ChatGPT draws on four information sources, and search is doing more of the work
ChatGPT's answers come from four places: its training data, live web search, saved memories and the user's approximate location. Which one carries the answer depends on the question. Simple definitions can be answered from training alone, while anything about prices, news or product recommendations sends the model to the web.
Training data is the slow, indirect route. OpenAI last published a detailed breakdown for GPT-3 in 2020 and now says only that its models were trained on diverse datasets. Sites that stay open to OpenAI's crawler may end up in a future model months later, but owners cannot see what was learned. GPTBot is the most-blocked AI crawler across roughly 140 million sites, so many publishers have opted out.
Search is the faster route and it is growing. ChatGPT's consumer product lead Adam Fry put its web searches at about one-fifth of Google's query volume, and OpenAI's own usage data shows 20.2% of consumer messages in September 2026 were people seeking information. The model rewrites prompts into several targeted searches, a process called query fan-out, and in a study of 66.6 million responses those that contained fan-out queries averaged about 1.6 each.
Citations are unstable. A four-week daily tracking of more than 1,300 prompts found only about a quarter of the pages cited for a prompt were cited again the next day, though the most consistently cited domains changed by about 14% day to day against 74% for everything else. For brands, that points to two jobs running in parallel: staying crawlable so future models absorb the site, and earning repeat citations in search results, where the same question rarely returns the same sources.
Reported by Ahrefs Blog ↗
Topics: ChatGPT, Citations, Measurement
In this directory: Ahrefs Brand Radar