Where Does ChatGPT Get Its Information? 4 Sources, Explained

Which one it uses depends on your question.

Before a model like the ones behind ChatGPT is released, it's trained on an enormous amount of information.

That training is where ChatGPT's general knowledge comes from: how language works, facts about the world, how to write code, what things look like, and so on.

The training happens in stages: preparing data, pre-training, post-training, and ongoing evaluation. (We go deeper on this in how answer engines work.) First comes pre-training on that huge mix of data. Then comes post-training, where human trainers and researchers (and increasingly the models themselves) teach it to follow instructions, reason, and behave safely.

According to OpenAI, its models are developed using three sources of training data:

What specifically is in this mix? The last time OpenAI published a detailed breakdown of its training data was for GPT-3, back in 2020. That mix was:

Since then, OpenAI has shared less and less. Its latest system card says only that the models "were trained on diverse datasets." So nobody outside OpenAI knows exactly what's in ChatGPT's training data today.

What does that mean for you? If your site is public and doesn't block OpenAI's training crawler, there's a good chance your pages end up in the data future models learn from. That's how ChatGPT can "know" your brand without searching for it. But it's slow and indirect: your content only shows up in the next model, months later, and you can't see what it learned.

OpenAI doesn’t use every piece of information it can gather. For example, OpenAI says that it doesn't intentionally gather data "from sources known to be behind paywalls or from the dark web," and that it filters out material like hate speech, adult content, sites that aggregate personal information, and spam.

OpenAI also says that they’re “increasingly" using synthetic data, training material its own models generate. OpenAI's examples include practice prompts and examples in languages with little training text. The original ChatGPT search model was partly trained on answers written by a stronger OpenAI model (o1-preview).

Yes. Some site owners don't want their work used to train AI for free, and big publishers often block it to keep leverage in licensing deals. In fact, GPTBot is the most-blocked AI crawler, according to our study of ~140 million websites.

The trade-off is that future models learn less about you from your own site. If you want ChatGPT to know your brand, you probably want to leave it open.

If you do want to opt out, there are two ways:

The knowledge cutoff is when a model’s training data stops. ChatGPT doesn't know about anything that happened after that date unless it searches the web.

For the models in ChatGPT today, the cutoff is spring 2026, and it depends on your plan:

So for anything since then, ChatGPT's answer has to come from search, not from training.

Training data is already out of date when a model is released. GPT-6 Sol's knowledge stops on April 20, 2026, but it only reached ChatGPT on October 7, almost six months later. To answer questions about prices, news, products, or anything recent, ChatGPT searches the web.

This is known as retrieval-augmented generation (RAG). The model retrieves information first, then generates its answer from it.

According to OpenAI, ChatGPT may search automatically "when your question would benefit from current information." You can also make it search by choosing Search yourself.

In my own test of 12 questions in a logged-out ChatGPT, 9 triggered a search: current events, prices, product recommendations ("best X in 2026"), and anything where we asked for sources. Only simple definitions, like "what is a backlink," were answered from training alone.

Search is now a big part of how people use ChatGPT. In November 2026, ChatGPT's consumer product lead, Adam Fry, told Big Technology that ChatGPT's web searches add up to roughly one-fifth of Google's query volume. He counted each question that sends ChatGPT to the web as one search. OpenAI's own usage data, OpenAI Signals, shows that about one in five consumer messages (20.2% in September 2026) are people looking for information.

ChatGPT doesn't paste your question into a search engine. OpenAI says it typically rewrites your prompt into one or more targeted searches, then may send more specific searches to other providers to refine the results.

This process is called query fan-out: one prompt "fans out" into several searches, each covering a different part of the question. The searches ChatGPT writes are called fan-out queries. For example, "best CRM for a small agency" might fan out into "best CRM for agencies 2026", "HubSpot vs Pipedrive pricing", and "CRM reviews small business".

In our ChatGPT query fan-out study of 66.6 million responses from free, logged-out ChatGPT (August to September 2026), responses that contained fan-out queries had about 1.6 each.

Some patterns stood out:

A separate academic study (80 questions, run 20 times each on GPT-5.3 and GPT-5.5 Instant, March to July 2026) found that those models never searched the user's exact wording. When questions were asked in Chinese, about 70% of the runs with visible searches were searched only in English. The same study found that even Google's results and OpenAI's own search API combined only matched about 27% of the pages ChatGPT cited, and about half of the sites.

Your own details can change the search too. OpenAI says ChatGPT can add details from your saved memories to its searches. For example, if you've shared that you're vegan and live in San Francisco, asking for nearby restaurants may produce a search for "good vegan restaurants San Francisco." ChatGPT can also use your approximate location, based on your IP address. It shares that general location, but not the IP address itself, with search providers.

Finally, the source it selects will change, even if you ask the same question. This happens because ChatGPT is non-deterministic: the same input doesn't always produce the same output.

For example, my colleague Glen Allsopp tracked 1,300+ prompts daily for four weeks and found that only about a quarter of the pages ChatGPT cited for a prompt were cited again the next day. The domains it cited most consistently were much steadier, changing by about 14% from one day to the next, against 74% for everything else.

