A large language model is trained on text up to a certain point. That point is its knowledge cutoff. Providers state the date for each model. In its models overview, Anthropic distinguishes two dates: the training data cutoff, the range of the training data used, and the reliable knowledge cutoff, the date up to which the model’s knowledge is most extensive and reliable.
Why does it matter?
Anything that changes after that date, the model does not know from its training. A new service, a move or a different price only reaches an AI answer if the assistant searches live and finds your page. Anthropic writes about its web search tool that it gives Claude access to current information from beyond its knowledge cutoff.
Example
Suppose a physiotherapy practice opens a second location after a model’s knowledge cutoff. Ask the assistant about the practice without web search and it only mentions the first location. With web search it can find the new location, provided the website states it clearly.
Common mistake
Thinking that an update to your website goes straight into the model’s memory. What a model knows from training only changes with a new model version. Until then, what counts most is that your current pages can be found and read by search bots.
More in How do AI assistants decide which websites to mention? and under retrieval-augmented generation.