The knowledge cut-off is the date up to which a language model’s training data extends. Anything published after this date is not part of the model knowledge. A model with a cut-off in autumn 2025 knows, from its own knowledge, no study, no product and no price from 2026.
How does the knowledge cut-off work?
A model is trained on a fixed dataset. After training, its knowledge no longer changes until the provider trains a new version. Cut-off dates typically lie several months before a model’s release; as of August 2026, only three of the widely used models had training data from 2026. The systems close the gap between cut-off and present via a web search, whose results flow into the answer through grounding.
Why does the knowledge cut-off matter for AI visibility?
It explains why grounding carries so much weight. A model that is not confident about a question — because the topic is new or changes quickly — triggers a web search. At exactly this moment it counts whether your page ranks for the query and whether its passages are extractable. For questions about prices, dates, product launches and current figures, web search is the rule, not the exception.
What does this mean for your website?
Everything you published after the cut-off of the common models reaches the AI only via web search. So make the currency of your content visible: publication and modification dates in the source code and in the schema markup, current years in the text. The term freshness describes why AI systems cite older content much less often.