Many people assume ChatGPT searches the web when they ask it something. Usually it doesn't. Ahrefs analyzed 66.6 million free, logged-out GPT-5.6 Luna responses and their 15 million fan-out queries from August and September. 14.4% of the 66.6 million responses contained fan-out queries. The other six in seven answers came from what the model had already learned during training. That changes how much you should trust an answer, when you should ask for a search, and what publishers should focus on.

What the study actually measured

On October 12, 2026, the SEO analysis tool Ahrefs published a large-scale study analyzing approximately 66.6 million ChatGPT responses. It counted how often ChatGPT sent background searches to the web before answering. These background searches are called "fan-out queries."

One detail is easy to miss. Ahrefs changed its own wording after publishing. They originally said 14.4% of responses searched the web. More accurately, 14.4% contained fan-out queries. In practice the two are close, but it's the more careful way to quote the number.

The sample also has limits. It covered free, logged-out users on one model over two months. Ahrefs notes that paid and reasoning models search more, so its 14.4% search rate and 1.6 fan-outs per searching response are not a description of every ChatGPT user. Treat it as a minimum. On a paid plan or with a reasoning model, you'll likely see searches more often. One market-commentary write-up put it this way: "Treat 14.4% as a snapshot that should inform testing, not as a fixed property of ChatGPT."

Training data vs. live retrieval, in plain terms

A chatbot like ChatGPT can answer a question in two ways:

  • From memory (training data). The model learned patterns from a huge body of text up to a cutoff date. When it answers from memory, nothing new is looked up. It produces the most likely answer based on what it absorbed. That works well for stable topics like how photosynthesis works or what a pivot table does. It's risky for prices, laws, product features, or anything that has changed recently.
  • From retrieval (web search). The model writes one or more search queries, reads the results, and builds its answer from those pages. This is where citations and source links usually come from.

When ChatGPT does search, it doesn't do much of it. Those responses had an average of about 1.6 fan-out queries each. With about 1.6 queries, your page has to match one of those one or two searches. That isn't a deep research pass. It's a quick lookup or two before writing the answer.

How to tell which kind of answer you got

The most obvious sign is visible sources. If the answer shows source links or a "searched the web" indicator, retrieval probably happened. If it gives a confident paragraph with no sources, assume it came from memory.

Even that rule has an exception. Search Engine Journal, citing Graphite's research, says a citation alone does not necessarily indicate a search occurred. In an earlier paper by Graphite regarding prompt tracking, it noted that as of May, ChatGPT could add inline links without displaying any evidence of a search. A link in the answer doesn't prove the model read that page today.

Some practical habits:

  1. Ask for a search on anything time-sensitive. Pricing, plan tiers, release dates, regulations, and "latest" anything should come from a live search. Say so directly in your prompt, or use the search option in the interface.
  2. Ask "what's your source and date for that?" If the model can't name a specific dated source, treat the claim as unverified.
  3. Pick a search-first tool for research-heavy questions. Perplexity is built around retrieval and shows sources by default. ChatGPT, Claude, and Gemini can all search, but how often they do it depends on the model, the plan, and the prompt.
  4. Click through on anything you'll act on. A summary of a page is not the page itself.

Where ChatGPT looks when it does search

The rest of the study is useful for anyone who publishes content. When ChatGPT searches, it often goes straight to the official source. 6.8% of fan-out queries used a site: operator to search a single website, spread across ~116K different domains. Those single-site searches tend to target brands. About 76% of ChatGPT's most frequent site: operator searches targeted brand websites. ChatGPT often looks for official product information, prices, and specifications.

Ahrefs also tracked which sites ChatGPT cited most in the US, and the list moved toward primary sources. PubMed moved up from 64th to fourth place among the domains ChatGPT most frequently cited in U.S. answers from March to September, while WebMD dropped from 18th to 79th. Community and reference sites dropped sharply too: Reddit was cited in 19.1% of U.S. ChatGPT answers in March and 2.3% in September, in Ahrefs' data. Wikipedia went from 14.1% to 2.9% over the same period. Ahrefs adds that both are still among ChatGPT's most-cited domains.

For publishers, this points to two separate goals:

  • Being part of what the model already knows. For most answers, which come from training data, your best chance is having been widely written about before the training cutoff. One analysis of the study notes that if 85.6% of ChatGPT responses come from training data, the content that matters is content that was already in the training set.
  • Being the page a search finds. For the minority of answers that do search, clear official pages with current prices, specs, and plain facts get picked. We covered how to structure pages for this in 7 ways to make your content easy for ChatGPT and Google AI to cite.

Some caution about the source: this is Ahrefs' study, and Ahrefs sells the citation-tracking tool that supplied part of the data, so read the recommendations with that in mind. Also, Search Engine Journal's coverage notes that other studies report different retrieval rates and that the setups are not directly comparable.

The takeaway

ChatGPT answers mostly from memory and checks the web only some of the time. For explanations, brainstorming, and drafting, that's fine. For anything that changes, like prices, features, rules, or recent news, you need to get it to search or check the answer yourself. A confident answer isn't necessarily a current one, and the sources it shows aren't always pages it actually read.

FAQ

Does ChatGPT search the web for every question?

No. In Ahrefs' sample of free, logged-out sessions from August and September 2026, only 14.4% of responses contained fan-out (search) queries. Paid plans and reasoning models search more often, so that figure is a minimum. Your own rate depends on your plan, the model, and how you phrase the question.

How can I make ChatGPT use current information?

Ask it to search the web, or turn on the search option, especially for prices, product features, laws, or recent events. Then ask for the source and date behind key claims and open the links yourself. If a question is mostly about current facts, a search-first tool like Perplexity is often a better place to start.

What does "query fan-out" mean?

It's the set of background searches an AI assistant runs before it answers. It might rewrite your question into one or two web queries, read the results, and combine them into one answer. In Ahrefs' data, searching responses averaged about 1.6 fan-out queries. That means a page usually has to match one of just one or two searches to be found.