Knowledge cutoff
The date after which a model has no training data, so anything later is unknown to it unless it retrieves the information at query time.
In depth
What it really means
Every model has a cutoff. Ask about something after it and an ungrounded model will either say it does not know or, worse, generate something plausible. Your company’s last two product launches may simply not exist to it.
The workaround is retrieval, and this is exactly why grounding matters so much. A model with a 2025 cutoff can still cite a page you published this morning, provided it can fetch it.
How it works
- Training data is collected up to a fixed date.
- The model learns from that corpus and the weights are frozen.
- Anything after the cutoff is absent from those weights.
- Retrieval at query time is the only route for newer information.
Pros & cons
Pros
- It clarifies which visibility problems retrieval can fix and which it cannot.
- Recent content is not disadvantaged, provided the page is retrievable.
- It explains stale brand descriptions without anyone needing to guess.
Cons
- Outdated facts about you persist in the weights until the next training run.
- Ungrounded queries about you return whatever was true at the cutoff.
- You have no control over when a retrain happens or what it ingests.
Common mistakes
- Confusing the cutoff with the context window. One is training recency, the other is per-query capacity.
- Assuming a rebrand or repositioning has reached the model. It has not.
- Publishing corrections only on your own site when the stale claim lives on third-party pages.
- Testing only in grounded modes, which hides what the model believes unaided.
Best practices
- Test your brand questions with web access turned off to see the unaided view.
- Keep a clear, current, crawlable page for every fact that has changed.
- Get updated facts onto third-party sources that get crawled and re-crawled.
- State changes explicitly, such as naming the old product name alongside the new one.
- Expect a lag of many months between a change and models absorbing it.
FAQs
What is a knowledge cutoff?
The date after which a model has no training data. Anything later is unknown to it unless the information is retrieved at query time.
Why does AI describe my company with old information?
Because the current description sits in its training data from before your change. Retrieval can override this, but only when the query triggers a fetch.
Is the knowledge cutoff the same as the context window?
No. The cutoff is how recent the training data is. The context window is how much text the model can read for a single query.
How do I get new information into a model?
You do not, directly. You make the current facts easy to retrieve, and you get them onto third-party sources so the next training run picks them up.
Keep reading
Related on LymLyt
Beyond LymLyt
Further reading
Want this working on your site?
We build the content behind the term, ranked in search and cited by AI.
Book a 30-min call →