The place Does ChatGPT Get Its Info? 4 Sources, Defined

Which one it makes use of relies on your query.

Earlier than a mannequin like those behind ChatGPT is launched, it is educated on an infinite quantity of data.

That coaching is the place ChatGPT’s basic information comes from: how language works, info concerning the world, write code, what issues seem like, and so forth.

The coaching occurs in levels: getting ready knowledge, pre-training, post-training, and ongoing analysis. (We go deeper on this in how reply engines work.) First comes pre-training on that vast combine of information. Then comes post-training, the place human trainers and researchers (and more and more the fashions themselves) train it to comply with directions, motive, and behave safely.

What particularly is on this combine? The final time OpenAI printed an in depth breakdown of its coaching knowledge was for GPT-3, again in 2020. That blend was:

Since then, OpenAI has shared much less and fewer. Its newest system card says solely that the fashions “have been educated on numerous datasets.” So no person exterior OpenAI is aware of precisely what’s in ChatGPT’s coaching knowledge right now.

What does that imply for you? In case your website is public and does not block OpenAI’s coaching crawler, there is a good probability your pages find yourself within the knowledge future fashions be taught from. That is how ChatGPT can “know” your model with out trying to find it. But it surely’s gradual and oblique: your content material solely reveals up within the subsequent mannequin, months later, and you’ll’t see what it discovered.

OpenAI doesn’t use every bit of data it may well collect. For instance, OpenAI says that it does not deliberately collect knowledge “from sources recognized to be behind paywalls or from the darkish internet,” and that it filters out materials like hate speech, grownup content material, websites that mixture private data, and spam.

OpenAI additionally says that they’re “more and more” utilizing artificial knowledge, coaching materials its personal fashions generate. OpenAI’s examples embrace apply prompts and examples in languages with little coaching textual content. The unique ChatGPT search mannequin was partly educated on solutions written by a stronger OpenAI mannequin (o1-preview).

Sure. Some website house owners don’t need their work used to coach AI free of charge, and massive publishers typically block it to maintain leverage in licensing offers. The truth is, GPTBot is the most-blocked AI crawler, in keeping with our research of ~140 million web sites.

The trade-off is that future fashions be taught much less about you from your personal website. If you’d like ChatGPT to know your model, you in all probability wish to go away it open.

If you happen to do wish to choose out, there are two methods:

The information cutoff is when a mannequin’s coaching knowledge stops. ChatGPT does not learn about something that occurred after that date until it searches the online.

For the fashions in ChatGPT right now, the cutoff is spring 2026, and it relies on your plan:

So for something since then, ChatGPT’s reply has to return from search, not from coaching.

Coaching knowledge is already old-fashioned when a mannequin is launched. GPT-6 Sol’s information stops on April 20, 2026, nevertheless it solely reached ChatGPT on October 7, virtually six months later. To reply questions on costs, information, merchandise, or something current, ChatGPT searches the online.

In response to OpenAI, ChatGPT might search routinely “when your query would profit from present data.” You can even make it search by selecting Search your self.

In my very own check of 12 questions in a logged-out ChatGPT, 9 triggered a search: present occasions, costs, product suggestions (“finest X in 2026”), and something the place we requested for sources. Solely easy definitions, like “what’s a backlink,” have been answered from coaching alone.

Search is now a giant a part of how folks use ChatGPT. In November 2026, ChatGPT’s shopper product lead, Adam Fry, informed Massive Expertise that ChatGPT’s internet searches add as much as roughly one-fifth of Google’s question quantity. He counted every query that sends ChatGPT to the online as one search. OpenAI’s personal utilization knowledge, OpenAI Alerts, reveals that about one in 5 shopper messages (20.2% in September 2026) are folks in search of data.

ChatGPT does not paste your query right into a search engine. OpenAI says it sometimes rewrites your immediate into a number of focused searches, then might ship extra particular searches to different suppliers to refine the outcomes.

This course of is named question fan-out: one immediate “followers out” into a number of searches, every protecting a unique a part of the query. The searches ChatGPT writes are referred to as fan-out queries. For instance, “finest CRM for a small company” would possibly fan out into “finest CRM for businesses 2026”, “HubSpot vs Pipedrive pricing”, and “CRM critiques small enterprise”.

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

A separate tutorial research (80 questions, run 20 instances every on GPT-5.3 and GPT-5.5 Prompt, March to July 2026) discovered that these fashions by no means searched the person’s precise wording. When questions have been requested in Chinese language, about 70% of the runs with seen searches have been searched solely in English. The identical research discovered that even Google’s outcomes and OpenAI’s personal search API mixed solely matched about 27% of the pages ChatGPT cited, and about half of the websites.

Your personal particulars can change the search too. OpenAI says ChatGPT can add particulars out of your saved recollections to its searches. For instance, if you happen to’ve shared that you simply’re vegan and reside in San Francisco, asking for close by eating places might produce a seek for “good vegan eating places San Francisco.” ChatGPT may also use your approximate location, based mostly in your IP deal with. It shares that basic location, however not the IP deal with itself, with search suppliers.

Lastly, the supply it selects will change, even if you happen to ask the identical query. This occurs as a result of ChatGPT is non-deterministic: the identical enter does not at all times produce the identical output.

For instance, my colleague Glen Allsopp tracked 1,300+ prompts each day for 4 weeks and located that solely a few quarter of the pages ChatGPT cited for a immediate have been cited once more the following day. The domains it cited most persistently have been a lot steadier, altering by about 14% from someday to the following, in opposition to 74% for every part else.

