Claude Web Search: How It Works and Cites

Claude web search is the feature that lets Claude look things up on the live web while it answers, and cite the pages it used. In the Claude apps you switch it on from the + menu, or in the newer Claude experience it simply searches when that helps. Developers get the same capability through the web search tool in the Claude API. Anthropic says every response that uses search includes citations, which makes Claude one more AI engine that can send readers, and credibility, to your site.

This guide covers both sides, from Anthropic’s own help center and API documentation, checked September 22, 2026. The first half is for people using Claude. The second half is for site owners who want Claude to be able to find, read and cite their pages.

On this page

• How do you turn on Claude web search?
• How Claude decides to search
• How Claude cites its sources
• Getting better answers from Claude web search
• Claude web search in the API
• Which crawlers does Anthropic use?
• Making your pages citable in Claude web search

How do you turn on Claude web search?

You turn on Claude web search from the + button in the lower left of the chat window: select Web search and a checkmark appears. Anthropic’s help article on web search adds three conditions worth knowing:

• The newer Claude experience has no toggle. “Claude searches the web when it helps.”
• Team and Enterprise plans need an admin first. An Owner or Primary Owner must enable web search for the workspace in Organization settings > Capabilities before members can use it.
• The model has to support it. The help page lists Opus 5.5, Fable 5.1, Opus 5, Sonnet 5, Fable 5, Opus 4.8, Opus 4.7, Sonnet 4.6, Opus 4.6 and Haiku 4.5.

If search seems “disabled,” those three are the usual causes. On free accounts there is a fourth consideration: Anthropic says usage limits reset every five hours and that web search and web fetch both count toward them, so switching search off for tasks that do not need it saves capacity.

With web search on, Claude can also fetch a specific URL you paste in, a feature Anthropic calls web fetch. The help page warns that asking Claude to summarize a long article pulls the whole thing into the context window, which uses much more of a free plan’s allowance than a normal search.

How Claude decides to search

Claude searches when a question depends on information that is current, changing or outside its training data. Anthropic’s web search tool documentation lists the typical triggers: current prices, rates, scores or statistics; information about organizations, people or products “that might have changed”; and explicit requests to look something up. You can force it by writing “Search the web” in your prompt, or tell Claude not to search.

The docs also give a sense of scale: simple factual queries typically use one to three searches, while “comparative or multientity research can use 10 or more.” The diagram follows one search from question to cited answer.

Scroll the diagram sideways to see all of it.

One thing Anthropic does not document is which search index powers Claude’s text results. The help page says image results are “powered by Bing.” It does not name a provider for web results, so we will not guess. What it does say is that Claude “may use your location (inferred from your IP address)” for local questions.

How Claude cites its sources

Claude cites sources inline with a link, and Anthropic says “Every response includes citations” when search is used. In the apps, answers include “direct citations to sources,” “source links for further reading” and “relevant quotes when appropriate.”

The API documentation shows what a citation contains, which tells a site owner exactly what gets lifted from a page.

Scroll the diagram sideways to see all of it.

The 150-character limit on cited_text is the useful detail. A citation quotes a sentence or less, so a page whose key fact is spread across a paragraph gives Claude nothing clean to quote. That is the same logic behind content chunking for AI: lead with the fact, keep it in one sentence, then explain.

Getting better answers from Claude web search

Better answers from Claude web search come from telling it what kind of sources you want and then checking the ones it cites. Anthropic’s own tips are short: ask about recent information, say “Search the web” when you want to be sure it searches, request multiple sources, and “cross-reference cited sources” before relying on them for important decisions.

A few habits build on that:

• Name the source type. “Use the manufacturer’s documentation and at least one independent review” produces a different answer from a bare question.
• Put dates in the prompt. “Changes announced since January 2026” keeps it from mixing old and new information.
• Paste the URL when you already know the source. Web fetch reads that page directly instead of searching for it, though on a free plan a long page uses more of your allowance.
• Open the citation behind any number. The quoted span is short, so check that the surrounding page says what the answer implies.

Anthropic also lists limitations: availability can vary with connectivity, some links may not work, and search times vary with query complexity.

Claude web search in the API

In the Claude API, web search is a server-side tool you add to a request, and the API runs the searches and returns results to Claude. The API page lists three versions: web_search_20250305 for basic search, web_search_20260209, which adds dynamic filtering, and web_search_20260318, which adds response inclusion control. Dynamic filtering lets Claude “write and run code that filters the search results before they reach the context window,” and is available with Claude 4.6 and later models.

The table compares the two ways of using Claude web search.

For builders, the cost note matters: Anthropic says search results “are counted as input tokens” in the turn where they are retrieved and in later turns, so long conversations with many searches add up.

Which crawlers does Anthropic use?

Anthropic uses three bots, and each does a different job. Its crawler help article lists them with what happens if you block each one, summarized in the matrix below.

Scroll the diagram sideways to see all of it.

In Anthropic’s own words, blocking Claude-User “prevents our system from retrieving your content in response to a user query, which may reduce your site’s visibility for user-directed web search.” Blocking ClaudeBot, by contrast, is documented only as a training signal: your future content should be excluded from training datasets.

Anthropic says its bots honor robots.txt, support the non-standard Crawl-delay directive, and “will not attempt to bypass CAPTCHAs.” It also warns that blocking by IP address “may not work correctly or persistently guarantee an opt-out,” because it stops the bot from reading your robots.txt. The published IP list lives at claude.com/crawling/bots.json.

That robots.txt promise is worth comparing across vendors, because the user-triggered fetchers differ. OpenAI’s bots page says that for ChatGPT-User, which acts on a user’s request, “robots.txt rules may not apply.” Perplexity’s crawler documentation says Perplexity-User “generally ignores robots.txt rules.” Anthropic’s article makes no such exception for Claude-User; it presents Claude-User as the control that lets site owners decide which sites user-initiated requests can reach. If you want to be strict about which AI tools can read your pages on demand, Anthropic’s documented setup looks the simplest of the three to control with robots.txt alone.

For most businesses that want AI visibility, the sensible setting is to allow Claude-SearchBot and Claude-User and decide separately about ClaudeBot. The syntax is in robots.txt for AI crawlers, and the trade-offs are in should you block AI crawlers.

Making your pages citable in Claude web search

To be cited in Claude web search, a page has to be reachable by Anthropic’s search bots, readable without guesswork, and written so a single sentence can carry each key fact. In order of impact:

1. Allow Claude-SearchBot and Claude-User, in robots.txt and at your CDN or firewall. A bot that gets a challenge page cannot cite you.
2. Serve real HTML. If your main content only appears after JavaScript runs, check it with JavaScript rendering for AI crawlers.
3. Write quotable sentences. With a 150-character quote span, put the fact, number or definition in one plain sentence near the top of its section.
4. Use descriptive titles. The citation shows your page title; “Services” tells a reader nothing.
5. Keep facts consistent across the web. An engine that reads several sources favors facts it can confirm, a point covered in retrieval vs citation.

None of this is Claude-specific, which is the good news. The same work makes you eligible across ChatGPT, Perplexity and Google, as mapped in how AI search engines find sources. Measuring the result is the harder part, because Claude does not offer a site-owner report; referral traffic and manual prompt checks are the practical options, set up in tracking AI traffic in GA4 and Search Console.

If you want to know whether Claude, ChatGPT and Perplexity can reach your site today, and which of your pages they cite, our GEO service starts with that access and citation check.