How Often Does ChatGPT Update Its Knowledge? Cutoffs vs. Live Search

A model's knowledge never changes after training. It changes only when OpenAI releases a new or refreshed model, and each one answers from a cutoff that sits months before launch. GPT-6 Astra, for example, has an April 30, 2026 knowledge cutoff and reached the API on September 3, 2026.

The release rhythm is fast. OpenAI's API changelog shows GPT-6 Astra on September 3, GPT-6 Sol and Luna on September 22, and GPT-6.1 Sol on September 29, 2026: four models in 26 days. It also logs two refreshes of the chat-latest snapshot, which tracks the Instant model used in ChatGPT, on May 5 and May 28, 2026.

A knowledge cutoff is the last date covered by a model's training data. The two newest releases we checked shipped about four months after their cutoffs, so a fresh model is already four months behind on day one.

Plan around that gap, not around a release calendar. If your pricing, leadership or positioning changed in the last four to eight months, assume the newest model's memory has not seen it.

There is no single ChatGPT cutoff. OpenAI's page for chat-latest, the alias for the latest Instant model used in ChatGPT, lists an August 31, 2025 cutoff, while the flagships on its models page run to April 30 and May 18, 2026. That is an eight-month spread inside one product.

Which model a given buyer gets depends on their plan and settings, and you cannot see that from outside. So the same question can return different "memories" on different accounts.

Treat the model's own claim about its cutoff as a hint, and OpenAI's documentation as the record. Test with several accounts or tiers before you decide a fix has failed or worked.

Only when it searches, and searching is the model's call. OpenAI's web search guide says the model can choose to search or not based on the prompt. When it does, the answer draws on pages fetched at that moment and carries inline citations by default.

The same guide describes three depths:

• Non-reasoning search: the query goes to the search tool and the top results come back, with no planning step.
• Agentic search: the model searches as part of its reasoning, reads results and decides whether to keep going.
• Deep research: a longer investigation that often taps hundreds of sources and can run for several minutes.

Live search fixes the clock, not the source. In a February 2026 test of six AI chatbots on 2,100 questions drawn from same-day BBC reporting, the best systems scored over 90% on multiple-choice questions about events hours old. Even so, the authors attribute over 70% of all errors to retrieval, meaning the system landed on the wrong source, not that it reasoned badly.

Your pages can only be used if OpenAI's crawler may fetch them. Per OpenAI's crawler documentation, sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers, though they can still appear as navigational links. Search results can take about 24 hours to reflect a robots.txt change.

Anything that changed after the cutoff does not exist in the model's memory, and an old fact the web repeated for years can outweigh a new one. Brand errors fall into three types, and each needs a different fix:

1. Changed after the cutoff. A new price, CEO or product name is simply absent from memory. Only live retrieval can supply it.
2. Changed before the cutoff, restated rarely. Hundreds of old pages and a handful of new ones leave the old version dominant, even in a newer model.
3. Never stated plainly. If no page says what you do in one clear sentence, the model fills the gap with a guess from similar companies.

OpenAI does not publish a data mix for its current models, so how much weight repetition gets is our reading of the mechanism, not a measured number. The practical advice does not depend on it: the next model learns from the same web you left unfixed.

Fix the fast clock first, because it is the only one you can move this quarter. Make the correct fact easy to fetch and quote, keep crawlers allowed, and remove contradictions. The slow clock follows once the web agrees with itself.

1. Check crawler access. OpenAI recommends allowing OAI-SearchBot. In AEOeye's crawler access study (October 2026, 384 sites), only 16 sites (4.2%) blocked it, but 60 of 375 homepages (16%) challenged our self-identifying crawler, which a bot-protection layer can also do to retrieval bots. The study hub has the method and the per-site CSV.
2. State the fact in one dated sentence on pages you own: About, pricing, FAQ. "Acme costs $49 per month as of March 2026" is quotable. A paragraph of positioning is not.
3. Update what ChatGPT cites. Run the question, open the pages it cites, and correct or contact the pages it relied on. Publishers matter here: in our study 27 of 38 news publishers blocked at least one AI retrieval bot, so do not count on press coverage to carry your correction.
4. Retire contradictions. Update or 301-redirect every old page that repeats the wrong fact.
5. Re-test after a week or two for live answers, and judge memory only after a new model ships.

The full playbook is in our guide to fixing AI that gets your brand wrong. To see how sources get picked, read how ChatGPT chooses its sources.

Ask the same buyer question twice in fresh chats, once without search and once telling it to search the web, then compare. If only the first answer is wrong you have a memory problem; if both are wrong, a source it retrieves is wrong.

1. Write four questions: three unbranded buyer questions ("best CRM for a 10-person agency") and one about your brand's price or leadership.
2. Ask each in a new chat and note whether sources appear.
3. Ask again with an explicit instruction to search, and open every cited page.
4. Record the model and date. Compare against the documented cutoff before you blame the fix.
5. Repeat in other engines, since a fact fixed in one can stay broken in another.

AEOeye's free audit does a version of this for you: it asks ChatGPT three unbranded buyer questions live, shows the full answers and lists the competitors it recommends instead. It is a snapshot, not daily monitoring. For how this ties to getting recommended, see ChatGPT SEO, and for the three ways ChatGPT touches your site, does ChatGPT use my website?