How AI answers are made
Knowledge cutoff
Also called: Training data cutoff
A knowledge cutoff is the date after which a model’s training data stops, so the model knows nothing newer unless it looks it up.
In plain words
A model is trained on data up to a certain date, and then released. Everything after that date is missing from its memory.
AI companies publish these dates. Anthropic, for example, lists a "training data cutoff" and a "reliable knowledge cutoff" for each Claude model.
Example
Your shop moved to a new address after a model’s cutoff date. The model’s built-in knowledge still has the old address. Only a live lookup, grounding, can find the new one.
Why it matters for AI search
For anything recent, AI answers depend on retrieval rather than memory.
Google describes grounding as giving the model content from the Google Search index at prompt time, to improve factuality and relevancy.
What you can do
- Publish changes, such as new prices, products or addresses, on your own site, where a grounded system can find them.
- Show a clear "last updated" date on pages that change.
- Remove old, wrong information from your site, so a lookup doesn’t find it.
Often confused with
- Context window
- The context window is how much a model can read at once. The knowledge cutoff is how recent its built-in knowledge is.
Sources
Facts and sources last checked on 19 October 2026.
Related terms
How does your page do in AI search?
GEO-Score checks one page on 22 signals, from AI crawler access to structured data, and shows what to fix first.
Analyze a page