Methodology

How Tastibase ranks every dish

Tastibase gives every dish a single 0–10 scorebuilt from thousands of real customer reviews — not paid placements, not star averages. Here's exactly how that score is made.

  1. 01

    Read every review

    We analyse publicly available reviews and pick out every sentence that mentions a specific dish — only review text, never rating-only responses. A single restaurant can contribute thousands of dish mentions.

  2. 02

    Score the sentiment per dish

    A BERT AI model trained on 700,000 restaurant reviews reads each mention and scores how positive it is toward that exact dish — handling negation, sarcasm and food-specific language a keyword counter would miss.

  3. 03

    Blend in the stars

    Each dish’s quality is 85% review sentiment and 15% the restaurant’s star rating — light corroboration, so a great dish at an average venue still rises and a coasting reputation can’t carry a weak dish.

  4. 04

    Shrink toward fair

    Bayesian shrinkage pulls dishes with little evidence toward the average for that dish. One glowing (or fake) review can’t fluke its way to #1 — a dish needs consistent praise across many independent mentions.

  5. 05

    Weight recent reviews

    Only the last 3 years of reviews count, and more recent ones count slightly more. A kitchen that slipped years ago isn’t propped up by old praise; one that turned things around gets credit.

The score, in one line

quality = 0.85 × sentiment + 0.15 × stars
# pull low-evidence dishes toward the dish's average
adjusted = (n × quality + C × prior) / (n + C)
# gentle recency weighting, last 3 years only
dish score = adjusted × recency × 10

Volume is confidence, not points.A dish with two reviews can't outrank one with two hundred just by averaging higher — shrinkage makes the chart reward proven, consistent quality.

What the scores mean

Exceptional8.5 +
Excellent8.0 – 8.4
Very Good7.0 – 7.9
Good6.0 – 6.9
Average / Mixedbelow 6.0

Common questions

Why rank dishes instead of restaurants?▾

A 4-star restaurant can make a brilliant biryani and a forgettable naan — a single star average hides that. Tastibase scores each dish on its own evidence, so you order the thing a kitchen actually nails.

Can a restaurant pay to rank higher?▾

No. Rankings cannot be bought. No restaurant pays to appear, to rank higher, or to be removed. The score is the score.

Do you show the actual reviews?▾

No. We publish aggregate scores only — a numeric score, a sentiment reading and a mention count. We never reproduce the verbatim text of a review or the name of a reviewer.

How many reviews does a dish need to be ranked?▾

A dish needs at least 4 clean mentions across a restaurant to be scored, and a city ranking only lists a dish once 2+ restaurants qualify — otherwise the evidence is too thin to be fair.

For how we source data and protect reviewer privacy, see About & data sourcing.

See the rankings →