OpenAI says ChatGPT "ranks search results using multiple factors," and that placement isn't guaranteed. It doesn't list the factors.

The biggest factor is what you ask. For this article, we pulled 4,000 US ChatGPT prompts from Ahrefs' Brand Radar in October 2026 and sorted every cited source by type.

Our research gives some more clues:

This is the big one. We know ChatGPT gets a lot of its information from search results, but whose? If it's Google's, ranking on Google is your best bet. If it's Bing's, you need to watch Bing too.

Here's what OpenAI says:

However, in 2025, The Information reported that OpenAI had used a scraping service called SerpApi to pull Google results for real-time topics like news, sports, and finance. OpenAI hasn't publicly confirmed it.

Do ChatGPT's citations match what those search engines rank? Our data studies show that the answer is mostly no:

In other words, ranking on Bing predicts ChatGPT's citations no better than ranking on Google. Whatever search results ChatGPT gets back, it makes its own choices about which pages to cite.

That means you can't tell whether ChatGPT cites you just by looking at your Google or Bing rankings. To see that, you need to look at ChatGPT's actual answers with a tool like Ahrefs' Brand Radar, which shows which prompts mention and cite your site.

So the best answer to the question? Likely a mix of everything: Bing and other providers, OpenAI's own index, and partner content, followed by ChatGPT's own choices about what to cite.

When a question has shopping intent, ChatGPT can show product cards with images, prices, and links to buy. OpenAI's shopping help page says these products are chosen using "structured metadata from first-party and third-party providers (e.g., price, product description) and other third-party content," plus your context, like memory and custom instructions.

OpenAI says product results "are not ads, nor influenced by any OpenAI partnerships." Ads in ChatGPT are shown separately.

Since 2023, OpenAI has signed content deals with more than 20 publishers, including the Associated Press, Axel Springer, Condé Nast, Hearst, Future, People Inc. (formerly Dotdash Meredith), News Corp, Vox Media, The Washington Post, The Guardian, and the Financial Times.

Since June 2026, licensed Getty Images photos also appear in ChatGPT's search results, under a display deal. In November, media.gettyimages.com entered ChatGPT's top 10 most-cited domains for the first time, at #9, according to our list of the most cited domains in ChatGPT:

Each deal works roughly the same way: the publisher licenses its content to OpenAI, and ChatGPT shows summaries of that content with attribution and a link back to the original article. For example, OpenAI's Grupo Folha and Grupo UOL announcement says their journalism will appear in ChatGPT with summaries, attribution, and links to the original articles.

We checked 85 partner domains in Ahrefs’ Brand Radar for our analysis of OpenAI's publisher partnerships.

Ten partner domains are among the 50 sites ChatGPT cites most, and 47 are in the top 500. Partner sites' share of citations among ChatGPT's 200 most-cited domains also more than doubled between March and July 2026, from 7.5% to 15.8%, and was 12.5% in late September.

Interestingly, the partners that do best aren't mainly news sites. They're review and lifestyle sites, like Vogue, Good Housekeeping, TechRadar, and more.

As we can see from our data, a deal doesn't guarantee visibility. Vogue was cited 42,108 times; The Atlantic, also a partner, just 724. As our analysis puts it: "A partnership makes a publisher eligible to show up in ChatGPT, but it doesn't guarantee it."

There's also some evidence that partner content reaches ChatGPT differently from the rest of the web. David McSweeney, founder of QueryBurst, found that some high-authority publications, including Time, Business Insider, and Forbes, have pre-written, page-level summaries that appear to go straight into ChatGPT's context. Suganthan Mohanadasan spotted a separate retrieval source in ChatGPT's traffic called "labrador," which he describes as a "licensed tier".

Our fan-out data points the same way. Partner sites are cited far more often than ChatGPT searches for them by name: of 17 partners, Reddit was the only one among the 150 sites ChatGPT most often searched with site:.

Beyond its training and the web, ChatGPT uses whatever you give it:

Even your account matters. For example, when I asked ChatGPT "what is a backlink" in a logged-out, free account, it gave me a matter-of-fact answer.

However, when I asked the same question in my work ChatGPT account (part of Ahrefs' ChatGPT Business workspace), it suddenly mentioned Ahrefs.

By the way, it doesn’t just change the answer. It also changes what ChatGPT searches for. For example, the question “What are the best AI visibility tools” triggered a 26-website search in my Business account:

The free, logged-out version didn’t even trigger a search.

When someone pastes a link into ChatGPT, ChatGPT visits that page on their behalf. OpenAI's crawler documentation says these user-triggered visits use a separate agent, ChatGPT-User, not the search or training crawlers.

Now that we know how ChatGPT gets its information, one thing stands out: everything it says about you is secondhand. It's what it reads in training, what it finds when it searches, and what a partner publisher wrote.

Your website is one source of information, and even that only counts if it made it into training or ChatGPT picks it from a search.

So, here’s what that means in terms of real action steps:

To see how well this is working for your brand, Ahrefs’ Brand Radar shows how often ChatGPT (and other AI search engines) names and cites you across thousands of prompts, and your AI share of voice against competitors.