Where ChatGPT Actually Gets Its Information
Contents
ChatGPT gets its information from two separate systems. One is a training corpus that was frozen on a fixed date, February 16, 2026, for the frontier models running in July 2026. The other is live web retrieval, which runs during the conversation itself, only when the question needs something newer than the training data can offer. Everything else about what ChatGPT knows, and when it knows it, follows from that split.
What Are the Two Places ChatGPT Gets Information From?
ChatGPT pulls from two places that work nothing alike: a frozen training corpus and live web retrieval. Confusing the two is the most common mistake in how people describe ChatGPT's knowledge.
The training corpus is text and patterns learned before the model shipped, compressed into its parameters. It doesn't update itself: a model trained through a given date knows nothing that happened after it, no matter how much time passes after release. Live retrieval works differently: the model reaches out to the web mid-conversation, through its own search system, pulling in pages that may not have existed during training.
The table below lines up both systems side by side, the same comparison behind the "chart" many people are searching for on this exact question.
| Training corpus | Live retrieval | |
|---|---|---|
| When it happens | Learned before the model ships, during training | During the conversation, when the model needs something newer |
| What it holds | Text and patterns baked into the model's parameters | Pages fetched from the live web at query time |
| Is it current? | No, frozen at a fixed cutoff date | Yes, as current as whatever page gets fetched |
| Which OpenAI crawler | GPTBot | OAI-SearchBot, plus ChatGPT-User for user-triggered actions |
| How you get in | Public web content, third-party data, and trainer input from before the cutoff | Any public page OAI-SearchBot can reach at query time |
| How fast it changes | Only when a new model trains and ships | Every single conversation, in real time |
What Data Was ChatGPT Actually Trained On?
OpenAI names three sources for how its models are developed: "developed using three primary sources of information: (1) information that is publicly available on the internet, (2) information that we partner with third parties to access, and (3) information that our users, human trainers, and researchers provide or generate" (source: OpenAI Help Center, help.openai.com/en/articles/7842364, 2026).
The only training mix OpenAI has ever broken down in detail belongs to GPT-3, a 175-billion-parameter model described in "Language Models are Few-Shot Learners" (Brown et al., arXiv:2005.14165, published May 28, 2020). Filtered Common Crawl made up 60% of the mix (410 billion tokens), WebText2 added 22% (19 billion tokens), Books1 and Books2 contributed 8% each (12 billion and 55 billion tokens), and Wikipedia rounded it out at 3% (3 billion tokens). That Common Crawl data started as 41 monthly shards covering 2016 to 2019, 45 terabytes of compressed text before filtering, cut down to 570 gigabytes after (source: arXiv:2005.14165, Section 2.2).
Raw volume did not equal weight, either. OpenAI's paper states that "datasets we view as higher-quality are sampled more frequently, such that CommonCrawl and Books2 datasets are sampled less than once during training, but the other datasets are sampled 2-3 times" (arXiv:2005.14165, Section 2.2). The largest raw pool was also the most discounted one.
Why the Numbers You See Quoted Are From 2020
That 60/22/8/8/3 breakdown, recycled across most explanations of this topic, describes GPT-3's training mix from a paper published in May 2020, months before ChatGPT existed. It doesn't describe any model running today, and the mix behind OpenAI's current frontier models hasn't been published at that level of detail. That gap matters more now, given how much low-quality text is now in the pool models learn from.
Does ChatGPT Search the Internet, and What Is Its Knowledge Cutoff in 2026?
Yes, and it has for everyone since February 5, 2025. The training-corpus cutoff still matters, since live retrieval only kicks in when the model decides a question needs something newer than what it knows.
The frontier models running in July 2026, GPT-5.6 Sol, Terra, and Luna, share a knowledge cutoff of February 16, 2026 (source: OpenAI, developers.openai.com/api/docs/models/compare, 2026). A cutoff date isn't the same as a release date: it marks the last point the training data covers, not when the model became available.
Two claims still show up on pages ranking for this exact question, and both are wrong in 2026. "The dataset only went up to 2021" describes an old model generation, not anything running today. And "the standard version of ChatGPT cannot access the internet in real time" stopped being true once search rolled out to every signed-in user on December 16, 2024, then to everyone, including logged-out users, on February 5, 2025 (source: OpenAI, openai.com/index/introducing-chatgpt-search/; help.openai.com/en/articles/9237897).
Why ChatGPT Cites Different Sources for the Same Question
Retrieval isn't a single, stable pipeline. An analysis of 9,946 completed ChatGPT search runs, drawn from 1,000 prompts tested up to 10 times each and published by Search Engine Land in July 2026, found one retrieval back end handling 88.1% of primary sources, with the other three splitting 9.9%, 1.7%, and 0.3%. In 11.6% of prompts, the primary source changed between runs of the same question, and when it did, URL overlap dropped from 0.273 to 0.149, a 45% drop, while domain overlap fell from 0.265 to 0.155, a 42% drop.
