Methodology and limits

These are the same formulas that are wired into the service. If this page ever disagrees with the calculation, something is wrong and we will fix it.

How the answers are collected

Every question is put to an AI model through its official programming interface, in one of two modes. Model memory — the model answers from what it learned during training. Search and answer — the model searches the web and answers with sources. The mode is shown next to every answer. Consumer apps may answer differently: their output depends on the account, the chat history, the region and the version. We do not bypass any service's protections and we do not scrape their interfaces.

How many questions and runs

16–18 questions of five kinds: category, commercial, comparison, reputation, product. The base set is 16; two more are added when you name competitors — a comparison with the first of them and a comparison with the whole list. Commercial questions are asked 3 times: AI answers are not deterministic, and a single run spreads too widely to mean anything. The rest are asked once.

Formulas

Mention rate = answers containing the brand ÷ all runs × 100. The interval is Wilson's, 95%.

Share of voice = mentions of your company ÷ (mentions of your company + mentions of the competitors you track) × 100. Counted only on non-branded questions; the denominator is the same for every participant.

Share-of-voice threshold. We count share of voice only on questions where your company name does not appear in the wording of the question itself. If fewer than three such questions remain — which happens when the brand name coincides with a word from the industry description — we do not show share of voice at all and say why in the report. An inflated number is worse than a missing one.

The 0–100 index is made of five parts:

• Presence — 40
• Share of voice — 20
• Role in the answer — 15
• Sentiment — 15
• Model coverage — 10

Estimated missed demand is an order of magnitude, not a revenue forecast: the search volume of commercial queries in your industry × 0.43 (the share of commercial queries on Yandex that receive an AI answer, per Yandex data for Q2 2026) × (1 − your mention rate).

Limits

Language-model answers are probabilistic and may be inaccurate, incomplete, outdated or invented; we do not verify them. The report is not a ranking, an assessment of the quality of goods or services, a recommendation, or advertising. Sentiment is assigned by a classifier model and carries an error — a confidence value is printed next to every assessment.

Common questions

Why is the index not simply the percentage of mentions?

What matters to a director is not only whether the brand was mentioned, but in what role, with what sentiment, and across how many models it is visible at all. Mention rate weighs the most, but it is not the only part.

Can the index be compared across industries?

Carefully. The share of commercial queries that get an AI answer differs by category, and the competitive base is different too. The safer comparison is your own brand over time, and against direct competitors.

Where do I read about GEO versus ordinary SEO?

The terms and the difference are explained on “How GEO differs from SEO” — that page is in Russian.