We asked Gemini 2.5 Flash, with web search switched off, what it knew about a four-month-old website. The answer was one word: “UNKNOWN.” Not a hallucination, not a guess. The model simply had no memory of a brand that did not exist when its training data was collected.
Whether an AI model “knows” your brand depends on two separate layers. The first is training memory: knowledge baked into the model’s weights during training, frozen at a cutoff date and refreshed only when the lab trains a new version. The second is live retrieval: web search the model runs at answer time. New and mid-size brands live almost entirely in the second layer, because the first has never heard of them.
Confusing these two layers is the most common mistake in AI visibility work. They fail differently, they refresh on different clocks, and they need different tactics.
What is the difference between memory and retrieval?
Training memory is parametric. During training the model reads a vast corpus, and brands that appear often and consistently across credible sources get encoded into the weights (Neo4j). That knowledge is fixed at the training cutoff. Anything published after it is invisible to the base model unless retrieval fills the gap (knowledge cutoff explainer).
Retrieval is the live briefing. When the model searches the web mid-answer, it pulls current pages and works them into the response. Think of it as a packet handed to the model right before it speaks. If that packet is thin or missing your best pages, the model falls back on memory, and if memory is empty, you get “UNKNOWN.”
The clocks are what make this matter. Retrieval refreshes constantly, the moment your page is indexed. Training memory refreshes on the lab’s schedule, which for brand knowledge can be a year or more between the version that learned about you and the one before it. WordLift’s researcher argues models often skip live lookup for brand queries because retrieval costs money, meaning some of what a model “knows” about you is a slow-moving snapshot that may only update every 6 to 12 months.
Why do young brands only exist in one layer?
Because they were not in the training corpus. A brand launched this year did not appear in a dataset assembled last year. The base model has no parametric memory of it to recall. That is not a defect you can fix with better content. It is arithmetic. Our Gemini test with search disabled is the clean demonstration: no retrieval, no memory, no answer.
Switch retrieval back on and the same model can describe that four-month-old brand fluently, because now it is reading live pages instead of recalling frozen ones. For a young brand, retrieval is not one channel among several. It is the entire surface. Every ounce of AI visibility a new company has depends on being crawlable, indexed, and citable right now, because the memory layer will not carry any weight for another training cycle.
This reframes the anxiety productively. A founder worried that “ChatGPT doesn’t know us” is usually worried about the wrong layer. The base model not knowing you is expected and temporary. The urgent question is whether retrieval can find you, because that is the layer you can actually move this quarter.
Does that mean training memory doesn’t matter?
It matters, and it favors incumbents brutally. A brand that has been mentioned consistently across authoritative sources for a decade is baked into the weights. The model recalls it without searching, which is faster and, by several accounts, treated as more reliable than live-retrieved facts. GEICO and Coca-Cola do not need to worry about their retrieval packet. They are the memory.
For everyone else the strategy splits by clock speed. Retrieval is your near-term lever: publish, get indexed, earn citations, and the model can use you tomorrow. The mechanics of that live surface are in how ChatGPT chooses which sources to cite. Training memory is your long game: sustained, consistent presence across credible sources so that the next model version encodes you and stops needing to look you up. There is no shortcut into the weights. You earn your way in by being written about, repeatedly, over a training cycle you do not control.
The mistake to avoid is spending on the slow layer while ignoring the fast one. A new brand pouring budget into “being remembered by the model” is optimizing a layer that will not refresh for months, while its retrieval surface, the thing that decides today’s answers, sits un-indexed and un-cited.
Start by finding out which layer is actually answering questions about you. Run a check with retrieval on, then reason about what the base model would say with it off. Our free scan runs live questions through a grounded engine and shows the citations, so you can see whether you exist in the layer that refreshes daily or the one that refreshes yearly: check it.