Direct answer
An AI answers differently depending on whether or not it uses the web. Without web search, it draws on its internal memory, what it learned during training: the stock. With web search, it goes and fetches live pages: the flow. These two modes often produce very different answers about the same brand. The challenge isn't just to measure one or the other, but to control which regime you're querying the AI in — because if you don't control that setting, you don't know what you're measuring.
The problem
When you ask an AI "what are the best brands in [your category]," you assume it gives you one answer. In reality, it can give you several, radically different ones, depending on a setting you can't see: did it search the web, or did it answer from memory? Did it mix the two?
This is a huge blind spot. A brand can be everywhere when the AI digs through the live web, and almost absent when it answers from memory. Or the other way around. These two situations carry very different consequences — and most measurements don't even say which mode they were taken in.
The idea to grasp
You have to understand what an AI does when it answers, because everything follows from that.
The stock is its memory. During training, the model absorbed billions of texts and retained lasting associations from them: this brand goes with this category, this attribute. When it answers without the web, it draws on that. This stock is stable (it only shifts when the model is retrained) and deep (it reflects a reputation built over years). If an AI cites you spontaneously, without looking anything up, that's a solid asset.
The flow is what it finds live. When the AI searches the web, it pulls up current pages. This flow is reactive (a good recent article can make you appear) and volatile (it depends on what's online today).
The two are worked on differently: the flow is won with fresh content and SEO; the stock is won over time. A brand strong in the flow but weak in the stock is visible today, fragile tomorrow.
And here's the point almost no one states plainly. When a tool queries an AI to "measure your visibility," does it know which mode the AI answered in? Most of the time, no. It sends a query, gets back an answer, and doesn't control whether the model triggered a web search, for which part of the answer, or whether it blended memory and web. The answer obtained is therefore an uncontrolled mix of regimes.
The direct consequence, and it has to be said honestly: a tool that scrapes AI answers measures neither the stock cleanly nor the flow cleanly. It measures something indeterminate — and it can't say what. It's not that it measures "the flow" as people often believe. It's that it doesn't know what it's measuring, because it doesn't control the setting.
What you hear everywhere
"Your AI visibility is what ChatGPT says about your category." Says it how? From memory, on the web, or a mix of both? Without that detail, "what ChatGPT says" means nothing measurable.
"Our tools measure your presence in AI answers." In which regime? If the tool doesn't specify — and doesn't control — its result can't be qualified. There's nothing to anchor it to.
"To be visible in AI, publish fresh content." Good advice for the flow. No effect on the stock, which only shifts when the model is retrained. Confusing the two means promising something durable with a volatile lever.
My core stance applies fully here: facts only. And the first fact, an uncomfortable one, is that a measurement with no control over the regime isn't a usable fact. You can't draw conclusions from something whose very nature is unknown.
My vision: control the regime, and say where the control ends
From here on, the register changes: we describe the instrument.
What sets a serious measurement apart isn't "capturing the web." It's controlling the querying regime:
- Stock measurement: query the model with web search disabled, to capture its own memory.
- Flow measurement: query with web search enabled, to capture what it finds live.
- Clean separation: never mix the two in a single measurement; each regime is isolated and identified.
- Comparison: the stock/flow gap itself becomes an indicator (durable visibility vs. visibility borrowed from current events).
- Reproducibility: each regime measured through repetition (n=20) with a margin of uncertainty, all of it sealed.
Controlling the regime is exactly what a scraper doesn't do: it's subjected to the mix, whereas mAIr chooses it.
A word of caution, in the name of honesty. Controlling the querying regime (web enabled or not) doesn't mean controlling what the model does inside. An AI remains a black box: you never master 100% of its internal workings. What mAIr masters is the conditions of the measurement — the setting, the repetition, the calculation, the sealing. Not the inside of the model. Measuring rigorously isn't claiming to read the AI's mind; it's querying under known, constant conditions, and saying so.