The first time you notice it, you assume it is a glitch. You ask ChatGPT for the best new restaurants in London and it recommends three places, two of which have closed. You ask Gemini for the leading challenger banks in Europe and the answer omits the one whose IPO dominated financial headlines for most of last year. You ask Perplexity for the top project management tools and the list reads like a snapshot from a different year - credible, plausible, and quietly out of date.
It is not a glitch. It is the system working as designed. And once you understand the mechanism, the implications for any brand whose market position has changed in the last twelve months become significant.
The Two Modes Most Users Never See
AI systems answer questions in fundamentally different ways depending on the query, the model, the user's settings, and the system's configuration in that moment. Two modes matter most.
Training data recall. The AI generates an answer from what it learned during its last training cycle - often six to twelve months earlier, sometimes longer. This is the default mode for many queries. The answer is shaped by what the world looked like at the moment the model's training data was finalised, not what the world looks like today.
Live retrieval. The AI searches the web in real time, pulls in current sources, and synthesises an answer from them. This is the mode users assume they are always in. They are not.
Which mode a query triggers depends on multiple factors: the AI system in question, the specific model version, whether browsing or grounding is enabled, how the user has configured their account, and - most consequentially for users - the specific phrasing of the query. ChatGPT's web browsing behaviour differs from Perplexity's grounded retrieval, which differs again from Gemini's tool use, which differs from Claude when given web tools. The differences are technical, but they produce dramatically different answers from what looks like the same conversation.
The user almost never knows which mode they are in. The response looks identical in both cases - confident, well-structured, plausible. The signals that would tell them otherwise are buried in the model's behaviour, not surfaced in the UI.
What Triggers Each Mode
The mechanism is probabilistic rather than deterministic, but patterns are clear enough to be useful. Queries with explicit time markers - "latest," "newest," "this year," "in 2026," "right now" - are more likely to trigger live retrieval. Queries that name specific recent events, products launched in the last few months, or current circumstances will often trigger retrieval as well, because the model recognises the query cannot be answered from training data alone.
Generic categorical queries - "best running shoes," "leading CRM platforms," "top restaurants in London" - are more likely to draw on training data, because the model treats them as questions it already has answers to. The result is an answer weighted toward whatever was prominent when the model was trained, which can mean two-year-old reviews influencing today's recommendations.
The same buyer asking "best restaurants in London" and "best restaurants in London opened in 2025" can get materially different answers from the same AI system, in the same conversation, within seconds of each other. One draws from the training corpus. The other triggers retrieval. The buyer almost certainly does not know that this phrasing choice is the most consequential decision they are making.
The Brands That Lose
This mechanism systematically disadvantages four categories of brand.
Recently launched brands. If a company launched within the past twelve months - sometimes longer - it may simply not exist in the training data the AI is drawing on. Live retrieval can compensate when triggered, but for queries that default to training data, the brand is invisible. A challenger bank that launched last quarter, a restaurant that opened six months ago, a product that hit the market this year - all face structural underweighting until the next training cycle catches up.
Recently repositioned brands. A brand that has materially changed its proposition, expanded into a new category, or rebranded itself may still be described by AI in terms of its previous identity. The training data captures who the brand was, not who it is now. For brands mid-transformation, this can mean AI confidently misrepresents them to buyers conducting research.
Recently improved brands. A product that has resolved historical weaknesses - better quality, better service, expanded feature set - may continue to be characterised by AI based on the reviews and coverage it received before the improvements. The fix exists. The signal that the fix exists does not.
Currently leading brands in fast-moving categories. In categories where competitive dynamics shift quickly - fintech, consumer apps, restaurant openings, AI tooling itself - the gap between current reality and training-data reality is largest. The brand that is dominant today may be entirely absent from AI's default recommendations because dominance is recent.
None of these brands have done anything wrong. They are losing AI visibility for reasons that have nothing to do with content quality, infrastructure investment, or brand strength. They are losing it to a structural property of how AI systems decide which mode to operate in.
The Brands That Win - and Why They Should Not Be Comfortable
The flip side of this dynamic is that brands which were dominant at the moment the AI's training data was finalised will continue to enjoy disproportionate AI visibility long after their actual market position has shifted. A restaurant that earned heavy coverage two years ago will continue to be recommended even as newer establishments overtake it on critical reception. A tool that defined a category in 2023 will continue to appear in AI recommendations even after competitors have closed the gap or surpassed it.
This is comfortable for the brands benefiting from it. It should not be. The training data lag does not last forever. Each new training cycle resets the baseline, and brands relying on historical citation density without continuing to build present-day signals will see their AI visibility erode the moment the next training cycle reflects current reality. The advantage is rented, not owned - and the rent comes due.
What This Means for Brand Strategy
The implications for marketing and brand strategy follow directly from the mechanism. Three are worth surfacing explicitly.
First, the question is not just are we visible to AI. It is are we visible to AI in both training and retrieval modes. A brand that performs well only in live retrieval queries is dependent on the user triggering retrieval - which most users do not consciously do. A brand that performs well in training-data queries but has not invested in current retrievable infrastructure will lose ground in real-time queries. Both modes need to be addressed.
Second, the time horizon of AI visibility work matters more than most marketing planning accounts for. Content published today contributes to training data that may not be reflected in AI responses for six to eighteen months. Infrastructure changes made today affect live retrieval almost immediately but training-data presence on a multi-month delay. Brands that wait for proof of AI visibility impact before investing will always be measurably behind brands that move on conviction.
Third, the brands that benefit from training-data inertia today should treat it as a borrowed advantage rather than a permanent one. The AI that recommends them this year may not recommend them next year, and the difference will be determined by what they are building right now, not by what they built when they earned their original prominence.
The Honest Conclusion
Most users assume AI gives them the answer that is true today. The reality is more nuanced. AI gives an answer that is sometimes true today, sometimes true twelve months ago, sometimes a blend of both, and almost never accompanied by signals that would tell the user which is which.
For brands, this is a structural condition of the market they now operate in. It cannot be fully solved, but it can be addressed - by building both retrievable infrastructure for live queries and citable presence for training-data queries, and by understanding that AI visibility is not a single state to be achieved but two distinct conditions to be managed.
The brands that recognise this earliest will compound the advantage. The brands that treat AI visibility as a single problem with a single solution will keep being surprised by AI recommendations that do not match the world they think they are competing in.