Back to Blog
  • AI SEO
  • AI Search Game Plan
  • AI Search Visibility
  • Social Search
  • AEO
  • Citation-Ready Content

How ChatGPT Gets Its Information: Memory, Search, and Your Business

Hayden BondUpdated 10 min read
How ChatGPT Gets Its Information: Memory, Search, and Your Business
ChatGPT gets its information two ways: from what it learned in training, which works like memory, and from web searches it runs during a chat. Most of the time it doesn't search at all. Ahrefs went through 66.6 million free, logged-out ChatGPT answers from August and September 2026, and only 14.4% ran a search. Call it 1 in 7. If your business only shows up when ChatGPT searches, you're betting on that 1 in 7.

6 in 7

Free, logged-out ChatGPT answers that ran no web search

Source: Ahrefs, October 2026 ↗

The two ways ChatGPT knows things

Think about how you answer a question. Sometimes you just know. Sometimes you look it up. ChatGPT does both.
The first is memory. Before a model comes out, it's trained on a huge pile of text, and what it learns stays inside the model. People in the field call this parametric knowledge. It stops at a cutoff date and doesn't change until the next model is trained.
The second is search. When a question needs current facts, ChatGPT can search the web, read pages, and cite them. People in the field call this the retrieval layer.
The question decides which one ChatGPT leans on. And which one it leans on decides what you can do about it.

Two ways ChatGPT answers

Each answer leans on memory or on a web search. Each one has a different fix.

Memory

About 6 in 7 answers

What it uses
What the model learned in training
When it's set
At the model's training cutoff
When it changes
When a new model is trained
How you show up
Be described the same way in many trusted places

Search

About 1 in 7 answers

What it uses
Pages it finds and reads during the chat
When it's set
Each time it searches
When it changes
When your pages are crawled again
How you show up
Rank for the question, answer it plainly, name your sources
Shares: Ahrefs, 66.6 million free, logged-out ChatGPT answers, August to September 2026.

Why ChatGPT's memory matters for your business

Answer from memory, and nothing gets fetched. Nothing gets cited. There's no page for the model to visit, so your website can't help you in that moment, no matter how good it is. What helps is having been written about, plainly and the same way, in the places the model learned from.
Nobody outside OpenAI knows exactly what today's models trained on. The last time OpenAI shared its mix, in the GPT-3 paper from 2020, ordinary web pages made up 60% and Wikipedia made up 3%. We can't see today's mix. But if it looks anything like that one, the wider web shapes what a model remembers far more than any one site does.
That's why I think of entity SEO and AI search as one job. Your Google Business Profile, your directory listings, your social profiles, how trade press describes you, and the structured data on your site either agree about who you are or say different things. I wrote more about how AI search engines decide which brand you are.

When ChatGPT searches, it doesn't start fresh

A search doesn't wipe the model's memory. When what it remembers and what it just read disagree, the old belief can win.
Two 2026 studies point that way, though neither tested ChatGPT itself. The OAKS benchmark fed 14 models a stream of facts that changed over time. They averaged 39.4% accuracy on one test set, and 33.3% when the facts changed often. Some models clung to old answers. Others flipped too fast.
A team at Heidelberg tested 13 open models on sources that disagreed. The models trusted government sources over newspapers, and newspapers over people and social media. But repeat a claim from a weaker source and the order could flip. The authors compared it to the illusory truth effect, where people start to believe things they've heard often.
I take both as a warning more than a measurement. If a model learned something old or wrong about your business, fixing one page may not do it. The same correct facts, repeated in a lot of trusted places, give you a better shot.

Getting mentioned isn't the same as getting cited

BrightEdge ran thousands of prompts through ChatGPT in August 2025. It mentioned brands 3.2 times as often as it cited them with a link. ChatGPT talks about businesses it knows, link or no link. It's one reason AI can cite your content and still recommend your competitor.
Rand Fishkin of SparkToro goes a step further. In a May 2026 video, he said that most of the time the links are "post hoc rationalization": the AI picks the brands first, then goes and finds pages that back them up. He says it's not always that way, and it's his read, not a study. If he's right, being known matters more than any one page that gets linked.
It also means a buyer can hear your name in an answer and never click anything. Your traffic reports won't see that. Our measurement page covers what to watch besides clicks.

