How often does ChatGPT update what it knows about my business?

There is no file about you that gets refreshed

The question assumes ChatGPT keeps a record of your business and re-reads it on some schedule, and that is not how any of this works. There is no row with your name on it, no last-updated timestamp, and no queue your business sits in waiting to be re-read. There are two entirely separate mechanisms that can put a fact about you into an answer, and they update on different terms.

Training. A model is built from a corpus of text collected before the build. OpenAI documents GPTBot as the crawler that collects content which may be used to train its foundation models. Whatever that corpus contained about you is what the model carries, and it does not change afterwards. Publishing something today cannot alter a model that already exists. It can only be material for a corpus that has not been gathered yet.

Retrieval. When ChatGPT answers with a live search, it fetches pages at the moment you ask. OpenAI documents OAI-SearchBot as the crawler used to surface websites in ChatGPT's search features. The material behind that answer is whatever those pages said when they were last fetched, so a correction on a page can show up in an answer without anything about the model changing at all.

Most real answers are a blend of the two, which is why the same business can be described correctly and incorrectly in two consecutive replies. If what you actually want to know is how long until an engine names you for the first time, that is a different question and it has its own page: how long does it take to show up in ChatGPT.

How to tell which path a given answer used

Look for citations, because that is the tell. If the reply carries source links, retrieval ran and those specific pages are the material the answer was built from. If it carries none, you are looking at the model composing from what it already holds, and nothing you published this week is in it.

The check takes about a minute, and it is worth doing on every answer you are about to react to.

1. Ask the question in a fresh chat, so nothing earlier in the conversation is doing the work.
2. Look at the reply for citation chips, footnote markers or a sources list. Note whether there are any at all.
3. If there are, open them. Those pages are the answer's raw material, and correcting or getting onto them is the only lever that touches this reply.
4. If there are none, ask the same question again and add a request to search the web. If the answer changes materially once it searches, the first version came from memory.
5. Record both versions with the date. A single answer is one sample, and these systems are not deterministic.

Two honest limits on this test. A reply can blend both paths, citing one page while carrying unsourced detail from training around it, so citations tell you retrieval happened, not that everything in the answer came from those pages. And a citation is evidence that the page was used in composing the reply, not proof that every sentence traces back to it. Which sources are worth chasing once you have that list is covered in where ChatGPT gets its information about your business. One more thing this test cannot settle: an updated, accurate description of you is a different operation from being put forward to a buyer, which is the difference between an engine recognising you and recommending you.

Any specific number of days you have read elsewhere is unsourced

Nobody publishes the figure, so nobody can honestly quote it. OpenAI does not publish a crawl frequency for a given domain, a recrawl request mechanism for its crawlers, or a retraining schedule. Google does not publish an update cadence for what its models hold about a named business either. Those numbers are not withheld from the public and available to agencies. They are not published at all.

So when you see "ChatGPT refreshes every 30 days", or a median in days, or a quarterly cycle stated as fact, the question to ask is what was measured, over how many businesses, on which dates, and where is the raw data. If those four are not available, the number is decoration. We have seen agencies quote one.

The method that replaces the number you cannot have

Measure the two things you can actually observe: when a crawler last visited you, and whether the answers have changed. Neither requires a vendor to tell you anything.

Your server logs are the only direct evidence available to you. Every fetch by an AI crawler lands in them with a user agent string. Search your access logs for GPTBot and OAI-SearchBot, and for the other engines' crawlers while you are there. That gives you real dates for real visits to real URLs on your own site. It tells you nothing about when a model will be retrained, and it does not tell you whether the fetched page was used in any answer, but it is the difference between knowing and guessing about the retrieval path.

The answers themselves are the other observable. Ask a fixed set of questions, in the same words, on a fixed cadence, and record what came back with the date. Movement in that record is the only evidence that anything updated, whichever path caused it. The cadence worth running is in how often you should check your AI visibility, and what counts as real movement rather than noise is in how to know if the work is working.

One prerequisite sits underneath both. If your robots.txt is blocking these crawlers, the retrieval path is closed and nothing else on this page can apply to you. In our scan, 13% of the 300 Australian businesses we checked blocked AI crawlers outright in robots.txt. The trade-offs are laid out in should I block AI crawlers.

What to do while you are waiting, if the engine has a fact wrong

Fix the source rather than the chatbot, because there is nothing at the chatbot end to fix. A wrong fact that arrived by retrieval is on a page somewhere, and that page can be corrected. A wrong fact that arrived from training cannot be edited out of an existing model by anyone, including OpenAI, so the only thing that helps is making the correct version the one that retrieval finds and the one a future corpus would collect.

That is the whole reason the citation check above matters before you spend a week on anything. The full repair sequence, including what the in-product feedback controls realistically do, is in ChatGPT has wrong information about my business.