AI search · The panel's scope

Why we don't measure Perplexity, Mistral or Grok

Our weekly reading queries ChatGPT, Claude and Gemini. Together they account for roughly 86% of the traffic that AI platforms send to websites in France, according to SE Ranking data for January to April 2026, from a study covering January 2025 to April 2026. We prefer three engines measured in depth every week, with dates, named models and preserved evidence, over a long list measured once for a brochure. This page explains that choice with dated figures, and states precisely what would bring a new engine into the panel.

How an engine earns its place in the panel

A measurement panel is not a row of logos: it is a commitment. Every engine added multiplies the passes to repeat, the answers to archive and the cadence promises to keep, week after week, under our honest measurement charter. So we add an engine for one reason only: because French buyers actually use it to choose their suppliers. Not because a competitor lists it on a brochure.

Public figures say who carries weight today. Of the traffic AI platforms send to French websites, shares for January to April 2026: ChatGPT 77.2%, Perplexity 8.82%, Gemini 6.30%, Copilot 3.72%, Claude 3.01%. Mistral: 0.89% over the whole period, January 2025 to April 2026. Grok produces no measurable traffic there over time. Source: SE Ranking, aggregated data from more than 100,000 websites connected to Google Analytics.

One surface joins this three-assistant panel, on a different cadence: since 26 August 2026 we read Google's AI Overviews every quarter, the generated answers at the top of the results page. That reading is published separately, on the Barometer page, and never enters the weekly count.

Perplexity: the first candidate, and why it still waits

Perplexity is the second-largest source of AI traffic in France in that dataset. That is not nothing, and we will not pretend otherwise. Two things nevertheless keep it out of the weekly panel.

First, a share of traffic is not a share of usage. Perplexity shows sources under every answer, by design; mechanically, its share of outbound traffic to websites exceeds its real share of conversations. Second, our criterion is the buyer's: in our conversations with the leaders of French SMEs, the question "what does Perplexity say about us?" never comes up. The day it does, Perplexity enters the panel, and this page will say so. It is today the first candidate, and we re-examine this scope at every quarterly reading, against dated third-party data.

Mistral: the French champion, and an honest number

Mistral is the French champion, and we would be glad to see it carry weight. The data says otherwise today: 0.89% of AI traffic in France over the measured period. Many French companies do encounter Mistral's models without knowing it, embedded in business software; but a citation reading measures the engines a buyer actually asks questions to, and that usage remains marginal. Same rule as for everyone: if usage changes, the panel changes, and the change will be dated and explained here.

Copilot and Grok, for completeness

Copilot stands at 3.72% in the same dataset, ahead of Claude. We do not measure it, and that is not an oversight: Copilot is a Microsoft surface, but the engine answering inside it is OpenAI's, already in the panel through ChatGPT. Measuring the wrapper on top of the engine would say nothing new about what AI knows of your company. Grok currently produces no measurable traffic to French websites: there is nothing to measure honestly.

One broader reality deserves saying out loud: most of the AI systems French companies use daily, embedded assistants included, run on American models. Measuring ChatGPT, Claude and Gemini means measuring the engines that actually write the answers your buyers read.

Wide panels, measured once

Some competitors advertise wider panels: six engines, one hundred and fifty questions. Then look at what we look at in the market register: is the measurement dated, is the model named, were the questions frozen before the test, were the passes repeated, are the raw answers preserved? A wide panel measured once makes a brochure. A narrow panel measured every week, with its zeros published on the barometer, makes a measurement. We chose the measurement.

What would change this page

The panel's scope is re-examined every quarter, against dated third-party data, alongside our quarterly reading. Three precise triggers would bring an engine in: the leaders we meet start asking the question, a serious third-party source durably shows it carrying real weight in French usage, or our own readings show it shaping the answers buyers read. Every change to the panel will be announced as an instrument change, dated, with the old and the new rule side by side.