Why ChatGPT Still Talks About Your Old Offer
You changed your pricing a year ago, discontinued a product, or renamed an offer. Yet when a prospect asks ChatGPT what you sell, it answers with the old version. Your site is up to date. That's not where the problem lives — it lives in the gap between what a model memorized during training and what it actually retrieves when it answers.
A generative AI answers in two very different ways depending on the question. Either it draws on its training memory — billions of texts read up to a given date, then frozen into the model's parameters. Or it triggers a live search — fetching pages from the web at the moment of the question, as Perplexity does by default, or ChatGPT and Gemini in search mode. Which of the two happens determines the answer you get, and you have no way to force it from the outside.
An up-to-date site doesn't stop an AI from answering with information memorized before your change. Until it triggers a live search on your brand, it repeats what it learned — not what you've published since.
How stale info settles in without anything changing on your end
The scenario is always the same: an AI read your old pricing, your old lineup, or your old positioning during its training phase, before a certain cutoff date. Since then, you've updated everything on your site. But if the user's question doesn't trigger a web search — because the tool answers in quick mode, or because the topic doesn't seem to need real-time verification — the model answers with what's in memory. The result: a wrong answer, stated with exactly the same confidence as a correct one.
This isn't an isolated bug or negligence on your part. It's a direct consequence of how these models are built: they don't re-read the web for every question, they lean first on what they've already absorbed.
Two response modes, two possible outcomes
| Response mode | Source of the information | Risk of stale info |
|---|---|---|
| Training memory alone | Text read up to a cutoff date, frozen in the model | High if your offer has changed since |
| Live search (web) | Current pages fetched at the moment of the question | Low, if your current pages are well indexed and up to date |
Key point: the same tool can switch between the two modes depending on how a question is phrased, without the user ever noticing. Two people asking a slightly different variant of the same question to the same AI can therefore get an up-to-date answer for one and a stale one for the other.
What improves your odds of getting the current version
You don't control which mode an AI chooses when it answers. But you can make the current version of your brand easier to find whenever a live search does happen:
- A visible last-updated date on your key pages (offer, pricing, products) signals clearly that the content has changed recently.
- Genuinely updated content — not just a date that changes without the text following, which is easy to detect and counterproductive.
- Consistency across all your public sources — site, social profiles, partner listings — so an AI cross-referencing multiple sources finds the same current version everywhere, not a mix of old and new.
None of this forces an AI to always search live. But it narrows the gap between what it would find if it searched, and what it has memorized if it doesn't.
What you can't fix directly
There's no form to report an error to a model and have it disappear immediately from its training memory. That memory only updates at the model's next training run, on a schedule you don't control. This is a fundamental difference from a Google Business Profile listing or a Wikidata entry, which you can correct yourself in minutes. Facing a model, the only realistic option is to act on what's verifiable live — your current pages — and monitor over time, rather than hope for a one-off correction.
Free GEO audit — we check what each AI says about you today
We test ChatGPT, Perplexity, Claude and Gemini on concrete questions about your offer, flag stale or false information, and identify the freshness signals to reinforce first. No commitment, delivered in 24-48 hours.
I want my auditFrequently asked questions
Why does ChatGPT keep repeating false or outdated info about my brand?
Because the answer can come straight from the model's training memory, frozen at a given date, instead of a live search on your current site. If the model didn't trigger a web search for that question, it answers with what it memorized — even if that's no longer accurate.
Does the knowledge cutoff affect every AI tool the same way?
No. Tools like Perplexity, or the search modes of ChatGPT and Gemini, fetch live pages for most queries, which reduces the risk of stale answers. Other usage relies more heavily on the model's memory. The same product can answer correctly in one mode and with stale info in another.
How do you fix stale info an AI keeps repeating about your company?
There's no button to correct a model after the fact. The only reliable move is to reinforce freshness signals on your public pages, then monitor regularly how each AI describes you so you can react as soon as false info appears, instead of discovering it months later.