Direct answer
mAIr (Metrics of AI Responses) is a discipline of measuring what artificial intelligences say about a brand and a market. Unlike most AI visibility tools, mAIr does not sell optimization and never "judges" using an AI: it repeats every measurement, computes it mechanically, displays it with a margin of uncertainty, then timestamps and cryptographically seals it. Where many tools deliver a score on the day you happen to look, mAIr delivers a reproducible measurement you can stand behind.
The problem
Type "AI visibility tool" and you'll land on about thirty platforms. They all promise you the same thing: a nice score out of 100, a dashboard, competitors, recommended actions. They all tell you that you're more or less visible in ChatGPT.
One question, just one: how many times was that score actually measured?
In most cases, the answer is: once. A single query, on a given day, turned into a round number. Except an AI doesn't answer the same way twice. Ask it the same question again tomorrow, or even this afternoon, and the score moves. What you're sold as a measurement is really a photo taken in the dark: you see something, but you don't know what.
And it gets worse. Most of these tools are both the ones measuring your visibility and the ones selling you the services to improve it. The thermometer and the invoice in the same hand.
The idea to grasp
To understand what sets mAIr apart, you have to separate three things the market lumps together.
First, measuring is not optimizing. Optimizing your visibility in AIs — GEO — means taking action: working on your content and your presence so you get cited more often. Measuring means observing what AIs say, without changing anything. Two different jobs. mAIr does the second one, and only that.
Second, a single measurement is not a measurement. Because the models are probabilistic, an isolated answer tells you nothing reliable. You have to repeat — query the same thing several times — and look at how the results scatter. That's what turns an impression into data. mAIr measures at n=20 and observes the spread of the results.
Third, a score computed by an AI is not a neutral measurement. If you ask an AI to rate a brand's visibility, you're stacking one layer of approximation on top of another. In mAIr, the AI never computes anything. It produces the answers that get measured — and that's all. All of the scoring (presence, rank, share of voice, stability) is computed mechanically, in code, in a reproducible way.
These three principles aren't technical details. They're the conditions for a measurement to deserve the name.
What you hear everywhere
"Our AI visibility score is computed over millions of prompts." Impressive. But your brand, on your queries, how many times was it measured? The overall volume of the database doesn't replace repetition on your specific case.
"We measure AND we help you improve." That's exactly the problem. When the party doing the measuring cashes in on the optimization, the diagnosis is no longer independent. You don't ask the auditor to keep the books they certify.
"You get a real-time dashboard." Real time without repetition is blurry real time. A figure that changes with every refresh, with no displayed margin of uncertainty, doesn't help you decide — it creates the illusion of precision.
My fundamental stance, unchanged from the start: no trust in the AI, none in the vendor, only in the facts. A score is only a fact if it's repeated, computed mechanically, dated, and you can produce it as-is in the event of a dispute.
My vision: measurement as an instrument, not a sales pitch
From here on, the register changes: we describe the instrument. Plainly.
A credible AI visibility measurement rests on four requirements:
- Repetition: each query is asked several times (n=20 in production), and the spread of the answers is measured, not endured.
- Mechanical computation: presence, rank, share of voice, stability, and cited competitors are computed in code. The AI rates nothing.
- Dating and sealing: each report is timestamped and signed (HMAC-SHA256), so it can be verified and stood behind later.
- Model knowledge vs. live web: mAIr separates what the AI knows on its own (its own memory) from what it goes and fetches on the web at a given moment — a distinction that scrapers cannot reproduce.