ChatGPT personalizes answers to the person asking

The person asking changes what ChatGPT says, too.
ChatGPT has a feature called Memory. It's not the training memory I described above. This Memory is about the person asking. It saves things they've told ChatGPT. Since April 2025, it can also draw on their past chats, starting with paid plans.
That can change a search. OpenAI says that when Memory is on, ChatGPT may use saved details to rewrite what it searches for. Their example: tell it you're vegan and live in San Francisco, then ask for restaurants nearby, and it may search for "good vegan restaurants San Francisco." It may also use your rough location, based on your IP address, to find local results.
So two people can ask the same question and hear about different businesses. That's why I don't put much weight on one screenshot of what ChatGPT says about you. It's one person's answer. You can't control who's asking. You can work on the part every answer starts from: what the model knows about you and what it can find. I go deeper on personalization in what comes after persona marketing in AI search.

How ChatGPT picks pages when it does search

Google ranking seems to count for a lot here. Kurt Fischman's February 2026 working paper looked at 730 citations from ChatGPT and Gemini together, across 75 business questions. Pages in Google's top spot got cited in 43% of the questions where they showed up. By the seventh spot, 5%.

How often a page got cited, by its Google position

Top Google spot

43%

Seventh spot

5%

Kurt Fischman, working paper, February 2026. 730 citations from ChatGPT and Gemini across 75 business questions.

Schema markup by itself didn't seem to matter. Pages that got cited used it about as often as pages that didn't (43.1% against 44.8%). It's one paper, tested on older models, so I'd treat it as a lead, not a rule. I'd still get schema right. I just wouldn't count on it to carry you.
The basics of citation-ready content still matter. Each section answers one question. Each passage makes sense lifted out on its own. Every number names its source. And the crawlers aren't blocked.

ChatGPT and Perplexity pull from different places

Profound ran 100,000 prompts through ChatGPT and Perplexity in 2025. Only 11% of the sites they cited showed up in both. Each tool has its own crawler, too: OAI-SearchBot for ChatGPT, PerplexityBot for Perplexity, and plain Googlebot for Google's AI Overviews. They don't search the same way, either. Perplexity says it searches the internet in real time when you ask it a question, while free, logged-out ChatGPT usually answers from memory.
Some sources do travel. In Profound's study of 680 million AI citations from August 2024 to June 2025, Wikipedia got 7.8% of all ChatGPT citations, and 47.9% of the citations that went to ChatGPT's ten most-cited sites. I read that as a sign of what these tools lean on: pages that describe things plainly and that other people check.
So you can be named in one tool and missing from the next. Memory is the thread that runs through all of them, and I'd bet a business that's described the same way in many trusted places is easier for any model to recognize. That's the idea behind our AI search optimization work.

Every new screen runs on the same two things

At a search marketing conference in Seattle in 2008, I overheard someone say, "Bing likes my writing better and Google likes his writing better." That sounded like a lot of work to me then, and it still does.
I still see teams split up the same way, one person on Google, one on ChatGPT, one on voice. Add glasses that answer questions about whatever you're looking at, and that's four screens to chase. But every one of those screens draws on the same two things: what the model already knows and what it can find. That's why I'd work on those two instead of chasing each screen.

How to check where your business stands

This isn't the whole job. It's where I'd start.

Memory: does ChatGPT know you before it searches?

