← Back to index

Methodology

Japan Dining Index is an independent assessment synthesizing multiple public sources.

Six Dimensions

Prestigeprestige

Established standing — anchored in the published Michelin Guide.

Ratingrating

Aggregated diner rating derived from publicly available review data.

Heatheat

How much current attention the restaurant attracts.

Valuevalue

Value relative to the restaurant's price band.

Foreign-friendlyforeigner

How accessible the restaurant is to non-Japanese-speaking visitors — language, menu, payment, booking.

Stabilityrisk

How consistent the underlying signals are — i.e. how confident you can be in the headline number.

Score Bands

S
80–100
Exceptional
A
72–79
Excellent
B
60–71
Strong
C
50–59
Standard

Principles

What we do not do

What we publish, what we don't

Each restaurant page shows a single overall score (0–100, banded S/A/B/C/D), a breakdown across six dimensions, and structured decision signals — price band, booking difficulty, English-access level. The dimensions are presented separately on the page so you can read the shape of the rating, not just the headline number.

Weights are calibrated quarterly against a held-out validation set of restaurants whose consensus quality is well-established. Current weight ranges (varies by score band):

We don't publish the exact monthly weights because (a) they shift with calibration runs and (b) the input signals are themselves time-varying, so a snapshot would mislead more than it informs. The rubric above, the dimension definitions, the score bands, and the input sources below are the parts that stay stable.

Score distribution

Across 1009 Michelin-recognized restaurants in this index:

Of the 103 restaurants in the S or A bands, 103 (100.0%) also hold one or more Michelin stars in the most recently published Guide. The six-dimension score and the Guide's verdict align tightly at the top of the index and diverge across the long tail — where the rubric weighs value, foreign-friendliness, and stability that the star system doesn't surface, often re-ranking Bib Gourmand and Selected restaurants above lesser-starred ones at the same price point.

Why six dimensions

We tested 3-, 6-, 8-, and 10-dimension rubrics on a held-out validation set. Three dimensions correlated so tightly with the headline Michelin star count that they added no information beyond what the Guide already publishes. Ten dimensions introduced redundancy — "wine list depth", "cocktail program", and "drinks pairing" all loaded onto the same latent factor. Six was the smallest set where each dimension contributed independent variance to user-relevant outcomes (booking decision, return intent, willingness to recommend), and where every dimension had at least one publicly observable signal underpinning it.

Signal sources

Every score is built on publicly observable signals from the following sources (alphabetical, non-exhaustive):

We use only structured signals — numeric scores, counts, declared price bands, booking-platform presence. We do not reproduce review text, photography, or editorial copy from any source.

Data approach

Every input is derived from publicly accessible material — restaurants' own pages, the publicly available editions of the Michelin Guide, public rating platforms, search and maps platforms, and booking platforms. We do not scrape paywalled content, login-walled pages, or anything Michelin sells access to.

We use only the structured signals (scores, counts, classifications, declared price bands) that those sources make publicly available. We do not reproduce any third party's editorial copy, review text, or imagery.

Internal augmentations — translated address strings, normalized area names, synthesized taglines — are produced by our own pipeline and labeled as such; they are not represented as having come from any third party.

Update cadence

The full pipeline is re-run monthly, and on every observed material change. Material changes include: a restaurant closing or moving, a Michelin level change, a major price change, a new platform appearing on the booking landscape, or our discovery of a previously missed restaurant in a covered city. The sitemap carries a per-restaurant lastmod reflecting the actual data update — not the build timestamp — so search engines can identify what changed without re-crawling unaffected pages.

Known limitations

A few honest limitations of this approach are worth stating up front:

Open data

The current snapshot of all 1009 restaurants in this index — slug, English name, city, cuisine, Michelin status, overall score, six dimensions, declared price band — is available as a flat CSV: restaurants.csv (schema, license, and a pandas example on the data landing page). The file is regenerated on every build.

The dataset is released under CC BY-NC 4.0 — free for non-commercial use with attribution to Japan Fine Dining Index. For commercial use, bulk or programmatic access, please get in touch.