Definition
A knowledge cutoff is the date after which a model’s training data stops, so anything that happened later is absent from what the model learned. Unlike a context window, which limits how much text a model can hold in one conversation, a cutoff limits what it learned before the conversation started. Retrieval is how a model gets around it.
At a glance
- CategoriesCore Concept
- Related fieldsAI Search
- DifficultyBeginner
At a glance
- The cutoff is the last date in the training data, not the model’s release date.
- A model with a 2025 cutoff can still answer about last week if a live search hands it the text.
- Models often state their own cutoff wrongly, because that date is itself a learned fact.
- Every model inside a product has a different cutoff.
Where the cutoff date comes from
A model is trained once, on a fixed pile of text. The pile is assembled, frozen, then training runs for weeks. The cutoff is the last date in that pile, and the gap between it and the launch is usually months. A model released in September may have learned nothing after the previous December.
Nothing gets added afterwards. Training does not tick along in the background, and the model has no sense that time has passed.
What changes is the wrapper around it. Products like ChatGPT and Perplexity search first and paste fresh pages into the prompt, which is how AI search works at all. The cutoff has not moved. The model is reading something current that retrieval handed it, held in the context window for one answer and gone afterwards.
The version of your brand a model remembers
The worry is always about being left out of an AI answer. The quieter problem is being included with facts from two years ago.
Say you rebranded in March and dropped two product lines. A model trained before that still holds the earlier version. Ask about you in a way that does not trigger a search and the earlier version is the answer: old pricing, a departed founder, a discontinued product, stated without a hedge.
Stale does more damage than absent, because it reads as authoritative. Nothing in a trained memory carries a date stamp.
Two things pull against it. Repetition, because a corrected fact said across many sites gets picked up by the next training run. And retrieval, because a live fetch of your page usually outweighs what the model half-remembers. A reasoning model sometimes notices the conflict and searches to settle it, but that is luck, not a plan.
Keep your current facts reachable, not just published
Every fact about you exists in two forms: what a model learned, and what it can fetch now. You control the second one.
- Date the facts that age. Pricing, headcount, availability, leadership. A dated line gives retrieval a reason to prefer your page over an undated copy.
- Keep one canonical page per fact. One pricing page, one about page, one per product. Three half-answers means the fetch returns whichever it found first.
- Retire old pages properly. Redirect the discontinued product page and the pre-rebrand about page. Unlinked is still crawlable, and still quotable.
- State the correction in words. “Formerly known as Acme Group” does real work, tying the thing a model remembers to the thing that is true now.
- Ask, every quarter. Put a plain question about your company to ChatGPT and Gemini and read the answer for stale facts. Ten minutes, no budget.
Two ways a model can know something about you
| Learned in training | Fetched at answer time | |
|---|---|---|
| Freshness | Frozen at the cutoff | As current as your live page |
| How you influence it | Be widely written about, over time | Be crawlable and clear today |
| Speed of change | Next training run, months out | Next crawl, days or weeks |
| Failure mode | Confident and out of date | Wrong page pulled, or none |
Frequently asked questions
-
Is a knowledge cutoff the same as a context window?
No, and the two get mixed up constantly. A cutoff is about training: the last date in the data the model learned from. A context window is about capacity, the text a model holds in one conversation, which Anthropic’s documentation on context windows sets out per model. One is fixed at build time, the other resets every chat.
-
How do I find out a model's cutoff date?
Check the provider’s documentation rather than asking the model. A model’s answer about its own cutoff is a learned fact like any other, and often wrong by months. Providers publish the date per version, and a product like ChatGPT routes you between versions depending on the request.
-
Can I get a model to update what it knows about my company?
Not directly, and nobody can sell you that. You make the current facts easy to find and widely repeated, so the next training run learns them and today’s searches surface them. No submission form exists, and no vendor has a back door to one.
-
Does a knowledge cutoff still matter now that assistants search the web?
Less than it did, and not zero. Search covers the questions where the assistant chooses to search. Short factual questions about a company the model thinks it already knows often skip that step and answer straight from training.