How do ChatGPT, Claude and Grok search the web?

Published 29 October 2026. Updated 29 October 2026.

ChatGPT, Claude and Grok all search the web in the same basic way. The model decides a question needs current information, writes its own search queries, reads the pages that come back and cites some of them in its answer.

The engines differ in what they search, how many rounds they run and how they show their sources. This page takes each engine in turn, using what OpenAI, Anthropic and xAI publish about their own search tools. If the term is new to you, our guide to query fan-out explains how one prompt turns into several searches.

In this hub

Posts in this hub go deeper on one engine or one question at a time, and each one will be listed here. Every guide on the site is on the Only Fan Out blog.

What happens when an AI assistant searches the web?

Every search follows the same loop of five steps, though each company describes them in its own words.

Step	What happens	What the docs say
1. Decide	The model judges whether the question needs information it does not have.	OpenAI says ChatGPT "may search the web automatically when your question would benefit from current information". Anthropic says "Claude determines when to search based on the prompt."
2. Write queries	The model rewrites the prompt into its own search queries.	OpenAI says ChatGPT "typically rewrites your query into one or more targeted queries".
3. Search	The queries go to a search backend, which returns a list of results.	OpenAI says ChatGPT search "sometimes partners with other search providers".
4. Open and read	The model opens some results, reads them and may search again.	xAI says Grok's tools run in a loop that "continues until sufficient information is gathered".
5. Answer and cite	The model writes its answer and links a subset of the pages it read.	Anthropic says "Citations are always enabled for web search."

The queries in step 2 decide which pages the model ever sees. A page that ranks for none of them never reaches step 4, however well it answers the question the person typed.

How does ChatGPT search the web?

ChatGPT searches on its own when it judges a question needs current information, and you can force a search by picking Search from the tools menu. OpenAI's help article on searching the web with ChatGPT says it rewrites the question into one or more targeted queries, sends them to search providers it partners with, and may send follow-up queries after reading the first results. The article does not name those providers. It says ChatGPT does not pass your IP address or account details to them, though it may share a general location.

In captured sessions, ChatGPT writes four kinds of query: web, business, image and find. A single turn often mixes them. Shopping questions also return product cards with price, rating and merchant, which sit apart from the cited sources.

OpenAI's web search guide for developers lists three actions a search step can take: run a search, open a page and search within a page. The same guide separates the full list of sources the model consulted from the shorter list it cites inline. In the ChatGPT app, the Sources button shows the cited pages plus other relevant links.

Since 14 September 2026, ChatGPT sends the text of its search queries only in the live answer stream and drops it from the saved conversation. A tool that reads the saved conversation later sees the sources but not the queries.

How does Claude search the web?

Claude searches when a request depends on information that is current, changing or outside its training data, according to Anthropic's web search tool documentation. The API runs each search and hands the results back to Claude, and that can repeat several times within one request before Claude writes its answer.

In the Claude apps, Anthropic's help article on web search says you turn search on from the + menu, while the newest Claude experience has no toggle and searches the web when it helps. You can ask for a search in the prompt, or ask Claude not to search. Image results come from Bing, and for local questions Claude may use a location inferred from your IP address.

Claude attaches each citation to the sentence that relies on it. In the API, every citation carries up to 150 characters of the text it quotes. Captured sessions show web searches and places searches, grouped by turn, and Claude does not cite the places it plots on a map.

How does Grok search the web?

Grok has two search tools, according to xAI's search tools documentation. Web Search searches the web and browses pages, and X Search searches X posts, users and threads. The model works in a loop. It decides whether to call a tool or answer, reads what comes back, and keeps going until it has enough.

Grok cites inline with numbered links. xAI's citations documentation says the numbers start at 1, rise in order and repeat when a source comes up again. Its full citations list holds every source the agent came across, and xAI notes that not every URL in that list is referenced in the final answer. Captured sessions show three kinds of activity per turn: web searches, page browses and X searches.

How do ChatGPT, Claude and Grok compare?

All three decide for themselves when to search, but they search different places and show their sources in different ways.

Engine	When it searches	What it searches	How it shows sources
ChatGPT	When a question needs current information, or when you pick Search	Web, business, image and find queries, with follow-up queries after the first results	Inline citations, plus a Sources list of cited and other relevant links
Claude	When a request needs current or outside information, or when you ask	Web and places searches, repeated within one request when needed	Citations attached to the sentences that use them
Grok	Inside a tool loop that runs until it has enough	The web, the pages it browses and X posts	Numbered inline citations, with a fuller list that includes sources it did not use

Why do AI assistants read more pages than they cite?

Each engine keeps two lists: the pages it consulted and the smaller set it cites. OpenAI describes its sources list as "the complete list of URLs the model consulted", while inline citations "show only the most relevant references". xAI says the same of Grok's citations list. A page in the first list but not the second was found and read, then left out of the answer. The query fan-out guide covers that gap between fetched and cited sources in more detail.

How many searches does each engine run per answer?

None of the three companies publishes a figure, and third-party counts vary by an order of magnitude with the model, the account type and the prompt. Our guide to query fan-out keeps the published studies and their setups in one table. We have not found a published per-answer count for Claude.

Which crawlers let ChatGPT and Claude read your site?

OpenAI and Anthropic each run a search crawler, a fetcher for pages a user asks about, and a separate training crawler.

Company	Search crawler	Fetches for a user	Training crawler
OpenAI	OAI-SearchBot	ChatGPT-User	GPTBot
Anthropic	Claude-SearchBot	Claude-User	ClaudeBot

OpenAI's crawler overview says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", and that ChatGPT-User visits come from a user's action. Anthropic's article on its crawlers lists Claude-SearchBot for search quality and Claude-User for fetches a user asks for. Because the training crawlers are separate, a site can block training and still let both search crawlers in.

What does this mean for your site?

An assistant finds pages through the queries it writes, so the useful question is which queries it writes for your topic and whether you rank for them. A few habits follow from the docs above:

• Answer the question in the first sentence or two, so a page that gets opened has a passage worth citing.
• Back claims with your own data and named sources.
• Let OAI-SearchBot and Claude-SearchBot crawl the site.

How can you see what an assistant searched?

The chat apps show cited sources but keep most of the search out of view. Only Fan Out is a free Chrome extension that lists the queries ChatGPT, Claude and Grok ran for each turn, the pages they fetched and the pages they cited, in a side panel next to the chat. Nothing leaves your browser.

Related questions