How Claude searches the web: web search, web fetch, places and citations

Published 5 November 2026. Updated 5 November 2026.

Claude searches the web when it decides a question needs current information or facts it does not have. It writes its own search queries, reads the results, can open full pages with a separate fetch tool, runs a places search for local questions, and attaches each citation, with a short quoted passage, to the sentence that relies on it.

Our page on how AI assistants search the web compares Claude with ChatGPT and Grok. This page stays with Claude: when it searches, the tools it uses, where its results come from, what one search turn looks like in Anthropic's own data, and what that means if you want Claude to read and cite your pages.

When does Claude decide to search the web?

Claude searches when the answer depends on something recent, changing or specific, and answers from what it already knows when it does not. Anthropic's web search tool documentation lists both sides.

Claude searches for	Claude answers directly for
Recent events, news or announcements	Established facts, maths, science basics and coding concepts
Current prices, rates, scores or statistics	Creative writing and brainstorming
Organisations, people or products that may have changed	Analysis of content already in the conversation
Explicit requests to search or look something up	Greetings and small talk

In the Claude apps, Anthropic's help article on web search says you switch search on from the + menu, and that the newest Claude experience has no toggle and searches when it decides that helps. On Team and Enterprise plans, an Owner has to enable web search for the whole organisation first. A question Claude treats as settled fact may get no search at all, so no page is read for it.

Which tools does Claude use to look things up?

Claude has four ways to reach information on the web, and each leaves a different trace in the answer.

Tool	What it does	What comes back	Cited in the answer?
Web search	Runs a query Claude wrote	Results, each with a URL, title and page age	Yes, citations are always on
Web fetch	Opens one full page or PDF	The page's content and the time it was retrieved	Optional in the API, off by default
Places search	Looks up businesses and locations for local questions	Names, addresses, ratings and map positions	No
Image results	Adds images to a web search answer	Images, each linked to its source page	Each image links its source

The web search and fetch rows come from Anthropic's documentation for each tool, and the image row from its help article. The places row comes from conversations recorded with Only Fan Out, which show web and places searches grouped by turn and no citations on the places Claude plots on a map, and from the capture study described in the next section.

Which search engine does Claude use?

Anthropic's current documentation and help articles do not name the provider behind Claude's web search. The help article names one provider only: image results are powered by Bing.

When web search launched in March 2025, Simon Willison found Brave Search newly added to Anthropic's subprocessor list, and the ten pages Claude cited for one of his test searches matched Brave's results for the same query. We have not seen Anthropic confirm that this still holds.

Local questions take a different route. Nicolas Sitter logged Claude's event stream for hotel searches in April 2026 and found a places_search tool returning Google Places data, used in 8 of his 9 prompts, with no citations on the places it returned. He calls his sample small. For a local business, that makes the details in its Google Maps listing part of what Claude reads.

How many searches does Claude run per question?

Anthropic publishes no count for the Claude apps. Its guidance to developers, written for setting the max_uses limit, says simple factual queries typically use one to three searches and comparative or multi-entity research can use ten or more.

In the API, searching, reading and searching again can repeat several times inside one request, up to that limit. Sitter's hotel captures show Claude sending two or three places searches at once and sometimes a second round after reading the first results.

When does Claude open a full page?

Claude opens a full page only when its URL is already in the conversation: in your message, or in an earlier search or fetch result. Anthropic's web fetch tool documentation says it cannot fetch a URL that appears only in Claude's own output, so a page gets read in full only after a search has turned it up or someone has pasted the link.

The same documentation sets limits that matter to anyone who wants a page read:

• The fetch tool does not run JavaScript, so text that only appears after scripts run is missing from what Claude reads.
• It refuses URLs blocked by robots.txt, along with private addresses and other restricted URLs.
• Results are cached, and the content returned "might not always reflect the latest version available at the URL".
• It reads text, HTML and PDF only.

Those rules are written for the API. Anthropic's help article describes the same feature in the apps, where Claude reads full pages from links you give it.

What does one Claude search look like in the data?

In Anthropic's API, one search turn is a short chain of blocks, and each block answers a different question about your page.

Block	Fields	What it tells you
server_tool_use (web_search)	input.query	The exact query Claude wrote
web_search_tool_result	url, title and page_age for each result	Whether your page came back for that query
web_fetch_tool_result	url, retrieved_at	Whether Claude opened your page in full, and when
web_search_result_location citation	url, title, cited_text of up to 150 characters	Which sentence of your page the answer used

Because cited_text carries the quoted passage, a Claude citation shows which sentence on your page did the work. The gap between pages read and pages cited runs through all three assistants, and our post on fetched vs cited sources sets out where each one keeps its two lists.

How can you see what Claude searched in the Claude app?

The API hands you the blocks above. In the Claude app, Only Fan Out reads the responses claude.ai already sends to your browser and lists, for each turn, the web and places searches Claude ran, the pages that came back and the pages it cited. Places are split into those Claude used on the map and those it only named. To see how the extension reads that data with no server involved, read how Only Fan Out works.

What does this mean for your site?

Claude reaches a page through a search first and a fetch second, so start with access and ranking, then look at the page itself.

1. Let Claude-SearchBot and Claude-User in. Anthropic's article on its crawlers says blocking Claude-User "prevents our system from retrieving your content in response to a user query".
2. Rank for the searches Claude runs on your topic, which you can read in the data above.
3. Put your main facts in the HTML itself, since the fetch tool does not run scripts.
4. Write key sentences that make sense on their own, because a citation quotes at most 150 characters.
5. For a local business, keep the Google Maps listing complete and accurate.

None of this guarantees a citation. Anthropic publishes the rules above, not how Claude chooses which of the pages it read to cite.

Related questions