What search engine does Claude use for web search?

There is no single documented answer, and that matters for how you plan. Two pieces of evidence exist.

First, Simon Willison reported on 21 March 2025 that Anthropic's Trust Center had added Brave Search as a subprocessor. That listing appeared around the launch of web search in Claude, so the reasonable reading is that Brave supplies at least some of the results Claude sees when it searches. A subprocessor listing is a data-handling disclosure, though. It does not say every query goes to Brave, and it does not say Brave is the only provider. As of October 2026, we have no newer first-party statement that confirms or replaces it.

Second, Anthropic runs its own search crawler. Its crawler documentation (updated 7 April 2026) describes Claude-SearchBot as the agent that improves search quality, distinct from ClaudeBot, which collects training data, and Claude-User, which fetches pages when a user asks.

What this means: Claude's web results likely draw on Brave plus Anthropic's own crawling, and the mix is not public. You cannot optimise for an index you cannot see. You can make sure nothing on your side blocks either path.

A search engine only matters when Claude actually searches. On every platform the evidence base covers, retrieval is per query and discretionary.

• Claude.ai: web search is a user toggle. Anthropic publishes no default for it.
• The Claude API: developers add the web search tool, and the integrator's system prompt steers when Claude uses it.
• The model itself: even with search available, Claude decides whether a question needs it.

When search is off, Claude answers from what it learned in training, frozen at its knowledge cutoff. Nothing you publish this quarter reaches that answer. Our page on whether Claude searches the web by default covers the toggle in more detail, and you can learn how to tell if an answer used live web search before reading anything into a result.

The next move: when you test Claude, note whether search was on. A missing brand mention in an unsearched answer says nothing about your site.

Anthropic's web search tool documentation is the clearest public view of what Claude receives. Each search result exposes a url, a title and a page_age field, described as when the site was last updated. Citations point to specific text spans, with a cited_text of up to 150 characters, rather than whole pages. Anthropic prices the tool at $10 per 1,000 searches, plus tokens.

Two cautions apply. The docs do not name the index behind the results. And exposing page_age is not the same as weighting it: no vendor states that freshness changes which pages get picked. Our digest of the Claude web search tool documentation goes field by field.

What it means for you: Claude quotes short spans. A page that states a specific fact plainly, such as a price or a one-line definition, gives Claude a sentence it can quote. That is the only on-page pattern with a positive experimental signal, and it still promises nothing.

Whatever index sits behind Claude, these are the gates you control. They work by removing blocks, not by adding signals.

1. Allow Claude-SearchBot in robots.txt. Rules are per bot. A blanket AI block written to stop training often removes search crawlers too. If you want to stay out of training but stay eligible for search, keep the two separate:

User-agent: ClaudeBot
Disallow: /

User-agent: Claude-SearchBot
Allow: /

2. Check the firewall and CDN. robots.txt can say allow while your bot management returns a 403 or a challenge page. Anthropic warns that IP blocking impedes its ability to read your robots.txt file at all. Request a page with the Claude-SearchBot user agent and confirm you get a 200, not a block. The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.

3. Put key facts in the raw HTML. Vercel's crawler study found Claude's crawlers fetched JavaScript files on 23.84% of requests but did not execute them. Independent testing in June 2026 found Claude-SearchBot does not render JavaScript either. Pricing or product details that only appear after hydration may be invisible.

4. Leave snippet controls off by accident. Check templates for stray noindex or nosnippet directives.

The full breakdown of ClaudeBot, Claude-SearchBot and Claude-User explains each agent's role, and the training vs search crawler comparison lines them up against other vendors.

If Brave supplies part of Claude's results, then being present in Brave's index is plausibly upstream of appearing in Claude. That link is reasonable, not measured. Nobody has published how much of Claude's retrieval comes from Brave, or whether being indexed by Brave changes Claude citations.

So treat Brave as a secondary check, not a strategy. Make sure you are not blocking Brave's crawler, and see how Brave Search AI answers pick sources for that engine on its own terms.

What stays out of your control: which provider Claude queries, what queries it writes, and which retrieved pages it chooses to cite. Passing every access check buys eligibility, not citations.

Is Claude-SearchBot the same as ClaudeBot? No. Anthropic documents them as separate agents. ClaudeBot collects content that may be used for training. Claude-SearchBot improves search quality. Claude-User fetches pages when a person asks Claude to. You can block one and allow another in robots.txt, and blocking ClaudeBot for training reasons does not need to touch Claude-SearchBot.

Does Claude use Google or Bing for web search? The evidence base contains no first-party statement that Claude uses Google or Bing. The public signal points to Brave, from the March 2025 subprocessor listing reported by Simon Willison, plus Anthropic's own Claude-SearchBot crawling. Anthropic has not published a complete provider list, so treat any confident single-engine claim with caution.

Will allowing Claude-SearchBot get me cited by Claude? No. Allowing the crawler makes your pages eligible to be fetched. It does not guarantee retrieval for a given question or a citation in the answer. Claude decides when to search and which results to cite. Access is the part you control, and blocking it is the part that reliably removes you.

Does schema markup or an llms.txt file help with Claude search? Anthropic does not document reading llms.txt, and no engine in the evidence base does. Ahrefs' matched test of schema found no meaningful citation lift across platforms. Neither is a lever for Claude. Existing markup does no harm, but the evidence puts your effort on access, rendering and stating clear facts.