AI knowledge cutoff dates: ChatGPT, Claude, Gemini and Grok

A client asks why ChatGPT still describes their company the way it looked before the rebrand. And why the model names a competitor but doesn’t seem to know their product. Before we answer that the model “has old data”, let’s look at how this actually works.

What is an AI model’s knowledge cutoff?

AI model knowledge cutoff dates: the table

Clicking a header sorts the whole table and temporarily dissolves the groups.

How to read cutoff dates: three things that mislead

A declared cutoff is a soft boundary. Data in the training corpus thins out as collection approaches its end, so the final months before the cutoff are represented more sparsely than earlier ones. Anthropic describes this openly with two separate dates. Other providers give one, which does not mean the effect is absent there – they simply don't separate it out.

• A cutoff is not a release date

  A model released in 2026 may have its memory stopped in January 2025. For some models the gap between release and cutoff runs to well over a year, and it has no fixed value, so the version number says nothing about how fresh the data is.

• A cutoff is not the date in the model name

  Identifiers such as claude-haiku-4-5-20251001 or Bielik-11B-v3-Base-20250730 denote a version or a training checkpoint. That is not the boundary of the data.

• A model's answer about itself is not a source

  Hundreds of articles about earlier generations' cutoffs sit in the training data, so a model can reproduce someone else's date as its own. What counts is what the provider states in the documentation for a specific model identifier.

Effective cutoff: knowledge does not stop evenly across topics

A declared cutoff does not apply equally to every source and every topic in the training corpus. A team from Johns Hopkins University showed this in Dated Data: Tracing Knowledge Cutoffs in Large Language Models (Cheng et al., arXiv:2403.12958), an Outstanding Paper at the COLM 2024 conference.

The authors introduced the notion of an effective cutoff – an empirically estimated boundary of a model's knowledge, calculated separately for individual sources and topics, as distinct from the single date declared by the provider. They proposed a way to estimate it by probing the model across successive versions of the same data, without access to the training corpus. The finding: effective boundaries often differ markedly from the declared date.

Their analysis of open corpora pointed to two causes. The first is temporal misalignment in Common Crawl data – new dumps contain significant amounts of older content. The second lies in deduplication schemes, which struggle with semantic duplicates and near-duplicates at the lexical level.

Why ChatGPT doesn't know your new page

An answer is built from the knowledge held in the model's parameters and – depending on the product and its settings – from content added along the way: search results, an uploaded file, product memory, a connector or the conversation itself. The cutoff governs the first layer only. If your page appeared after the declared cutoff, don't assume its content sits in the model's parametric knowledge.

The catch is that retrieval fires conditionally. On general questions, category comparisons and recommendations the model often answers from memory, because it sees no reason to search. It then names what it knows – the competitors from before the cutoff. And even when it does search, what guides it in judging the sources it finds was shaped by data from before the cutoff. That is why being in a model's memory and being in an index are two separate things.

Three speeds of propagation

Generative systems learn about changes on your site at different rates, and that rate depends on architecture, not on the quality of the content.

This difference in speed has a consequence that is easy to forget during an audit: a base model with no web access and no additional context will not see your page unless its content was fixed during training. This applies to open-weight models run locally – Llama, Qwen, Bielik and Mistral. SEO has no direct influence on the base model, because the model itself does not visit the web. The two routes are the presence of your content in the training corpus, or delivering it through a RAG system, a search tool or another component deployed around the model. Which bots reach for your content at all is something you can check on the list of crawlers and AI bots.

How to check whether a model knows your brand

Asking a model for its cutoff date won't give you a reliable answer, but asking about specific facts will. Instead of asking "what is your knowledge cutoff", find out where the boundary of its knowledge about you actually runs.

1. Turn off search in the interface, if it can be turned off. Without that you are testing grounding, not the model's memory.
2. Ask about several well-documented facts from before the declared cutoff. A missing or wrong answer suggests that knowledge of the brand is thin or hard to surface – it does not prove the brand was absent from the corpus.
3. Ask about a change made after the cutoff and check where the model got it. A correct answer may come from search, a file, a connector, product memory or an earlier part of the conversation rather than from parametric knowledge.
4. Repeat the test on several models and in several fresh conversations. Divergences show differences in knowledge and in how information is retrieved, but a single question will not pin down the boundary.
5. Check several topics, not only your own brand. Effective boundaries differ between fields, so a result from one area says nothing about the others.
6. Record the conditions of the test: the full model identifier, the product or API, the state of search, the state of memory and connectors, the exact prompt and the date. Without that the result cannot be compared with the next measurement.

A single test says little, because models answer non-deterministically. A repeatable measurement over time says more. How I approach this with clients is described on the page about AI visibility.

Frequently asked questions