Direct answer
An AI visibility measurement is never final. Models change constantly: their providers apply daily adjustments and regularly switch versions. The result: what the AI says about your brand today — your presence, your rank, the competitors it cites — may be different in a month, and upended when a new model arrives. AI visibility measured once is a dated snapshot, not a settled fact. For it to remain useful information, you have to repeat the measurement over time.
The problem
You have your AI visibility measured. The result is good: you're cited, well placed, ahead of your competitors. You file the report away, satisfied. You consider the matter settled.
Mistake. You've just taken a snapshot of a landscape that keeps moving.
Because while you're filing the report away, the model keeps changing. Not in six months. Every day. And the day a customer asks the AI whether your brand is a good option, the answer may no longer be the one you measured — without you having changed anything on your end.
The idea to grasp
Picture a large dam holding back a huge reservoir of water. The reservoir is the AI's raw power. The dam is the set of rules the provider puts in place to keep it usable, law-compliant, acceptable. And every day, technicians come to patch leaks, adjust what the dam lets through, tweak settings.
You don't see these adjustments. You're not notified. But you feel their effects: what the AI used to let through about your brand last week may be filtered, reworded, or replaced this week.
Why the level shifts quietly, concretely:
- Daily adjustments. Providers continuously correct their models' behavior — new rules, new filters, adaptations to a shifting legal context. Most go unannounced.
- Version changes. When a new model is released, its memory and its associations change. A brand well established in the old model can be less present in the new one — or the reverse.
- The drift is invisible to the naked eye. It doesn't show up in a single answer. It manifests statistically, over time, across a large number of measurements.
- It's localized. It often appears first in one language, with one provider, on one type of question — before spreading.
And here's the most common reasoning error: believing an AI model is stable like an installed piece of software. People assume that something that worked yesterday still works today as long as no one touched it. That intuition doesn't apply to AIs. A brand resting on its laurels is like a house built on ground that alternates between flood and drought: the walls eventually crack, even if no one did anything.
What you hear everywhere
"We measured our AI visibility, we know where we stand." You know where you stood on the day of the measurement. That's already something. But it's not where you stand today, and even less so tomorrow.
"Models improve, so our visibility will improve too." Nothing guarantees that. A new model may well cite you less than the old one. "Better" for the model doesn't mean "better for you."
"If it's always shifting, what's the point of measuring?" It's precisely because it shifts that you have to measure regularly. You don't give up weighing a patient because their weight varies — you weigh them at regular intervals to see the trend.
My core stance, and it makes full sense here: facts only — repeated over time. A fact measured once and never rechecked is no longer a fact, it's a memory. Confidence in AI visibility isn't declared once and for all: it's measured, and it's measured over time.
My vision: measure the drift, don't be subjected to it
From here on, the register changes: we describe the instrument.
Tracking AI visibility over time rests on a few principles:
- A baseline set of measurements, timestamped, that serves as a point of comparison.
- Regular repetition (for example every two to four weeks depending on the stakes), to measure the gap against the baseline.
- Cross-comparison across providers and across languages, since drift often emerges first in one specific channel.
- A quantified measure of the gap, computed mechanically (the AI doesn't judge), with a margin of uncertainty.
- A dated sealing of each cycle, to have a traceable record that can be held up and a verifiable history.
The usual word of caution. Measuring the drift doesn't mean mastering it. mAIr doesn't control what providers do to their models, or when. What mAIr masters are the conditions of the measurement and their repetition over time. We observe the movement, we date it, we quantify it — we don't claim to stop it.