Reading time: 12 minutes
ChatGPT becomes outdated because it learns from text up to a knowledge cutoff and does not always search the web. Even when it searches, it may read old pages or the stale copy that a search engine stored. So the AI can repeat a company’s outdated information for months, even after the company fixes its site.
In short
- Without search, ChatGPT only knows what it learned up to its knowledge cutoff. In the API documentation, the models it uses today have cutoffs in April and May 2026.
- By our count, the current ChatGPT, Gemini and Claude models reached the public three to seven months after their cutoff.
- Web search helps, but AI may read the outdated version that stays in search engine indexes. According to Google, a new crawl can take from several days to several months.
- In an Ahrefs study, pages cited by AI assistants were 2.9 years old on average. So the fix starts with the sources AI reads, not with the chat.
What is an AI model’s knowledge cutoff?
The knowledge cutoff is the limit of what a model learned in training. Anything that happened after it stays out of the model’s memory. It only reaches the answer if the AI searches the web. That is why every model becomes outdated at some point, and ChatGPT is no exception.
OpenAI calls this limit the “knowledge cutoff” in its Help Center. According to the company, answers do not include events after that point unless the model uses tools, such as search. The same page notes that ChatGPT may fail to read a site because of technical issues, paywalls or a robots.txt block.
The cutoff is not the same for every topic, either. In the Gemini 3.8 Flash model card, Google gives a March 2026 cutoff. However, it warns that in some domains the model’s knowledge is limited to January 2025. So a brand that changed its name before the cutoff may still show up with its old name. The full path of an answer, from training to search, is in the guide on how ChatGPT works.
What is the knowledge cutoff of ChatGPT, Gemini and Claude today?
In the API documentation, the ChatGPT knowledge cutoff ranges from April to May 2026, depending on the plan. In the newest Gemini models, the cutoff is March 2026. Meanwhile, the current Claude models stop at June 2026, according to Anthropic. So anything that changed after these dates only reaches the answer through search.
| Model | Where it appears | Knowledge cutoff |
|---|---|---|
| GPT-6 Sol | ChatGPT Plus, Pro, Business and Enterprise | April 20, 2026 |
| GPT-6 Luna | ChatGPT Free and Go | May 18, 2026 |
| Gemini 4 Argon | Gemini app, on Google AI Pro and Ultra | March 2026 |
| Gemini 3.8 Flash | Gemini app | March 2026 |
| Claude Opus 5.5 and Sonnet 5.5 | Claude | June 2026 |
According to the release notes, Plus, Pro, Business and Enterprise have used GPT-6 Sol since October 2026. Free and Go, on the other hand, use GPT-6 Luna. However, the note does not give the ChatGPT knowledge cutoff. That is why the table uses the OpenAI API documentation.
The Gemini data comes from Google’s model cards, and the Claude data comes from the Anthropic models overview. In practice, ChatGPT stays outdated on anything that came after April or May 2026 if it does not search the web.
Why does AI use a company’s old name or price?
Because the old data has more support than the new data. The old name shows up in years of pages, news, profiles and reviews, while the new one shows up in only a few. So an answer without search repeats what the model learned. An answer with search, on the other hand, may cite a page that nobody updated.
Four causes explain why ChatGPT seems outdated about a company:
- Knowledge cutoff: the change came after the cutoff, or close to it, and the model did not learn it.
- Answer without search: the AI answered from memory, without checking the web.
- Old index: the search engine still keeps the previous version of the page.
- Old sources: news stories, directories, marketplaces and PDFs stay online with the outdated information, and the AI cites those pages.
Think of a company that changed its name in April 2026. The change came after the Gemini 4 Argon cutoff and right at the edge of the GPT-6 Sol cutoff. Even in Claude, with a June cutoff, the new name would have only two months of text on the web.
In search, old sources also carry weight. In a study by Ahrefs, which sells SEO tools, pages cited by AI assistants were 2.9 years old on average, counting from publication. The average covers ChatGPT, Perplexity, Copilot and Gemini. In Google’s organic results, on the other hand, it was 3.9 years. The 2025 study analyzed about 17 million citations.
Old data, however, is different from made-up data. When AI creates something that was never true, the problem is a different one. The guide on why ChatGPT makes so many mistakes explains it. The step-by-step fix, source by source, is in the guide on finding the source of the error and fixing it.
Does web search fix outdated information?
Only in part. Search brings pages that are newer than the training data, but not every answer uses the web. In addition, search starts from an index, which may keep the old version of the page. Finally, the AI may cite an outdated source that is still online.
Each assistant decides on search in its own way. According to OpenAI, ChatGPT may search on its own when current information would help the answer. Users can also ask for a search. In Claude, according to Anthropic, search kicks in when the question calls for current information. Perplexity, in turn, searches the internet for every question, according to the company’s help center.
Without search, ChatGPT stays outdated because it answers from training alone. Even with search, OpenAI warns that results and citations can be incomplete, outdated or incorrect. In other words, the AI may present an outdated page as if it were the current version. That is why OpenAI recommends opening the source and checking the page’s publication or update date. To tell if an answer came from training or the web, see the guide on when ChatGPT searches the internet.
Why does AI with search still read the outdated page?
