Key Takeaways
- Treat the cutoff as a photograph, not a memory. A model learned what it knows up to one date and stopped. Nothing you did after that date is in there unless something else puts it there.
- Assume the picture is older than the cutoff. Training data is gathered from pages that were already crawled, so most of what a model holds about you was written well before the date on the label.
- Know which mode you are watching. An assistant either answers from the frozen picture or reads the live web first. Those two modes give different answers about you, and only one of them can see this week.
- Publish the change where it can be read again. Updating a page is what lets the live mode find the new fact. It is also the only thing that puts the new fact in front of the next round of training.
- Check the picture rather than guessing at it. You cannot tell from the outside which facts about you are stale. Asking the question your buyer asks, and reading what comes back, is the only honest way to find out.
Ask an assistant about your business and you may get an answer that was true two years ago. The old address. The service you stopped offering. A price you have not charged since before you rebranded.
This is not a bug in the way people usually mean it. It is the direct result of how these systems are built, and it has a name that turns up in every model maker's documentation: the knowledge cutoff.
Here is what the term means, why it leaves your business described in the past tense, and which of the two things you can do about it actually works.
What is a knowledge cutoff?
A knowledge cutoff is the date after which a model learned nothing new. The company gathers a very large body of text, trains the model on it, and stops. Everything the model knows by itself comes from before that line.
That is the whole definition, and it is worth naming the term once so you recognize it in the small print. From here on we will call it what it is: a fixed date, after which the model is working from memory.
OpenAI describes this plainly in its own explanation of what ChatGPT is, and the material being learned from is mostly public web text of the kind gathered by projects like Common Crawl. Your website, if it was reachable and had been crawled, is somewhere in there. The version of it that existed at the time.
Why that freezes the picture of your business
Because your business kept moving and the picture did not. Every change you made after the cutoff exists in the world and not in the model, which will still answer confidently using what it has.
The gap is usually wider than the date suggests. Training material is collected from pages that were already crawled and stored, so a page that was last rewritten a year before collection is a year older again by the time it is learned from. The label says one date. The contents are older.
An illustrative twelve month run. The three changes below the line are the kind that matter most to a buyer, and they are exactly the ones a model trained before them cannot know about.
Those three changes are illustrative rather than measured, but the shape is the common one: a certification, a move, and a price change, all made in the months a model cannot see. Each is a fact a buyer asks about directly.
What actually updates what an assistant knows
Two things, and only two. The assistant reads the live web while it answers you, or the model is trained again from scratch on newer material. Nothing else moves the picture.
The first happens constantly and is within your reach. When an assistant fetches pages to answer a question, it is reading what is published today, which means a page you updated last week can be quoted this week. Google describes the same arrangement for its own AI features in Search: answers assembled from pages its systems have already crawled.
The second happens on the model maker's schedule, not yours, and it is measured in many months. You cannot request it and you cannot pay for it. The only influence you have is making sure that when it does happen, the material about you that gets learned from is current and says what you want it to say.
| Question | The frozen picture | Reading the live web |
|---|---|---|
| How current is it? | Stops at the cutoff date, and the contents are older still | As current as the last time the page was crawled |
| Who controls the timing? | The model maker, on a schedule measured in many months | You, by publishing, plus the crawler's own pace |
| Does it run on every question? | Always, it is the default | Only sometimes, depending on the tool and the question |
| What is it good at? | Background, categories, who is generally worth naming | Specifics: prices, hours, the current offer, this year's news |
| How do you influence it? | Publish now so the next training round learns the right thing | Publish now and keep the page reachable |
How to tell which mode you are looking at
Look for citations. An answer that names sources and links out was assembled by reading pages during the answer. An answer with no links, delivered in a smooth confident paragraph, most likely came from the frozen picture.
The second tell is specificity about the recent past. Ask about something that changed in the last few months. If the answer is vague, hedged, or simply wrong in the direction of the old facts, you are reading memory rather than a live look.
The third tell is inconsistency. Ask the same question twice and watch whether the answer moves. Variation between asks is normal and has its own causes, which is the subject of why the same question asked twice gives different answers, but a wild swing between a current fact and an old one usually means one ask browsed and the other did not.
What to do when the picture is out of date
Fix the sources, not the answer. You cannot edit a model, and no amount of telling it that it is wrong in a chat window changes anything for the next person who asks. What you can change is what is published about you.
Start with the page that states the fact plainly, in a sentence a machine can lift whole. Then get the same fact onto at least one page you do not own, so the live read finds it in two places instead of one. The full repair sequence is laid out in how to fix an out-of-date picture of your business in AI.
Then be patient in the right proportion. The live reading mode can pick up a change within days or weeks. The frozen picture will not catch up until the next training round, and how often that comes around is covered in how often ChatGPT updates what it knows. Publishing today is what makes both of those go your way.
Before you rewrite anything, it is worth knowing which facts are actually stale, because it is rarely the ones you expect. The free check runs 25 real buyer questions from your niche through ChatGPT at $0 and emails a plain-English readout within the hour. It runs as soon as you ask for it.
Questions we hear the most
What is a knowledge cutoff in simple terms?
It is the date an AI model stopped learning from new material. Everything it knows on its own was written before that date, so anything about your business that changed afterward is missing unless the assistant looks it up while answering.
Can ChatGPT see my website right now?
Sometimes. When it browses, it reads live pages while answering your question. When it does not browse, it answers from training alone, and your current site is invisible to it no matter what the page says today.
How do I find out what an assistant thinks it knows about me?
Ask it directly, in a logged out session, and read the answer for facts rather than tone. Wrong prices, an old city, or a service you retired are all signs you are being described from the frozen picture.
If AI can browse the web, does the cutoff still matter?
Yes. Browsing is not guaranteed on every question, and a model still leans on training for background and for judgments about who is worth naming. The frozen picture shapes the shortlist even when the live read supplies details.
How long does a change on my site take to show up in AI answers?
For the live reading mode, as soon as the page is crawled again, which is usually days to weeks. For the frozen picture, only at the next training round, which is measured in many months and is nobody's decision but the model maker's.
Does deleting the old page fix a wrong answer?
Not by itself. Deleting removes the source from the live web but leaves whatever was already learned in place, and it can make things worse by removing the page that would have corrected the record. Replace the page rather than removing it.
How do I check whether AI recommends my business at all?
Ask several assistants the questions your buyers actually type, repeat each one a few times, and write down who gets named. A dozen questions asked three times each gives you an honest baseline in an afternoon.
What does the free 25-question check include?
It asks ChatGPT 25 real buyer questions from your niche at $0, then emails a plain-English readout within the hour: how often you were named, who was named in your place, 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 →