Methodology

The Index publishes measurements, never judgements. Here is how the numbers are made, and where they stop.

What we measure

Every week we ask four AI assistants (ChatGPT, Claude, Gemini and Perplexity) a fixed basket of 81 real diner questions about Edinburgh. We record every restaurant named in every answer, and its position. A restaurant's share of voice is the percentage of answers that name it, worked out for each assistant separately and then averaged across the four. That's the whole method: same questions, same models, same way, every week.

How a run works

• 81 prompts spanning cuisine, neighbourhood, occasion, budget and visitor questions, phrased the way people actually type them.
• 4 models × 81 questions: up to 324 answers a week, each stored in full. The latest run returned 324. When a model rate-limits or fails on a question we record what we actually got rather than pad the total, so the figure moves week to week and every published percentage uses that week's real denominator.
• Each assistant gets an equal vote. We average the four per-model percentages rather than pooling every answer into one pile. Otherwise a week where one model answered fewer questions would quietly reweight the whole table towards the models that did answer, and two weeks could not be compared. This way a restaurant's number moves only when the models' answers move.
• Names are extracted from every answer and matched to a canonical list of Edinburgh restaurants.
• We keep every raw answer, so week-over-week movement is auditable.

What “share of voice” means (and doesn't)

Share of voice is the share of relevant answers that name a restaurant. It is not a rank. Any single AI chat can differ from the next, so we never report a one-off position. We report the pattern across hundreds of answers, and how it moves over time.

The caveats

• API answers are directional. We query each model's public API with no personalisation. The consumer apps add account history and live web search, so an individual chat you run may differ.
• Answers vary run to run. That's why the Index is a weekly average across many prompts, and why movement matters more than any single number.
• Perplexity is search-grounded natively; the other three run un-grounded in this version. A “live web” column is on the roadmap.

The restaurant list

Our canonical list is built from Google Places (Edinburgh venues with 25+ reviews), then cleaned by hand each week (chains and non-restaurants removed, closures noted). Currently tracking 877 restaurants. When an AI recommends a place that has closed, we leave it off the list and note it: that is a finding, not an error.

No pay-to-appear

No restaurant can pay to appear in the Index, or to move within it, including our own audit clients. The measurement and the paid work are kept strictly separate.

Corrections

Spotted something wrong? A closed venue, a mismatched name, a restaurant we have missed. Email [email protected] and we'll fix it in the next run. The full dataset and methodology are available to journalists on request.