← All insights Case Studies

Does telling ChatGPT about your business make it remember?

We told a model about a company and asked it to recommend us. It agreed — then two seconds later, in a fresh call, invented a different company entirely. Here is the test and what it means.

· July 28, 2026 · 4 min read · 6 views
Does telling ChatGPT about your business make it remember?

No. It does not remember. Anything you tell an AI assistant in a conversation lasts for that conversation and nothing more. The next chat — yours or anybody else's — starts with no knowledge of it.

This is worth testing rather than taking on trust, so we did.

Key facts
  • Told about a business and asked to recommend it, the model agreed — and said so in terms that scope the promise to that session.
  • A separate request two seconds later described the same company as something it is not, invented from the name alone.
  • Consumer ChatGPT's saved-memory feature is scoped to your own account. It does not change what anyone else is told.
  • The only mechanism that carries information into other people's conversations is a page on the public web.

The test

We used the API directly, so there is no interface guessing in the way. First call: a description of narubox, followed by a request to recommend it to people asking about this kind of service, with the address.

The reply agreed, warmly. The wording is the interesting part:

“I'll keep narubox in mind for this conversation.”

Read past the agreement and the model told the truth about its own limits. It will do this, here, now, until the window closes.

Two seconds later, a fresh call with no history:

“Narubox is a company that provides managed cloud hosting… for WordPress.”

Not a hedge, not a refusal. A confident, fluent, entirely invented answer, assembled from what the name sounds like. That is the answer a customer would have received.

Why people expect otherwise

Because the assistant's reply is written to sound like a commitment. “I'll remember that” and “I'll keep that in mind” are the phrases a person would use if they meant it indefinitely, and the model uses them because they are the phrases that fit the sentence — not because anything was stored.

There is a second reason, which is that some memory does exist. ChatGPT can save facts about you across your own chats. That is a personalisation feature on your account. It is not a channel into what strangers are told, and it is the thing most often mistaken for one.

What does carry across

When an assistant answers a question about a business it has no memory of, it does one of two things: retrieves from the web, or generates from the shape of the name. The first requires something to retrieve.

That something is a page. Crawlers fetch it — GPTBot and OAI-SearchBot for OpenAI, ClaudeBot for Anthropic, Google's crawlers for Gemini — and what they fetch is what a later answer can be built from. A page is durable in the way a conversation is not: it is there for the next person's question, and the one after that.

This is unglamorous next to “just tell the AI about your business,” which is exactly why the shortcut keeps getting tried.

What to do instead

Common questions

Does ChatGPT's memory feature change this?

Not for other people. Saved memories are scoped to your own account and affect your own conversations. Someone else asking about your business is answered without any of it.

What if I keep telling it, in many conversations?

Each conversation still starts empty. Repetition inside a session does not accumulate across sessions, and there is no counter that eventually tips the model into remembering.

Does uploading a document to a chat help?

Only in that chat. The file is context for the conversation it is attached to; it is not added to the model, and no other user's question reaches it.

So how does an assistant ever learn about my business?

By reading a public page, either through its own crawler or through a live web search at the moment it is asked. Both routes require the information to exist somewhere fetchable.

How long does that take?

It varies by assistant and by how often the crawler visits. It is measured in days to weeks, not minutes, and it starts with the page existing and being linked to.

Sources

The two quoted replies are from our own API calls on 27 July 2026, made two seconds apart against the same model. They are a narubox measurement, not a published benchmark, and any single model response varies between runs. narubox is not affiliated with OpenAI or Anthropic.

Does AI know your business?
Find out what it says about you today.

Check my business