Direct answer
An AI answers from two very different sources. The first is its training memory: what it "learned" once and for all, frozen, without going online. The second is web search: what it fetches live at the moment of answering. These two sources don't give the same answers — and above all, a brand can be present in one and totally absent from the other. A very well-known, long-established brand, abundantly cited online (often in English), has a good chance of being "etched" into the model's memory. A more recent, more local brand, or one in a sector poorly represented in the training data, may exist only if the AI goes and fetches it on the web — and vanish the moment it doesn't. Knowing which of the two sources you're present in means knowing whether the AI thinks of you spontaneously or only when helped to search.
The problem
When you ask "does ChatGPT talk about my brand?", you imagine a single answer. In reality, the answer depends on a setting you don't see: did the AI search the web, or did it answer off the top of its head?
Both modes exist in all real-world use. Sometimes the AI answers from memory, instantly, without citing any source. Sometimes it consults web pages and relies on them. Depending on the mode, your brand may or may not appear — for the same question. Ignoring this distinction means measuring an "average" visibility that matches no real situation, and missing the essential point: does the AI know you, or does it have to rediscover you each time?
The idea to grasp
Here's the distinction that changes everything: memory and web search are two separate sources, and a brand's visibility isn't the same in the two.
- Training memory (the "model knowledge"). An AI model is trained at a given moment on an enormous quantity of texts. What it retains is frozen: that's its memory. When it answers without going to the web, it draws solely from there. A brand present in this memory is cited "spontaneously," effortlessly — the AI thinks of it on its own.
- Web search (the "live web"). When the AI goes online at the moment of answering, it relies on pages found right then. A brand can then appear even if the model didn't know it from memory — simply because a page mentions it. And conversely, it can vanish the moment the AI doesn't run that search.
That leaves the real question: why is one brand in memory, and not another? Several factors come into play, and none depends on some magic optimization:
- The volume of existing content. The more a brand has been written about, cited, and discussed online before training, the better its chances of being "learned."
- Fame and seniority. A long-established, globally known brand leaves a stronger imprint than a recent or niche player.
- Language. The large models are trained on massively English-language data. At comparable fame, a brand heavily present in English content starts with an advantage over a brand present mainly in another language. (Do AIs favor American brands?)
In our tests (internal measurements, on a few brands and two AIs), the contrast is stark: a French professional neobank never appeared from memory on business-creation questions — it emerged only a little when the AI went to search the web. Conversely, a very well-known American productivity app came up systematically from memory, with no search at all, the moment the question touched its domain. Same AI, same method: two opposite behaviors, explained not by the quality of the brands, but by their imprint in the training data.
An honest caveat: these are observations on a small number of cases, not a statistical law. They illustrate the mechanism; they don't claim that "every French brand is disadvantaged." The fact that we measure is the observed gap — not a general rule.
What you hear everywhere
"If ChatGPT cites me, the AI knows me." Not necessarily. If the answer comes from a web search, the AI doesn't "know" you from memory — it found you on a page at that moment. Cut off the search, and you can vanish. It's a presence borrowed from the web, not a presence in memory.
"I just need to produce content and the AI will eventually know me." Content published today doesn't enter the memory of an already-trained model. It can make you visible via web search (the live web), but the memory (the model knowledge) only updates at the model's next training, over which no outsider has any control.
"My AI visibility is a single number." No: it's at least two. Present from memory and present via the web are two distinct situations, with neither the same causes nor the same levers. Merging them into a single score erases the most useful information.
Our stance: only the facts. And the fact is that there are two sources, that they often diverge, and that the only way to know is to measure them separately.
Our approach: measure model knowledge and live web separately
From here on, the register changes: we describe the instrument.
Since memory and web search are two distinct sources, a serious measurement never mixes them. For each AI, mAIr measures your presence in two modes:
- Without web search (the model knowledge) — what the AI says about you from memory, spontaneously.
- With web search (the live web) — what it says when it goes to fetch live.
The gap between the two is precisely the information: it tells you whether your visibility rests on fame anchored in the model, or only on web pages the AI consults at the moment of answering. A strong presence on the live web but none in memory tells a very different story from a solid presence in both.