sentence as equally good.
• Listed, not ranked. Nothing we have read about this venue separates it from the others, so we print no number at all — not a low one, not a last place. These venues appear below a labelled line, in stable alphabetical order, and the page says that is what the order is.

A city with no scoreable evidence gets a list, not a ranking. If we have read a city’s restaurants and found only general marketing language, there is no honest way to say which is best, so that page carries no ranking anywhere on it and says so at the top. The restaurants are still there, still searchable, still filterable by cuisine, and every one of them still shows what we read. What is missing is the one thing we cannot support: the claim that the one at the top is better than the one at the bottom.

Alphabetical order inside a tie is deliberate, and it is not a ranking. It is what keeps a position from jittering between sweeps. Positions are frozen at scoring time and update as new evidence and new venues are reviewed — they are not recomputed live on each visit, so nothing shifts under you while you read. Every claim on a venue page links to the source it came from.