  • Open a Temporary Chat in ChatGPT and choose Unpersonalized before your first message, so it skips your Memory, and ask about your business. Ask it not to search the web, and check that the answer shows no sources or links. Ask Gemini the same thing, then check both answers against what's true.
  • Search your business name on Google. Does a knowledge panel show up?
  • Do your Google Business Profile, directory listings, LinkedIn page, and social profiles describe you in the same words?
  • Do trade publications or local news mention you, with the same name and description?
  • Wikipedia has strict rules about who gets a page. If you don't meet them, skip it. I'd put the time into the rest of this list.

Search: can ChatGPT find and cite you when it looks?

  • Does your robots.txt let in OAI-SearchBot, PerplexityBot, and Googlebot?
  • Does each section of your main pages answer one question on its own?
  • Does every number on your site name its source?
  • Is your schema right, with Product or Review details filled in where they fit?
  • Do you show up on Google's first page for the questions your buyers ask?
If you'd rather we take a look, the AI Search Game Plan is a one-time review of where you stand and what to fix first. For tracking over time, we compare the tools in our AI visibility tools directory.

What stays the same when the screens change

Think about the phone. Rotary, cordless, flip phone, smartphone, smartwatch. The device kept changing, but your number stayed put, and so did your name in other people's contacts. People could reach you because the system already knew who you were.
I think AI search will go the same way. The screens will keep changing, and the businesses that are known and easy to find will be the ones people can still reach. Start by asking ChatGPT what it knows about you.

Common questions

Does ChatGPT search the web every time?

No. In Ahrefs' study of 66.6 million ChatGPT answers from August and September 2026, only 14.4% ran a web search. The rest came from what the model learned in training. Ahrefs only looked at free, logged-out sessions, so other plans may search more or less often.

How do I know if ChatGPT knows about my business?

Open a Temporary Chat and choose Unpersonalized before your first message, so your own history doesn't color the answer, and ask what it knows about your business and what you sell. Ask it not to search the web, and check that the answer shows no sources or links. If it doesn't, you're looking at the model's memory of you. Check it against the facts, then ask Gemini the same thing. If it's thin or wrong, that's the side to work on.

Can I get my business into ChatGPT's training data?

There's no form to fill out. Models learn from text that was published before their training cutoff. What you can do is make sure your business is described clearly and the same way in the places that get read: your own site, directories, trade press, and the profiles you keep up.

What if ChatGPT describes my business wrong?

Then it likely learned something that's no longer true, like what you sold before you changed direction, and a new search may not override it. That's a different job from getting known in the first place. I'd focus on getting the correct description into the places a model reads, like your Google Business Profile, directory listings and trade press, and keeping it the same everywhere.

How long does it take ChatGPT to learn about changes to my business?

Memory only changes when a new model is trained, and companies don't say exactly what each one learned from. Search can pick up a change once your page has been crawled again. So you'd expect a fix to show up in search answers before memory answers, and in memory only if the new facts are repeated widely.

Does schema markup help with ChatGPT?

Not much on its own, going by Kurt Fischman's February 2026 paper, though pages with filled-in Product or Review details did better. I go through what Google, Microsoft and OpenAI each say about schema in my post on schema markup for AI search.

What's the difference between parametric knowledge and retrieval?

Parametric knowledge is what a model learned in training and keeps inside itself. It's the model's memory. Retrieval is when the model fetches outside information, like web pages, while it answers. ChatGPT uses both. In Ahrefs' study of free, logged-out chats, most answers came from memory.

Sources

Share this article
Add as a preferred source on Google

Topics

Ready to appear in AI search?

Plate Lunch Collective is an AI SEO agency that gets brands named when buyers ask, in search, on social, and in AI. If you have questions about where you stand, we're easy to reach.

Work with us

Hayden Bond

Hayden Bond has been doing SEO since 2004. He founded Plate Lunch Collective in Aiea, helping brands get cited by AI platforms rather than just ranked by Google.

Stay current on AI search

New research and analysis, direct to your inbox.

Related articles

Explore our AI SEO services

Previous

ChatGPT Brand Mentions: Why AI Cites Your Content and Recommends Your Competitor

Next

GEO vs SEO: What's different, from an agency that does both