Search trained us to think vertically.
You ranked third. Then fifth. A competitor moved above you. The position was imperfect, personalised and changeable, but the mental model was wonderfully simple: visibility lived somewhere on a list.
AI search is making that model harder to defend.
OtterlyAI’s September 2026 analysis tracked 1,600 prompts across seven AI search engines and found strikingly different movements in the social sources those systems cited. LinkedIn citations rose through September across several engines while YouTube fell sharply overall. Gemini barely cited LinkedIn at all. Perplexity behaved differently from ChatGPT, which behaved differently from Google AI Mode.
Those are OtterlyAI’s measurements, not universal laws. The company explicitly limits the findings to its dataset.
But another OtterlyAI study points to the deeper measurement problem. Across 252,407 AI answers, brand mentions were generally more stable than citations. In its datasets, ChatGPT and Google AI Mode reused only about a quarter of the sources they had cited the previous day, while Perplexity retained roughly three quarters. Nearly half of the distinct URLs in one dataset were cited only once during a 30-day window.
If those patterns hold more broadly, asking “Where do we rank in AI?” may be the wrong question.
There may be no single where.
A brand’s visibility is now potentially distributed across engines, prompts, answer formats, source choices and time. The brand can remain present while the evidence underneath that presence changes. Or a source can disappear from citations without the brand itself becoming less visible.
That distinction matters because marketers are going to be tempted to treat every movement as feedback.
We changed the page on Tuesday. Citations rose on Thursday. Therefore the optimisation worked.
Maybe. But if citation turnover is naturally high, the movement may simply be noise.
This is familiar territory for researchers. Before interpreting a change, you need some idea of the normal variance of the thing being measured. A single observation is not a trend, and a metric is not useful merely because a dashboard can count it.
AI visibility may therefore need to be treated less like a rank and more like a distribution.
How often does the brand appear across a meaningful set of prompts? On which engines? In what contexts? How stable is that presence over time? Which sources repeatedly support it? How much movement is normal before we decide something has actually changed?
That is a different measurement architecture.
It also changes how we think about optimisation. Google’s current guidance for its generative AI search experiences is refreshingly unexotic: create unique, valuable, non-commodity content and continue following foundational SEO practices. Google explicitly warns against chasing “AEO/GEO hacks” for its own systems.
So perhaps the most useful response to AI-search volatility is not to optimise harder against every observed fluctuation.
It is to separate what the brand controls from what the intermediary controls.
You control the usefulness and distinctiveness of your content, its technical accessibility, the evidence behind it and the clarity with which it represents what you know.
You do not control which sources an AI system retrieves tomorrow, how it combines them, whether an experiment changes that behavior, or whether another engine uses a different information ecosystem entirely.
That creates a slightly uncomfortable consequence for measurement teams: visibility may become more observable at exactly the same moment it becomes less attributable.
We can collect enormous amounts of data about mentions and citations. The danger is mistaking that abundance for certainty.
The job, then, is not to find the one AI visibility number.
It is to learn the shape of the system: its baseline, its variance, its differences across engines, and the changes large and persistent enough to deserve an explanation.
Search gave us positions. AI may give us probabilities.