That volatility is the practical reason a single citation shouldn't be treated as a locked-in placement, the same caution worth applying to any answers that never send a click back.
This article, Blue could have written it for you: content optimized for Google + AI, without you writing a single word.
Try mentionLABHow Does ChatGPT Decide Which Websites to Pull From?
Any public website can appear. OpenAI states plainly that "any public website can appear in ChatGPT search" (source: OpenAI, help.openai.com/en/articles/12627856, 2026). Ranking among those sites "is based on a number of factors designed to help users find reliable, relevant information," and OpenAI is explicit that "there is no way to guarantee top placement" (same source).
Getting into consideration comes down to two technical conditions: letting OpenAI's search crawler reach the site, and allowing OpenAI's published IP ranges through. Miss either one and the content simply isn't eligible.
The Three Crawlers, and Which One You Actually Need
OpenAI runs three separate bots, and mixing them up is the costliest mistake here. GPTBot crawls content "that may be used in training" OpenAI's models; disallowing it opts a site out of training only (source: developers.openai.com/api/docs/bots, 2026). OAI-SearchBot "surfaces websites in search results in ChatGPT's search features" and is explicitly "not used to crawl content to train" those models (source: help.openai.com/en/articles/12627856); opting out means a site "will not be shown in ChatGPT search answers." ChatGPT-User fires only when a user's own action triggers a fetch, and OpenAI says it "is not used for crawling the web in an automatic fashion" (source: developers.openai.com/api/docs/bots).
The point almost none of these pages make: blocking GPTBot doesn't remove a site from ChatGPT's answers, it only opts content out of training. Blocking OAI-SearchBot does the removing, since it's the only one of the three built to power citations. A full breakdown of which AI crawlers to allow and which to block and how AI models decide which sources to trust is worth reading before touching a robots.txt file.
How Do You Become One of the Sources ChatGPT Uses?
You can't get into the training corpus on demand, so the door that's actually open is retrieval, and it opens this week. Training and retrieval run on opposite calendars: training is frozen, retroactive, and out of a site's control until the next model trains and ships, a cycle measured in years. Retrieval is live right now, reversible, and directly affected by decisions a site makes today.
What makes a page retrievable and citable comes down to a short list: answering the question in the first sentence instead of building up to it, using facts that are verifiable and dated rather than vague, structuring the page so a machine can parse it cleanly, and addressing the question the way it's actually asked rather than a nearby topic.
We build MentionLab's own articles around that list: sourced to primary documentation, dated, and structured for clean retrieval, then we track which pieces get picked up in ChatGPT, Perplexity, and Claude. That's a first-hand view of how uneven the pickup is even with solid sourcing.
The retrieval side splits into platform-specific and general work. Optimizing a site for ChatGPT specifically covers what changes when OAI-SearchBot, not Google, is the audience. The broader practice of getting cited by AI systems applies the same principles across ChatGPT, Perplexity, and Google's AI Overviews at once. At the sentence level, that starts with putting the answer in the first sentence, and at the page level, FAQs written to land in AI answers give a retrieval system exactly the question-and-answer pairs it's built to lift.
Frequently Asked Questions
Who gives ChatGPT information?
OpenAI names three sources for developing its models: information publicly available on the internet, information from third-party partnerships, and information from users, human trainers, and researchers (source: OpenAI Help Center, 2026). On top of that fixed corpus, live web retrieval adds whatever OAI-SearchBot pulls in during the conversation itself.
Does ChatGPT get its information from Google?
No. ChatGPT doesn't query Google's search index. It runs its own retrieval system, built around OAI-SearchBot and OpenAI's own back ends, separate from Google entirely. That's different from Google's AI Overviews, which sit inside Google Search itself; see how AI Overviews work for that comparison.
Does ChatGPT get information from the internet?
Yes, in two distinct ways. Part of its training corpus is web content that was publicly available before the cutoff, and separately, live web retrieval fetches current pages during the conversation itself, a capability every user has had, signed in or not, since February 5, 2025 (source: OpenAI, 2025).
What data is ChatGPT trained on?
OpenAI names three categories: publicly available internet content, data from third-party partnerships, and material from users, human trainers, and researchers. The only training mix ever published in real detail belongs to GPT-3, from May 2020: filtered Common Crawl, WebText2, two book datasets, and Wikipedia. The mix behind today's models hasn't been made public.
Where does ChatGPT get its information about me?
Three places: whatever was publicly available about you before the cutoff, whatever it finds through live retrieval during the conversation, and whatever you tell it directly in the chat. ChatGPT doesn't maintain a separate profile or database about individual users outside of those three sources.
Blue handles your SEO and your GEO. On autopilot.
You approve, she produces content optimized for Google + AI.
Join the Lab · 5-day trial