OpenAI says ChatGPT “ranks search outcomes utilizing a number of components,” and that placement is not assured. It does not listing the components.

The most important issue is what you ask. For this text, we pulled 4,000 US ChatGPT prompts from Ahrefs’ Model Radar in October 2026 and sorted each cited supply by sort.

Our analysis offers some extra clues:

That is the massive one. We all know ChatGPT will get plenty of its data from search outcomes, however whose? If it is Google’s, rating on Google is your finest guess. If it is Bing’s, you want to watch Bing too.

Nevertheless, in 2025, The Info reported that OpenAI had used a scraping service referred to as SerpApi to pull Google outcomes for real-time matters like information, sports activities, and finance. OpenAI hasn’t publicly confirmed it.

Do ChatGPT’s citations match what these search engines like google and yahoo rank? Our knowledge research present that the reply is usually no:

In different phrases, rating on Bing predicts ChatGPT’s citations no higher than rating on Google. No matter search outcomes ChatGPT will get again, it makes its personal decisions about which pages to quote.

Which means you’ll be able to’t inform whether or not ChatGPT cites you simply by taking a look at your Google or Bing rankings. To see that, you want to take a look at ChatGPT’s precise solutions with a software like Ahrefs’ Model Radar, which reveals which prompts point out and cite your website.

So one of the best reply to the query? Possible a mixture of every part: Bing and different suppliers, OpenAI’s personal index, and accomplice content material, adopted by ChatGPT’s personal decisions about what to quote.

When a query has purchasing intent, ChatGPT can present product playing cards with photos, costs, and hyperlinks to purchase. OpenAI’s purchasing assist web page says these merchandise are chosen utilizing “structured metadata from first-party and third-party suppliers (e.g., worth, product description) and different third-party content material,” plus your context, like reminiscence and customized directions.

OpenAI says product outcomes “are usually not adverts, nor influenced by any OpenAI partnerships.” Advertisements in ChatGPT are proven individually.

Since 2023, OpenAI has signed content material offers with greater than 20 publishers, together with the Related Press, Axel Springer, Condé Nast, Hearst, Future, Folks Inc. (previously Dotdash Meredith), Information Corp, Vox Media, The Washington Put up, The Guardian, and the Monetary Occasions.

Since June 2026, licensed Getty Photos images additionally seem in ChatGPT’s search outcomes, beneath a show deal. In November, media.gettyimages.com entered ChatGPT’s prime 10 most-cited domains for the primary time, at #9, in keeping with our listing of probably the most cited domains in ChatGPT:

Every deal works roughly the identical method: the writer licenses its content material to OpenAI, and ChatGPT reveals summaries of that content material with attribution and a hyperlink again to the unique article. For instance, OpenAI’s Grupo Folha and Grupo UOL announcement says their journalism will seem in ChatGPT with summaries, attribution, and hyperlinks to the unique articles.

Ten accomplice domains are among the many 50 websites ChatGPT cites most, and 47 are within the prime 500. Companion websites’ share of citations amongst ChatGPT’s 200 most-cited domains additionally greater than doubled between March and July 2026, from 7.5% to fifteen.8%, and was 12.5% in late September.

Curiously, the companions that do finest aren’t primarily information websites. They’re overview and life-style websites, like Vogue, Good Housekeeping, TechRadar, and extra.

As we are able to see from our knowledge, a deal does not assure visibility. Vogue was cited 42,108 instances; The Atlantic, additionally a accomplice, simply 724. As our evaluation places it: “A partnership makes a writer eligible to point out up in ChatGPT, nevertheless it does not assure it.”

There’s additionally some proof that accomplice content material reaches ChatGPT in a different way from the remainder of the online. David McSweeney, founding father of QueryBurst, discovered that some high-authority publications, together with Time, Enterprise Insider, and Forbes, have pre-written, page-level summaries that seem to go straight into ChatGPT’s context. Suganthan Mohanadasan noticed a separate retrieval supply in ChatGPT’s site visitors referred to as “labrador,” which he describes as a “licensed tier”.

Our fan-out knowledge factors the identical method. Companion websites are cited way more typically than ChatGPT searches for them by identify: of 17 companions, Reddit was the one one among the many 150 websites ChatGPT most frequently searched with website:.

Past its coaching and the online, ChatGPT makes use of no matter you give it:

Even your account issues. For instance, after I requested ChatGPT “what’s a backlink” in a logged-out, free account, it gave me a matter-of-fact reply.

Nevertheless, after I requested the identical query in my work ChatGPT account (a part of Ahrefs’ ChatGPT Enterprise workspace), it instantly talked about Ahrefs.

By the way in which, it doesn’t simply change the reply. It additionally adjustments what ChatGPT searches for. For instance, the query “What are one of the best AI visibility instruments” triggered a 26-website search in my Enterprise account:

The free, logged-out model didn’t even set off a search.

When somebody pastes a hyperlink into ChatGPT, ChatGPT visits that web page on their behalf. OpenAI’s crawler documentation says these user-triggered visits use a separate agent, ChatGPT-Person, not the search or coaching crawlers.

Now that we all know how ChatGPT will get its data, one factor stands out: every part it says about you is secondhand. It is what it reads in coaching, what it finds when it searches, and what a accomplice writer wrote.

Your web site is one supply of data, and even that solely counts if it made it into coaching or ChatGPT picks it from a search.

So, right here’s what meaning by way of actual motion steps:

To see how properly that is working in your model, Ahrefs’ Model Radar reveals how typically ChatGPT (and different AI search engines like google and yahoo) names and cites you throughout hundreds of prompts, and your AI share of voice in opposition to rivals.