Methodology · Measured AIs
The AIs we measure, and how.
mAIr measures brand visibility inside AI assistants. This page lists the models we query, how we do it, and why we do not add more. We measure. The client interprets.
Two ways to answer
Stock and flow: two measures, neither better than the other
An AI answers in two ways. Either from its training memory, or by looking information up on the web in real time. These are not two levels of quality: they are two different dimensions of a brand’s visibility.
What the AI has learned
The answer comes from the model’s memory. It is a brand’s established reputation: stable, durable, revealing what the AI retains over the long term.
What the AI looks up
The answer relies on a real-time web search, with verifiable citations. It is the news of the moment: shifting, sensitive to the latest publications.
The scope
The AIs we measure
Three statuses. Tested AIs are measured today. Upcoming ones are being integrated. The others exist on the market but are not measured — the last section explains why. ChatGPT and Copilot are grouped together: same underlying engine.
| Model | Status | Measured modes |
|---|---|---|
| Tested at launch | ||
| ChatGPT · CopilotOpenAI · Microsoft | Tested | Stock + Flow |
| GeminiGoogle | Tested | Stock + Flow |
| ClaudeAnthropic | Tested | Stock + Flow |
| GrokxAI | Tested | Stock + Flow |
| MistralMistral AI | Tested | Stock + Flow |
| DeepSeekDeepSeek | Tested | Stock |
| Coming soon | ||
| PerplexityPerplexity | Soon | Stock + Flow |
| QwenAlibaba | Soon | Stock |
| Available on the market, not measured | ||
| DoubaoByteDance | Available | — |
| ERNIEBaidu | Available | — |
| YuanbaoTencent | Available | — |
Our limit, owned
Why not more flow
An AI’s web mode is not a fixed fact. Often, it cannot be measured precisely.
When an AI searches the web, it queries a search engine. But that engine is not the same everywhere: it changes with the user’s region, the version of the app, and the publisher’s choices — which can change overnight. A single AI may rely on one engine in Europe and an entirely different one in Asia.
A web mode that fluctuates cannot be measured precisely or reproducibly.
For several models, notably Asian ones, web access goes through engines specific to their ecosystem, which cannot be reproduced faithfully from the outside. We can then measure their stock — the model’s memory — rigorously and verifiably. But not their flow as the end user experiences it.
Conversely, when an AI backs its web search with verifiable citations — like ChatGPT, Gemini, Claude, Grok or Mistral — the flow becomes reproducible: we know what was consulted, and we can re-measure it. That guarantee is what separates a measurable flow from one we would rather not put a number on.
Our rule is simple: a measurement we cannot guarantee has no place in an mAIr report. We would rather measure less, but measure right.
Usage breakdown
Who uses what, on the consumer side
Share of global web visits among the seven assistants most used by the public. This measure reflects how humans actually use assistants — the very ground on which a brand’s visibility plays out.
Share of global web visits, all devices — Similarweb, May 2026
This breakdown measures web visits and includes neither mobile apps, nor embedded uses (an assistant preinstalled on a phone or inside an office suite), nor markets where other assistants dominate locally. It is the most reproducible public view, not a measure of total usage.