Because AI search depends on indexes that crawlers build. In AI Overviews and AI Mode, a page can only appear as a source if it is in the Google Search index. The rule is on Google’s page about AI features. ChatGPT, for its part, combines partner search providers with its own search crawler, OAI-SearchBot. Perplexity, in turn, indexes the web with PerplexityBot.
The index works like a cache: it keeps what the crawler saw on its last visit. According to Google, Search stores the information about each page in the Google index, a large database. Until the crawler comes back, the AI may read the outdated version. On the same page about AI features, Google notes that this new crawl can take from several days to several months.
Sometimes ChatGPT and Perplexity open the page at the moment of the question, with the ChatGPT-User and Perplexity-User agents. Even so, nothing guarantees that they will pick the new page.
How long does it take for AI to reflect a brand change?
It depends on where the answer comes from. In search, the change shows up after the search engine crawls the page again, which takes days, weeks or months. But in the model’s memory, the change only enters a model trained after it. In addition, current models reached the public months after their cutoff.
We did the math with the official dates. GPT-6 Sol, with an April 20, 2026 cutoff, arrived in the ChatGPT Chat tab in October, almost six months later. Gemini 4 Argon, with a March 2026 cutoff, also reached the Gemini app in October, about seven months later. Claude Sonnet 5.5, with a June 2026 cutoff, came out in September, close to three months later.
So a change made after the ChatGPT knowledge cutoff only enters a future model, which may take months to arrive. Meanwhile, without search, ChatGPT stays outdated about the company. In search, the AI may show the outdated version until the next crawl.
A domain change adds another wait. According to Google, most pages on a small or medium-sized site move in the index within a few weeks. Larger sites, however, take longer.
What should you do after a rebrand so AI updates?
Write the new name in every source and link it to the old one. Then notify search engines and track the answer with a fixed set of questions. After all, search finds what is in the index. The model, on the other hand, learns best what appears in many texts, as the guide on AI hallucination shows.
- Say on your site that the name changed. On the home page and the About us page, write, for example, “XPTO, formerly ABC,” with the date of the change.
- Redirect the old domain. Google recommends permanent redirects, such as 301 and 308, kept for at least one year.
- Let search crawlers in. OAI-SearchBot, PerplexityBot, Googlebot and Bingbot need to read the new pages, as the guide on how to show up in ChatGPT shows.
- Notify search engines. Submit the sitemap, request a new crawl in Search Console and turn on IndexNow, as the guide to indexing your site on Bing explains.
- Fix third-party sources. Update your Google Business Profile, social media, directories and marketplaces. For old pages, see how to remove outdated content from Google.
- Build mentions of the new name. Press releases, interviews, partners and customers spread the change in texts that AI reads.
- Track the answer. Ask about the old name and the new one, with and without search, every week, and write down the source of each answer.
This way, even with an outdated model, ChatGPT can find the new name when it searches. Correcting ChatGPT in a chat, however, is not enough. With memory on, the correction may apply to future chats of the person who made it, but not to your customers’ chats.
What if the news is a product launch or a new price?
The plan is the same, with a dedicated page for the news. It should state in text the product name, the price, the launch date and what changed. Next, remove the old price from PDFs, tables and forgotten pages, because AI can still find that outdated information.
If Gemini says a product launched in August does not exist, check whether the answer shows the Sources button. Without it, according to Google, Gemini did not provide links for that information. In other words, the answer may have come only from the model’s memory. If ChatGPT has outdated facts about the product, OpenAI suggests asking for a new search with a specific source, date or location.
What we found when we analyzed 8 pages in Portuguese about outdated ChatGPT answers
We analyzed 8 pages in Portuguese about the knowledge cutoff and an outdated ChatGPT. They came up in 5 searches in Portuguese. Only 1 gives the cutoff of the models ChatGPT uses today, and 2 still cite September 2021. Four describe the current ChatGPT web search, but only 1 warns that AI may cite an outdated page.
None explains that search engine indexes keep the old version of a page until the next crawl. Likewise, none says how long a change takes to reach the answer. Finally, only 1 gives recommendations for companies, and they do not cover rebranding.
Frequently asked questions
Not in training. In the API documentation, GPT-6 Luna, used in ChatGPT Free and Go, has a May 18, 2026 knowledge cutoff. GPT-6 Sol, used in Plus, Pro, Business and Enterprise, has an April 20, 2026 cutoff. In addition, web search is available on all plans, according to OpenAI.
You can ask, but the answer is not the most reliable source. Google, for example, warns that Gemini can misrepresent how it works, including when people ask about its training. That is why you should check the knowledge cutoff in each company’s documentation, as in the table in this guide.
Not right away. A removed page drops out of the index after the next crawl, and then AI stops finding the outdated version in search. However, a model that already finished training keeps what it learned up to its knowledge cutoff. So the old name tends to disappear only in models trained after the change.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Does ChatGPT tell the truth? (2026).
- OpenAI, ChatGPT release notes (2026).
- OpenAI, GPT-6 Sol page in the API documentation, with the knowledge cutoff (2026).
- Google DeepMind, Gemini 3.8 Flash model card (2026).
- Anthropic, Claude models overview (2026).
- Google Search Central, AI features and your website (2025).
- Google Search Central, site moves with URL changes (2026).
- Ahrefs, study on the age of pages cited by AI assistants (2025).
