What search engine does Claude use?

Claude can search the web: the feature launched in March 2025 for paid users in the US, and after a May 2025 update Anthropic says it's available on all Claude plans worldwide.

Developers get the same ability through the API's web search tool, which Anthropic prices at $10 per 1,000 searches plus normal token costs.

In the Claude apps, Claude decides when a question needs a search (current events, prices, anything that changes) and answers from what it already knows when it doesn't. You can also just ask it to search.

Claude searches the web by writing its own query, running it, reading the results and citing what it used, and Anthropic's documentation says it can repeat that loop several times in one answer: simple questions typically take one to three searches, comparisons can take ten or more.

In our October 7, 2026 study, Claude Haiku 4.5 (through Anthropic's API with search switched on, the same setup our fan-out tool uses) ran exactly one search for each of 100 prompts.

Those searches were plain. They averaged 6.5 words, shorter than the prompts themselves (7.7 words), only 3 of the 100 repeated the prompt word for word, and 31% leaned on the word "best". For "best CRM for a 10-person sales team", Claude searched "best CRM for small sales team 10 people 2026" and cited two independent roundups, while ChatGPT named four vendors in its search and cited their pricing pages (that comparison is in what search engine ChatGPT uses).

Claude's searches were also the most repeatable of the three engines we tested. Asked the same prompt three times, Claude ran an identical first search in all three runs for 8 of 20 prompts; ChatGPT and Gemini managed that for 0 of 20. If you want to know what Claude will search for your topic tomorrow, today's search is a decent guide, which isn't true of the other two. The full results are in our fan-out queries study.

Claude doesn't use Google as far as any evidence shows: nothing Anthropic publishes mentions Google Search, and our own numbers point the same way.

Of the 130 pages Claude cited across 30 prompts, 4 (3%) were in Google's top 10 for the search Claude had just run. For Gemini, which grounds its answers in Google Search, the figure was 18 of 130 (14%) on the same prompts.

Both numbers are rough, because our Google results came through a SERP feed that sometimes left brands' own pages out. The gap between the two engines is the point: Claude's sources look like they come from a different index, which is what you'd expect if that index is Brave's.

Claude's crawlers are three separate bots with three separate jobs, and Anthropic's help article on its crawlers describes each one:

• ClaudeBot collects web content that may be used to train Anthropic's models. Blocking it keeps your future content out of training.
• Claude-SearchBot crawls to improve the quality of Claude's search answers. Blocking it stops your content being indexed for that, which Anthropic says may reduce your visibility in user search results.
• Claude-User fetches a page when a person's question in Claude needs it. Blocking it stops Claude retrieving your page for that user.

Anthropic says its bots honor robots.txt, supports the non-standard Crawl-delay line, and publishes the IP addresses its crawlers use (claude.com/crawling/bots.json), so you can tell the real bots from fakes. Notice what none of the three is: Brave's crawler. Claude-SearchBot improves Claude's side of search, but the index Claude queries appears to be Brave's, and Brave crawls under its own rules.

To get your site into Claude's search results, you need to be in Brave's index, let Anthropic's search bots read the page, and rank in Brave for the plain question, which is what Claude actually searches.

1. Get crawled by Brave. Brave's crawler doesn't announce a user agent of its own, and its crawler help page says that if Googlebot can't crawl a page, Brave's bot won't crawl it either. So a robots.txt rule that shuts out Googlebot also shuts you out of Claude. Brave also learns about pages from browser users who opt in to its Web Discovery Project, and site owners can ask it to re-fetch a page.
2. Let Claude-SearchBot and Claude-User in, both in robots.txt and in your firewall. Blocking ClaudeBot (training) is a separate decision and doesn't take you out of search.
3. Rank for the plain question. Claude's searches were short and literal, usually the question with "best" or a year added, so the classic head term matters more for Claude than for ChatGPT, which names brands instead.
4. Use noindex, not robots.txt, to keep a page out of Brave. Brave says robots.txt doesn't stop a page being indexed; a noindex directive does.

None of that is new. It's crawling, indexing and ranking, in an index most SEOs never check.

To check if Claude can find your page, look at the Brave side and the Anthropic side separately, because either can fail without telling you.

Our free AI visibility checker does both for Claude: it checks whether Brave has your page and whether the page comes up in Brave for your topic, reads your robots.txt the way Claude-SearchBot, Claude-User and ClaudeBot would (and the way Googlebot would, since Brave follows Googlebot's rules), and sends real requests with Anthropic's user agents from our server to catch a firewall that blocks them.

To see the searches Claude actually runs for a prompt in your market, put that prompt into our query fan-out tool, which shows Claude's searches next to ChatGPT's and Gemini's.

What most people overlook about Claude search is how dated some of its searches are.

In our 100-prompt run, Claude added a year to 43 of its 100 searches, and 13 of those included 2024, close to two years stale by October 2026: "best budgeting apps after Mint shutdown 2024", "best free courses learn data analysis 2024", "most durable carry-on luggage 2024 2025". It looks like the model reaching for the year its training data ends in, and the search engine then returns pages that match the year it was given.

Two things follow from that. A page still carrying "2024" in its title can match a Claude search that a fresh 2026 page doesn't, which is an odd reward for not updating. And a brand that only appears in this year's roundups can be invisible to a Claude search pinned to 2024. You can't change the model's calendar, but you can make sure your key pages answer the question in their titles and opening lines without leaning on the year to do it.

Here are short answers to the Claude search engine questions people ask next.

Is Claude a good search engine?

Claude is a better research assistant than search engine: it runs a search, reads the results and writes a cited answer, so its quality depends on Brave's results and on how well it phrases the search. For finding one specific page, a search engine is still quicker.

Does Claude cite its sources?

Claude cites its sources whenever it uses web results; Anthropic's announcement says it gives direct citations so you can check them, and in our study it cited 4.3 pages per prompt on average.

Can I submit my site to Brave Search?

You can ask Brave Search to re-fetch a page, or to delist one that no longer exists, through the options on Brave's crawler help page. Most discovery still comes from Brave's crawler following links and from Web Discovery Project data, so links from sites Brave already crawls do the heavy lifting.

Does blocking ClaudeBot remove me from Claude's answers?

Blocking ClaudeBot doesn't remove you from Claude's search answers, because ClaudeBot collects training data. The bots that matter for search are Claude-SearchBot and Claude-User, plus Googlebot, whose rules Brave follows.

Does Claude use Bing?

Claude doesn't use Bing as far as the evidence shows: Microsoft appears nowhere in Anthropic's published documents or in the independent tests, all of which point to Brave.

How many searches does Claude run per question?

Claude can run several searches per question (Anthropic's documentation says one to three for simple questions and ten or more for comparisons), but Claude Haiku 4.5 ran exactly one search for each of our 100 prompts. Larger Claude models in the Claude apps may search more; our numbers describe the setup our tool uses.