Training data cutoff
The training data cutoff is the date an AI model's training data ends. Anything about your brand after it, the model only knows by searching.
The training data cutoff is the date where a model's training data ends. Anything that happened after it, the model didn't read about in training. Your launch, your new prices, your new name: if they came later, the model only knows about them if the platform searches the web and finds them.
AI companies usually publish the cutoff for each model. Anthropic, for example, lists a training data cutoff and a reliable knowledge cutoff for each Claude model on its models overview.
Why it's earlier than you'd think
Two things make the gap wider than the date suggests.
Models are released some months after their cutoff, and people go on using them long after that. So the model answering a question today may have stopped reading the web a year or more ago.
And the last months before the cutoff are thinly covered, because the web hadn't caught up yet. That's why some companies give a second, earlier date for when the model's knowledge is most reliable, and why the idea of an effective cutoff exists.
An example
For example, Harrow & Pine renamed its best-selling table range in May. A model whose cutoff is the previous December has never heard the new name. Asked about the new range by name, it may say it doesn't know it, or guess. Asked about Harrow & Pine's tables, it describes the old range under the old name.
What you can do
- Put anything new on a clear page of its own, in words people search for
- Say what changed on the old pages too, for example "now called..."
- Get the change into the places the platforms read: directories, review sites, the press
When the platform searches, these pages fill the gap. When the next model is trained, they're part of what it learns.
In AnswerOrbit
Each answer in Answers shows the model the platform used, along with the pages it read. If an answer is out of date and used no pages, it's likely answering from what it learned before its cutoff.