How Do LLM Training Cutoff Dates Work, and Can Models Include Newer Data Later?

How do training cutoffs work?

1. Data collection stops at some point: web pages, books, code and other sources gathered up to a date.
2. Training and testing take months.
3. The model is released, often with a stated knowledge cutoff.
4. The model's built-in knowledge stays fixed for that version. Later changes in the world are not in it.

A 2024 study, "Dated Data", found that effective cutoffs often differ from the reported ones and vary by topic. One reason is that web crawls include large amounts of old pages alongside new ones. Another is the way duplicate content is removed. So a model may know less about recent events than its stated cutoff suggests.

Can models include newer data later?

Yes, in three ways:

• Web search at answer time. ChatGPT, Gemini, Claude, Perplexity and Copilot can search the web when a question needs current information. The answer then draws on today's pages, not only on training data.
• New model versions. Developers train and release new models on more recent data. Built-in knowledge updates only when a new version replaces the old one.
• Fine-tuning and updates between major versions can add some newer information, but developers do not usually describe these in detail.

What does not happen is live editing of a model's knowledge about a specific company because its website changed.

Can a new brand appear in AI answers if it was not in the model’s training data?

Yes, through web search. When a question triggers a search, the assistant can find and cite a brand that did not exist when the model was trained. For that to happen:

• the brand's pages must be crawlable by each assistant's search crawler, such as OAI-SearchBot for ChatGPT;
• they need to be findable in the search sources the assistant draws on: Claude uses Brave Search, and ChatGPT uses a mix of providers and OpenAI's own index (studies of how closely its citations match Bing disagree, from 87% within Bing's top 20 in one 2025 study to about 8% overlap with Bing's top 10 in another);
• mentions in other sources, such as reviews, directories, lists and articles, make it much more likely to be found and recommended.

For questions that do not trigger a search, a brand missing from training data will usually be absent, or the assistant may say it does not know. Asking in ways that prompt a search, such as naming the current year or asking for recent options, often changes the result.

How often does ChatGPT update its knowledge of new brands?

Two clocks run at different speeds:

• Search-based answers update as the web does. Once OpenAI's search crawler and search partners have indexed new pages, ChatGPT can cite them, often within days or weeks.
• Built-in knowledge updates only with new model versions. OpenAI has released new models frequently, but each has a cutoff months before release. Built-in knowledge of a new brand typically lags by many months.

Allowing GPTBot, OpenAI's training crawler, affects whether your pages can be used to train future models. Blocking it does not stop ChatGPT's search from citing you, which uses OAI-SearchBot.

What does this mean for businesses?

• For new brands: focus on search visibility and third-party mentions first. They work now.
• For rebrands and changes: expect search-based answers to update first, and model-knowledge answers to lag.
• For outdated facts: fix the sources on the web. Future models will learn from them, and search-based answers will pick them up sooner.
• For measurement: track answers with and without web search separately, where possible, since they change on different timetables.

How to tell whether an assistant uses your content or only indexes it is covered in how to tell whether LLMs use your content or only index it. Why repeated prompts give different answers is covered in why LLM results change when you repeat the same prompt.

What might an illustrative case look like?

Consider an illustrative startup that launched a project management tool six months ago. Asked "what is [product]?" in a way that does not trigger a search, ChatGPT says it has no information. Asked "best project management tools launched in 2026", it searches and cites a launch review and the startup's own site. The startup makes sure OAI-SearchBot and Bingbot can crawl its pages, gets listed on two review platforms and appears in a category roundup. Search-based mentions grow over the next quarter. A year later, a newer model describes the product even without searching, reflecting coverage from before that model's cutoff.

What do people ask about training cutoffs?

How do I find a model's cutoff?

Developers usually state it in model documentation. Treat it as approximate, since effective cutoffs vary by topic.

Can I submit my site to be included in training data?

There is no general submission process. Allowing training crawlers and being widely mentioned increases the chance of inclusion.

Why does ChatGPT sometimes know about recent events and sometimes not?

It depends on whether it searched. Without a search, it relies on training data up to its cutoff.

Does updating my website change what models already know?

No. It changes what search-based answers find now and what future models may learn.