Key Takeaways
- Stop thinking of it as one memory. An assistant draws on a trained-in picture, a live fetch, and whatever you publish. The three move at completely different speeds.
- Expect the trained-in picture to lag by months. Anything absorbed during training only changes when a new model version ships, which is why an old business name can survive long after you changed it.
- Use the live fetch when you need a fast correction. Pages an assistant can retrieve at the moment of answering are the fastest lever you have, and the only one that moves in days rather than quarters.
- Fix the source, not the answer. There is no place to submit a correction. You change the material an assistant reads, then wait for the answer to follow.
- Re-check on a schedule, not on a hunch. Answers vary between runs for reasons that have nothing to do with you, so a change is only real if it holds across several runs a month apart.
You changed your business name eight months ago, or moved, or dropped the service you no longer offer. An assistant is still describing the old version, and there is no obvious place to go and fix it.
The frustrating part is that the answer to "how often does it update" is not one number. An assistant is drawing on several different things at once, and they refresh at speeds that differ by orders of magnitude.
Here is what those channels are, how fast each one actually moves, and which of them you can push on when something about your business is wrong.
How often does ChatGPT update what it knows about your business?
There is no single update schedule. What the model absorbed during training changes only when a new model version ships, which is months apart. What it fetches from the web while answering can reflect a change within days of you publishing it.
That split explains most of the confusion. People ask a question, get a current answer, ask a slightly different question and get a stale one. Both are the same assistant. One answer went out to the web and one did not.
OpenAI's own description of what ChatGPT is frames it as a system trained on a large body of text that can also use tools to look things up. Those are two different clocks, and knowing which one you are looking at is most of the diagnosis.
The three channels, and why they move at different speeds
Three things feed an answer about your business: what the model absorbed in training, what you have published where it can be read, and what it fetches at the moment of answering. Slowest to fastest, in that order.
They are not ranked by importance. The trained-in picture is what fills in when nothing is fetched, which is often, so a stale one does real damage. The fetch is fast but only reaches pages that are reachable. What you publish sits in the middle, because it feeds both of the others over time.
Ticks when: A new model version ships
Ticks when: You post a page or update a profile
Ticks when: Someone asks a question that triggers a search
The tick spacing is a texture showing relative speed, not a count of anything, and the three grains are illustrative rather than measured. What is not illustrative is the ordering: training is the slowest channel, the live fetch is the fastest, and what you publish feeds both.
Written out, because nothing here may live only inside a figure. The trained-in channel ticks when a new model version ships, months to years apart. What you publish ticks when you post a page or update a profile, with weeks between meaningful changes. The live fetch ticks every time somebody asks a question that triggers a search. All three grains are illustrative.
Why the trained-in picture lags by months
Because training happens in large batches, not continuously. A model is trained on a body of text collected up to a point in time, and everything it learned that way is fixed until a later version is trained on newer material. That is the knowledge cutoff, and it is the source of most stale answers.
Volume makes it worse than the dates suggest. If your business traded under an old name for six years and a new one for eight months, the older name appears in far more of the material a model learned from. Even after the new name is in the mix, the old one has weight, so it keeps surfacing.
Much of that material comes from large public web collections. Common Crawl, whose archive is described at commoncrawl.org, is one widely used example, and it works in periodic crawls rather than a live feed. A page you published last week was not in a crawl that finished last year, and no amount of waiting changes that particular snapshot.
| Channel | What it holds | How fast it changes | What you can do about it |
|---|---|---|---|
| Trained-in picture | Whatever the public web said about you before the cutoff | Only when a new model version ships | Nothing directly, only change the balance of what exists |
| Live fetch during an answer | Pages it can reach and read at that moment | Days, sometimes the same day | Publish a clear, reachable page and keep it crawlable |
| Your own site | Exactly what you wrote, if it can be fetched | Immediately on your end | Update the page, keep the facts consistent everywhere |
| Profiles and directories | Your listing details, often cached by the platform | Days to weeks after you edit them | Claim them and correct every stale copy of your details |
| Third-party mentions | What other people wrote about you, whenever they wrote it | Never, unless the author edits it | Earn newer mentions rather than chasing old ones |
The bottom row is the uncomfortable one. An article describing your old positioning is permanent, so the fix is not correction but dilution: enough newer material saying the current thing that the balance tips.
What changes in days instead of quarters
Anything an assistant can fetch while it is answering. That is the fast lane, and it is the only channel that responds to work you do this week rather than work you did last year.
Getting into that lane has a mechanical requirement: the page has to be reachable by the crawlers these products use. OpenAI documents its own at its bots page, including how they identify themselves and how a site can allow or block them. A robots file that blocks them is a decision, and it should be a deliberate one rather than something a plugin did for you.
The second requirement is that the page states the fact plainly. A fetched page still has to be read and understood. If your new service is implied by a testimonial halfway down a case study, it will not get picked up as a fact about you, no matter how quickly the page was retrieved. The full repair sequence is in how to fix an out-of-date picture of your business in AI.
What to do about the lag
Change the sources and then measure on a schedule. There is no correction form, no support ticket, and no record of your business to edit, so everything you can do is upstream of the answer.
Start with consistency. The same business name, the same city, the same description of what you do, on your site and on every profile that carries your details. Contradictions are the single most reliable way to keep an assistant hedging, because it has no basis for choosing between two versions of you. Where those details are read from is covered in where ChatGPT gets information about local businesses.
Then set a re-check rhythm rather than looking whenever you feel anxious. Monthly is enough for most businesses, plus an extra pass after anything material changes. Run each question several times, because answers vary between runs on their own, and a difference you saw once is not a result. The cadence argument is in how often you should check if AI recommends your business.
Our free check runs 25 real buyer questions from your niche at $0 and emails a plain-English readout within the hour, including which stale details are still showing up in answers about you. It runs as soon as you ask for it.
Questions we hear the most
How often does ChatGPT update what it knows about your business?
It depends on the channel. The picture absorbed during training changes only when a new model version ships, which is months apart. What it fetches from the web while answering can reflect a change within days of you publishing it.
What is a knowledge cutoff?
The date after which a model saw no training data. Facts that changed after that date are not in the model itself, so anything it says about them comes from a live fetch or from nothing at all.
Can I submit a correction to ChatGPT about my business?
No. There is no correction form and no record of your business to edit. The only route is to change the published material an assistant reads, then wait for later answers to reflect it.
Why does AI still use my old business name?
Because the old name appears in more of the material it learned from than the new one does. Old references outnumber new ones for a while, and the trained-in picture only shifts when a later model version absorbs the corrected balance.
Does updating my website change AI answers immediately?
Sometimes, if the assistant fetches your page while answering. It will not change what the model absorbed during training, so expect a split for a while where one route is current and the other is not.
How often should I re-check what AI says about me?
Monthly is enough for most businesses, with an extra check after anything material changes: a new name, a move, a new service, a merged listing. More often than that mostly measures normal variation between runs.
How do I find out what AI says about my business today?
Ask several assistants the questions your buyers would ask and record every name and detail that comes back. Run each question more than once, because a single answer is noise rather than a measurement.
What does the AnswerHalo free check include?
It costs $0. We ask ChatGPT 25 real buyer questions from your niche, then email a plain-English readout within the hour: how many answers named you, who got named instead, and the first thing to fix. It runs as soon as you ask for it.
Your buyers are already asking. Find out what AI tells them.
The free check asks ChatGPT 25 real buyer questions about your niche and sends you the report within the hour, at $0. If it shows you are already getting named everywhere, we will say so in plain words.
Get my free check →