Why AI Still Describes Your Old Brand: Knowledge Cutoffs Explained

If an assistant describes your company as it was two years ago, the cause is usually a knowledge cutoff rather than a bad source. Free ChatGPT sessions currently run on a model whose training knowledge ends in October 2023, and a model answering from memory has no way to know that anything changed since. The test that separates this from a genuine source problem takes about ninety seconds.

The rebrand went live. The new office opened. The product was discontinued. Your site says all of it clearly. And the assistant still describes the old version of your company with complete confidence.

This is the most common AI reputation complaint we hear, and it is also the one most often misdiagnosed. Teams assume a stale page is feeding the engine and go hunting for it. Often no such page exists.

Two sources, one answer

Every answer you get is built from two ingredients.

Training memory is what the model absorbed before its cutoff date, compressed into weights. It is not a database of documents. There is no row to update and no cache to clear. It behaves like a person recalling something they read a while ago.

Retrieval is what the engine fetches from the live web during your conversation. It reflects the world as of now, and it is the part you can influence.

When those two disagree, the engine blends them, and the blend is where the old version of you survives.

The ninety-second test

Run your question three ways and compare.

1. ChatGPT with search off. Pure memory. Whatever it says here came from training.
2. ChatGPT with search on. Memory plus retrieval. If this fixes the answer, the current fact is findable and the earlier answer was memory.
3. Perplexity. Retrieves for nearly every query and numbers its citations. If it also gets you wrong, a live page is producing the error and the citations will name it.

Three readings and three different fixes:

• Wrong without search, right with search. Knowledge cutoff. Your published record is fine. Nothing you write today removes the memory, and it will fade with the next model release.
• Wrong both ways, and Perplexity names a source. A retrieval problem. Go and fix or supersede that page.
• Wrong both ways, and Perplexity cites nothing relevant. A gap. The engine found nothing solid and filled the space. Publish the clear, dated answer that is missing.

Why cutoffs bite harder than people expect

Three things make this worse than the raw dates suggest.

Free tiers lag paid ones. The model serving a signed-out user is often not the newest one available. Your buyer is more likely on the free tier than you are.

Retrieval is not guaranteed. An engine decides whether a question needs a search. Ask something that sounds like general knowledge and it may answer from memory without checking anything. Claude in particular will often answer without searching, and whether it searches is itself a signal about how strong your entity is.

Old material outweighs new material. If your former name was written about for six years and your new name for six months, the older version has more supporting text behind it. Volume and time both count.

What actually shortens the lag

You cannot edit training memory. You can make sure that every time the engine does retrieve, it finds something unambiguous and current.

Publish the change as a fact with a date

Not a press release from three years ago. A page that states plainly what changed, when it changed, and what the position is now. Visible dates matter because engines weigh recency, and an undated page is hard to prefer over an old one.

Say the old name once, on purpose

A single clear line connecting the previous name to the current one gives the engine the bridge it needs. Delete every trace of the old name and you leave two disconnected entities, which is how a rebrand turns into a merged-identity problem.

Fix the machine-readable record

Structured data, Wikidata, business listings and industry registers. If these still carry the old name, address or description, retrieval keeps confirming the old version. This is dull work and it moves answers more reliably than content does.

Update the third-party pages you do not own but can reach

Review platforms, directory listings and partner pages. Profiles on Trustpilot, G2 and Capterra are roughly three times more likely to be cited than an equivalent page elsewhere, and most of them let you correct your own entry.

How long it takes

Retrieval-driven answers can move within weeks once the better source is published and crawled, and Perplexity usually shows it first. Memory-driven answers wait for a model release, and nobody outside the lab knows that date. Anyone offering you a fixed timeline for the second category is guessing.

This is why we report each engine separately rather than as one score. Perplexity improving while ChatGPT stays wrong is not a contradiction, it is exactly what a memory problem looks like in the data.

Frequently asked questions

What is ChatGPT’s knowledge cutoff?

It depends on which model is serving the session. Free sessions currently run on GPT-4o, whose training knowledge ends in October 2023. Paid tiers use newer models with later cutoffs. Assume the person asking about you is on the older one.

Can I ask the model to forget the old information?

Within one conversation you can correct it, and the correction applies to that conversation. It does not carry to anyone else, and it does not change what the model knows.

Will the next model release fix it automatically?

Only if the corrected version of you was well represented in the material available before that model was trained. Which is the argument for publishing the correction clearly now rather than waiting.

Does blocking the crawlers help?

It does the opposite. Blocking removes you from the retrieval pool while leaving whatever is in training memory untouched, so the outdated version becomes the only version. Google-Extended in particular is worth understanding before anyone touches it, because it governs generative grounding without affecting Search crawling.

Where this fits

Running this test across your real prompt set, engine by engine, is the first step of our AI reputation management engagement. If you want to run the baseline yourself first, we published the prompt set we use.

Related guides

• What Is AIO? AI Optimization vs AI Overviews, Explained (2026)
• What Is LLMO? Large Language Model Optimization Explained
• How to Build a Knowledge Panel and Brand Entity for AI Search
• Google’s Search Generative Experience (SGE) and AI Overviews: What Egyptian Brands Need to Know in 2026