A knowledge cutoff is the date after which a model's training data ends. Without retrieval, the model's picture of a market is frozen at that date.
A model trained through last year does not know this year's product launches, rebrands, or shutdowns, unless it retrieves. That is why the same model gives different recommendations with browsing on versus off, and why new brands break into grounded rankings long before ungrounded ones.
For challenger brands the cutoff is the moat to cross: the incumbent is in the training data; you have to arrive through the sources the model reads at answer time.
Telling cutoff from retrieval
Ask about something that unambiguously postdates the training data — a recent launch, a rename, a shutdown. An answer that handles it fluently is retrieving; one that misses it, or hedges about its knowledge, is answering from training data alone.
The same model gives different recommendations with retrieval on and off, which is why any measurement has to state which mode it used. This index queries the official APIs with web search enabled, and says so.
What the cutoff means for a newer brand
An incumbent is in the training data; a challenger has to arrive through the sources the model reads at answer time. That asymmetry is why new brands appear in grounded answers long before ungrounded ones, and why retrieval-mode visibility is the leading indicator.
It also decays: as models retrain, today's retrieved coverage becomes tomorrow's training data. Coverage earned now compounds in a way that a one-off mention does not.