Where Does ChatGPT Get Its Data? Training Data and Sources

ChatGPT uses publicly available internet content and licensed datasets created or reviewed by human trainers. GPT-4, OpenAI's flagship large multimodal model, was reportedly trained on roughly 13 trillion tokens of text and code (The Decoder, July 2023). To put that in perspective, if you read non-stop at 250 words per minute, it would take you over 70,000 years to get through it all. And that's just one model from one company.

So where does ChatGPT actually get its data? If you're a business owner or marketer trying to understand why AI recommends certain brands and ignores others, the training data is where the story starts. If your brand isn't represented in those training sources, you're essentially invisible to one of the fastest-growing discovery channels in the world.

Why Training Data Matters for Your Brand

ChatGPT's training data determines which brands, products, and ideas the model "knows" about. If your business isn't well-represented across the trusted sources that feed into LLM training, AI tools are unlikely to recommend you, no matter how good your product is.

Research from Seer Interactive (2025) shows that LLMs tend to reflect information appearing consistently across multiple independent sources. A single glowing review on your own website carries little influence. What matters is whether your brand appears across third-party directories, publications, forums, and review platforms, the same types of content that form the backbone of LLM training datasets. This is the core of answer engine optimization (AEO): getting into the answers that ChatGPT, Gemini, and Perplexity generate, not just ranking on Google.

The Five Layers of ChatGPT's Knowledge

Pre-training is the model's education. Common Crawl, a non-profit archive of the public web, contributes over 250 billion pages totaling roughly 468 terabytes. OpenAI also draws from digitized books, academic papers, and GitHub code repositories. The original ChatGPT model was trained on approximately 570GB of text, though newer versions use vastly more.

Fine-tuning uses much smaller, curated datasets to teach the model how to follow instructions and format responses. The goal isn't to add new knowledge; it's to turn a raw text predictor into something useful. RLHF then layers human judgment on top: reviewers rank responses by accuracy, clarity, and safety, and the model learns to favor those qualities. Neither stage adds new facts; they shape how the model communicates.

Web search is available on every ChatGPT plan, including free, and uses retrieval-augmented generation (RAG): the model decides when a question needs fresh information, searches the web, reads the results, and cites them in its answer. This is the layer that decides which sites get linked, and it can change much faster than training data. System prompts are hidden instructions at the start of each conversation that shape the model's behavior, allowing developers to specialize it for specific tasks.

How Big Is ChatGPT's Training Dataset?

GPT-4 was trained on trillions of tokens drawn from web crawls, books, and code. Training it cost over $100 million, according to OpenAI CEO Sam Altman. Common Crawl's raw archive contains at least 100 trillion tokens, though heavy filtering brings usable data down significantly: NVIDIA's filtered version holds 6.3 trillion tokens; HuggingFace's FineWeb holds 15 trillion. The model itself has an estimated 1.76 trillion parameters in a Mixture of Experts architecture, where only a fraction activate for any given query.

What ChatGPT Doesn't Know

Despite its massive training set, ChatGPT has significant blind spots. It has no access to real-time information unless it searches the web, cannot read paywalled or private content, and has a hard knowledge cutoff beyond which it has no training data.

Each model has its own fixed cutoff, and newer models move it later. ChatGPT 5.2 (released December 2025), for example, has a knowledge cutoff of August 31, 2025 (ALLMO, 2026), while GPT-4o's training stopped in October 2023. When gaps exist, ChatGPT doesn't always say "I don't know." As researchers at the NIH noted, the model "predicts the next token, not the next fact" (PMC, 2024), so it can fill gaps with plausible-sounding but fabricated information.

The Copyright Question: Who Owns Training Data?

The legal ownership of AI training data is one of the most contested questions in technology right now. As of October 2025, there were 51 active copyright lawsuits against AI companies, and no court is expected to rule on the core fair use question before summer 2026 (ChatGPT Is Eating the World, October 2025).

The highest-profile case is The New York Times v. OpenAI, filed in December 2023. The Times alleges OpenAI used millions of its copyrighted articles to train GPT models without permission. In January 2026, a federal judge ordered OpenAI to produce 20 million anonymized ChatGPT logs as part of discovery.

OpenAI also signed a deal with Stack Overflow providing access to 15 years of developer Q&A data. Not everyone was happy: some Stack Overflow users deleted their top-rated answers in protest, only to have those posts restored and their accounts banned. For businesses thinking about LLM visibility, these deals signal a clear trend: the era of scraping everything freely is closing, and which content makes it into future training datasets will increasingly depend on licensing arrangements.

ChatGPT Now Rarely Uses Reddit as a Source

ChatGPT now rarely uses Reddit as a source in its live answers. In Searchable data on the same prompts, the share of ChatGPT answers with a Reddit source fell from 24.1% (July 20 to August 9, 2026) to 1.9% (August 17 to October 5, 2026). That matters because live sources, unlike training data, decide which sites get linked in an answer.

When ChatGPT searches the web, it picks a fresh set of sources for each answer, and that mix can change within days. Before the change, about one in four ChatGPT answers that listed sources included a Reddit page, and afterward it was roughly one in 50. The decline started around August 8 and was complete by August 14.

The change affected almost every brand in the sample. Of 283 brands whose ChatGPT answers regularly included Reddit sources before August, 275 saw that share fall by at least half, and none saw it rise. Over the same period, LinkedIn's share of ChatGPT answers barely moved.

Other trackers reported the same drop in the same week. OpenAI has said it doesn't set a fixed visibility level for any site and that ChatGPT still cites Reddit, and our data agrees that Reddit is still in ChatGPT's answers, though now in roughly one answer in 50.

Training data and live sources are separate layers. OpenAI's licensing deal gives it access to Reddit content, which can shape what the model learns, but that doesn't mean ChatGPT will link to a Reddit thread when a buyer asks a question today.

If your plan for showing up in ChatGPT relied on Reddit threads being cited, that route is much narrower than it was in July. Reddit is still worth your time where your buyers ask questions, because forums still feed training data and other AI engines, though it works better as one channel among several than as your main way into ChatGPT's answers.

Source: Searchable, July 20 to October 5, 2026. ChatGPT answers collected from the ChatGPT app for prompts that brands track in Searchable, comparing the same prompts in both periods (about 60,000 prompts; 451,494 answers before and 282,151 after, counting answers that listed at least one source). An answer counts when ChatGPT lists a reddit.com page among its sources.

How Training Data Shapes AI Search Recommendations

LLMs don't evaluate sources the way humans do. They look for patterns and consensus across their training data. When multiple independent, authoritative sources mention your brand in a positive context, the model learns that association.

This is the core principle behind AI search optimization. It's not about gaming the system. It's about making sure your brand is genuinely well-represented across the open web, in the kinds of places that training data crawlers actually index.

Common Misconceptions About ChatGPT's Data

FAQs

Conclusion

Where ChatGPT gets its data, from Common Crawl's 250 billion web pages to licensing deals with Reddit and Stack Overflow, directly shapes its answers to millions of user queries every day. What it cites in live answers is a separate layer, and as the August 2026 Reddit drop showed, that mix can change in a week. For businesses, this creates both a risk and an opportunity. If your brand isn't well-represented across the open, crawlable web, you're leaving your AI visibility to chance.

The brands that show up in ChatGPT's recommendations aren't there by accident. They've built a consistent presence across trusted, independent sources: industry publications, review platforms, forums, and structured directories. Understanding where the data comes from is the first step. Making sure your brand appears in those sources is what actually moves the needle.