Claude's Web Search: Where Its Results Come From

By Muntazir · Published 30 October 2026

A regional appliance repair shop asks Claude, through its web search tool, which other shops in town also fix induction hobs. The answer names two competitors and links to a listing updated eleven days earlier — a page that did not exist when Claude’s training data was gathered, so no amount of memory could have produced it. An hour later, a customer asks a plainer question: who installed the shop’s own boiler-brand certification. Nothing comes back. The page answering that sits three clicks deep on the shop’s own site and never surfaces.

Same tool, same model, two different outcomes — because “Claude searched the web” describes at least three separate systems, and only one of them is well documented.

What actually happens when Claude searches the web?

Claude’s web_search tool is a live, per-request action, not a lookup against anything baked into the model. Anthropic’s own developer documentation describes Claude deciding, mid-answer, whether a question depends on information that is “current, changing, or outside its training data” — a price, a recent announcement, a fact about a specific business that might have moved. When it decides yes, it issues a search query, gets back a list of results — each with a URL, a title and a page_age showing when the page last changed — and writes its answer with citations pointing at those exact URLs. None of that comes from memory. It comes from whatever the search just returned, which is why the same brand can get a sharp, current answer in one reply and nothing at all in the next.

Which search engine actually supplies those results?

Not Claude’s own. Google Cloud’s documentation for running Claude on Vertex AI states it plainly: “The Claude models use a third-party search provider for the web search feature,” adding that the data sent to it is governed by “the third-party’s own terms of service and privacy policy” rather than Google’s. Anthropic’s own materials describe the same arrangement without ever naming the company.

The clearest evidence of which one points to Brave Search, and it is evidence rather than confirmation. In the same week Claude’s web search launched in March 2025, Anthropic’s own subprocessor list added Brave Search — a dated change to Anthropic’s own disclosure, not a third party’s guess. Independent testing the next day found Claude’s cited results matching Brave’s results for an identical query, and a parameter in Claude’s own tool definition named “BraveSearchParams.” Anthropic has not put the name in its documentation since, so that remains circumstantial, dated evidence rather than an official statement.

Is this the same thing as Claude-SearchBot crawling a site?

Not provably, and that gap matters more than it first looks. Anthropic’s crawler documentation describes Claude-SearchBot as a bot that “navigates the web to improve search result quality” and indexes content “for search optimization.” That reads like a description of exactly the retrieval step above. But nowhere does Anthropic’s documentation say Claude-SearchBot’s index is the one the web_search tool actually queries, rather than a separate signal feeding some other internal ranking step. The two capabilities sit in different documents, and Anthropic never draws the line between them. Claude-SearchBot is one of three distinct Claude crawlers with three distinct jobs — training, live user fetches and this indexing role — and what each one does, and how to block only the one meant is covered there rather than repeated here.

How does a page get into Claude’s search, or stay out of it?

Mostly the same way it gets into any crawler-fed system: by being reachable and not disallowed. Claude-SearchBot and Claude-User each take their own line in robots.txt, and disallowing one does nothing to the other — a site that blocks Claude-SearchBot still gets live, user-triggered fetches through Claude-User, and the reverse holds too. A developer calling the API separately can restrict web_search to an allowed_domains or blocked_domains list, but that list belongs to whoever built the integration asking the question, not to the site being searched — there is no form a business fills in, no verification step, nothing resembling the licensing deals a handful of named publishers have struck with other AI companies. A site’s only lever is access and clarity, applied evenly, with no guarantee either one gets picked up on a given search.

Can anything confirm whether Claude’s search will cite a specific page?

Partly, and the honest half of that answer is where this gets limited. A free crawler check can send a real request identifying as ClaudeBot, Claude-SearchBot and Claude-User, separately, and see whether each one is actually let through at the server level rather than just permitted on paper — the same split the free robots.txt checker runs for several named crawlers at once, because a CDN’s bot-blocking rule and a robots.txt file do not always agree. That establishes whether a page is even reachable.

It does not establish what gets cited. A deeper AI-visibility review’s model probes run through a general-purpose chat completion with no browsing tool attached at all — they test what a model’s training already holds about a brand, never what a live search would retrieve for it right now. They never call web_search and watch what comes back, so they cannot say whether Claude’s own search would cite a given page today, the same honest gap already documented for ChatGPT’s training-memory-only probe design.

Being easy for a crawler to reach and clear about what a business does and where — the practice this site’s AI-SEO page sets out — is still the only lever that touches all three systems at once. It doesn’t guarantee a citation from any of them.

Three systems sit behind what looks like one Claude answer: frozen training memory, a third-party search index nobody outside Anthropic has named on the record, and Anthropic’s own crawler indexing separately for a purpose its documentation does not quite finish explaining. A site can block all three, address none of them, and still show up by accident — or get everything right and still not know which of the three answered today’s question.