Where Does ChatGPT Get Its Information?
Where does ChatGPT get its information? A 2026 breakdown of its training data, knowledge cutoff, live web search, and where LLMs really pull answers from.
ChatGPT gets its information from two places: training and live search
Ask where ChatGPT gets its information and the honest answer is that it comes from two places that work nothing alike. Most of what it knows came from an enormous batch of text it was trained on once, then frozen in place. The rest, when you have search switched on, it goes and fetches from the live web in the moment you ask. That is the whole thing. Everything below is those two layers and, more usefully, what they mean if you are trying to get your brand into the answer.
The distinction matters because people ask the question expecting one answer and there are two. “Trained on the internet” and “reads the internet right now” are not the same claim, and confusing them is how you end up trusting a confident answer about something that happened last week.
What ChatGPT is trained on, and why it has a knowledge cutoff
The training corpus is the big one, the source of most of what ChatGPT can say without looking anything up. It is assembled from four buckets: a broad crawl of the public web (the Common Crawl dataset and similar), licensed data bought through deals, digitized books, and code. Reference-grade sources like Wikipedia are weighted heavily because they are clean and structured.
The licensing piece is the part most people miss, and it is public. OpenAI has signed content deals with Reddit and with a long list of publishers, the Associated Press, Axel Springer, News Corp, the Financial Times, Condé Nast, and others, to train on and surface their material. So “who gives ChatGPT information” is partly answered by a contract. Some of the internet was crawled for free, and some of it was paid for.
Here is the catch, and it answers a question people search constantly. That corpus is a snapshot, not a live feed. It was collected up to a fixed date, the knowledge cutoff, and then frozen. Ask ChatGPT about anything after its cutoff and, with no browsing, it cannot know. It will often answer anyway, confidently, which is exactly the failure mode to watch for. The training data is why ChatGPT sounds like it knows everything. The cutoff is why it is sometimes wrong about the recent past and does not flinch.
When ChatGPT searches the web, it uses Bing, not Google
Second layer. When search is on, ChatGPT does not rely on the frozen corpus. It retrieves live results from the web, reads them, and writes an answer grounded in what it just pulled. This is how it answers questions about today.
One detail trips people up: it is not Google. ChatGPT’s live search runs on Bing’s index and a set of third-party search providers, because of the Microsoft partnership behind it. So when someone asks whether ChatGPT gets its information from Google, the answer is no. It sees a Bing-shaped view of the web, which sometimes surfaces different pages than a Google search for the same query would.
Training data tells you what ChatGPT remembers. Live search tells you what it can look up. If you want to influence the answer, you have to win in both.
Where LLMs actually pull their answers from
This is the part that matters for anyone doing marketing, and it is where the generic explainers stop. So I pulled the real numbers. I track which domains get cited in AI answers across the prompts that matter for six B2B SaaS accounts I run, spanning cybersecurity, legal, HR tech, observability, marketing tech, and creator tools.
The first surprise is how little any single source matters. Those answers were assembled from more than 76,000 distinct domains, and the most-cited platform of them all, YouTube, accounts for just 2.7% of citations on average. The top ten sources combined are under 12%. There is no one page to win here. The answer is stitched together from a very long tail.
The second surprise is how consistent the top of that tail is. Strip out the industry-specific sites and the same recognizable platforms show up in every one of the six verticals:
| Platform | What it is | Avg share of a brand’s citations |
|---|---|---|
| YouTube | Video: how-tos, demos, reviews | 2.7% |
| Community threads and opinions | 2.5% | |
| Professional social | 1.0% | |
| G2, Capterra, GetApp, TrustRadius | Software review directories | ~0.6% each |
| Gartner | Analyst coverage | 0.6% |
| Microsoft, IBM | Big-tech docs and thought leadership | ~0.5% each |
| Medium, Forbes, TechRadar | Editorial and trade press | ~0.4% each |
| Wikipedia | Reference | in all six |
Video, community, review directories, analysts, reference. That is the independent layer AI reaches for first, and it is remarkably stable across industries a buyer would swear have nothing in common.
Then it gets uncomfortable. Your competitors’ own sites usually out-cite yours. In the legal account, three competitor platforms each individually beat the client’s own domain as a source. In cybersecurity, the two largest competitors did, and the brand’s own site limped in ninth. Even in the one vertical where the brand was the single most-cited domain about itself, that was still only 4% of the picture. The other 96% was everyone else.
The AI is building the answer about your category more from your competitors’ sites and a handful of review platforms than from anything you have ever published.
And the long tail is not all quality. The same auto-generated “statistics” sites, the kind that spin up a page of invented percentages for any topic you name, showed up as cited sources across all six industries. The models are pulling from junk right next to Gartner. Read that as a warning or an opening, depending on how cynical you are.
Which is the whole problem in a line: you can be crawled, cited, and still absent from the recommendation, a separate issue I have written about in AI citations are a vanity metric.
What this means for B2B SaaS
Put the two layers together and the playbook writes itself. You cannot edit ChatGPT’s training data, and you do not control Bing’s index. What you can control is whether the sources those systems trust already name you.
For B2B SaaS that reframes the whole job. Polishing your own product pages helps you get retrieved, but the answer to “best project management tool” or “top expense management platform” is assembled from the third-party sources the model reaches for first. If the independent roundups, the Reddit threads, and the trade coverage name a competitor, that is who gets recommended, no matter how good your own page is. There is a hard number on this: when an AI answer cites a brand’s own “best” listicle, it recommends a competitor 69% of the time (Lily Ray, via Search Engine Land). Your own listicle can lose you the deal.
So the work is off your site as much as on it: get named in the roundups that rank, earn the community and PR mentions the model treats as evidence, and publish first-party data and a point of view worth citing. That is the same muscle behind how to rank in AI Overviews and the broader generative engine optimization playbook, and it is why the self-promotional listicle is a dead format. You do not win the AI answer by talking about yourself. You win it by being the brand everyone else’s sources already talk about.
FAQ
Does ChatGPT get its information from Google?
No. When ChatGPT searches the web, its live results come from Bing’s index and third-party search providers, not Google, because of the Microsoft partnership behind it. That is a real difference: a Bing-shaped view of the web can surface different pages than Google would for the same query. Its training data, separately, was crawled from the open web rather than pulled from any single search engine.
Who gives ChatGPT information?
Three groups, effectively. The open web, through broad crawls like Common Crawl. Licensed partners, through paid content deals (Reddit and publishers such as the Associated Press, Axel Springer, News Corp, and the Financial Times). And, at answer time, whatever live pages its Bing-powered search retrieves for your specific question. Plus you: anything you type or upload into a chat becomes information for that conversation.
Does ChatGPT get information from the internet in real time?
Only when search is on. With browsing enabled it retrieves live pages and answers from what it just read. With browsing off it works entirely from its frozen training data, which stops at the knowledge cutoff. If a question depends on recent events, an answer with no browsing is working from memory, not the live internet.
What data is ChatGPT trained on?
A mix of public web text (crawled datasets like Common Crawl), licensed data from content deals, digitized books, and code, with reference sources like Wikipedia weighted heavily for being clean and structured. It is a fixed snapshot collected up to the model’s cutoff date, not a continuously updating feed.
Does ChatGPT own the information you put into it?
You keep ownership of what you enter. Depending on your settings and plan, though, your conversations can be used to improve future models, which is why you can turn that off and why business and enterprise tiers exclude your data from training by default. For anything sensitive, treat consumer ChatGPT as a place whose inputs may be seen, and check the data controls for your account.