Every answer Claude gives from the web carries citations, and those citations are drawn from a search pipeline that Anthropic's own subprocessor list names: Brave Search and TurboPuffer. Claude's share of AI referral traffic is anywhere from 0.95% to 18.5% depending on whose panel you read — which says more about audiences than about Claude.
This is how Claude actually finds, fetches and cites a page, and what a site can do about each step.
How does Claude decide when to search the web?
Claude searches when a question depends on information that changes or that it cannot know from training: recent events, current prices and statistics, and specific organisations, people or products. For settled facts and creative tasks it answers from the model. The decision is made per question, not per conversation.
That rule is spelled out in Anthropic's web search tool documentation, which lists "current prices, rates, scores, or statistics" and "specific organizations, people, or products that might have changed" as triggers. A simple query typically runs one to three searches; comparative research can run ten or more.
The practical consequence is that commercial and comparison queries are the ones most likely to trigger retrieval — "which agency", "how much does it cost", "X versus Y". Those are exactly the queries a business wants to be cited on, and they are where the page itself matters rather than whatever the model absorbed in training.
Which search engine does Claude use?
Anthropic lists two web-search subprocessors: Brave Search and TurboPuffer. Brave supplies an independent web index; TurboPuffer is a search database company, which suggests Anthropic now runs retrieval infrastructure of its own alongside Brave. Image results in Claude.ai come from Bing.
The Brave connection was first reported by TechCrunch in March 2025, after developers noticed Claude's citations matched Brave's results exactly. The subprocessor list now confirms it. The TurboPuffer reading above is our interpretation of the vendor's role, not something Anthropic has described.
This is the single most overlooked fact in AI visibility work. Most sites optimise for Google and check Bing for ChatGPT. Very few have ever looked at whether Brave indexes them at all, or how they rank there. If Brave does not have the page, Claude's primary web retrieval cannot find it.
What are ClaudeBot, Claude-User and Claude-SearchBot?
They are Anthropic's three crawlers, and they do three different jobs. ClaudeBot collects content that may be used for model training. Claude-User fetches pages live when a person asks Claude a question. Claude-SearchBot indexes content to improve search result quality. Blocking each one has a different consequence.
Anthropic's crawler support article is explicit about the trade-off. Blocking Claude-User "may reduce your site's visibility for user-directed web search"; blocking Claude-SearchBot "may reduce your site's visibility and accuracy in user search results". Blocking ClaudeBot only excludes future content from training.
All three respect robots.txt, including the non-standard Crawl-delay, and
Anthropic notes the opt-out must be set on every subdomain. IP blocking, it
says, "may not work correctly or persistently guarantee an opt-out".
A reasonable position for most businesses is to allow Claude-User and Claude-SearchBot even if you block ClaudeBot. Training and citation are separate decisions. This site allows all three by name in its robots rules, rather than relying on a wildcard, so the intent is unambiguous.
Can Claude read every page it finds?
No. Claude's web fetch tool does not execute JavaScript, so a page whose content only appears after client-side rendering reaches Claude as an empty shell. It also respects robots.txt, refusing URLs a site disallows. Server-rendered HTML is the only version of a page Claude reliably sees.
The web fetch documentation states that it "does not support websites dynamically rendered with JavaScript". A typical 10 kB page costs about 2,500 tokens to read, which is why clean, compact HTML with the answer near the top is cheaper for an agent to use, not just easier.
This is where technical SEO stops being a Google-only discipline. A single-page app that renders product details in the browser can rank in Google, which runs JavaScript, and still be invisible to Claude, which does not. Check what your pages return with JavaScript disabled; that is roughly what Claude reads.
What does a Claude citation actually contain?
Each citation carries the page URL, its title, and up to 150 characters of the
quoted text. Search results also include a page_age field recording when the
page was last updated. The citation is therefore a short extracted passage, not
a summary of the whole page — which is why passage structure matters.
Those fields come from the web search tool's documented response format. For developers using the API, "citations are always enabled for web search", and the tool costs $10 per 1,000 searches plus tokens. On Claude.ai, Anthropic's help centre says every web-informed response includes citations.
A 150-character quote is about one sentence. That is the strongest argument for answer-first writing: the sentence directly beneath a question-shaped heading should answer it completely, because that sentence may be the only part of your page that appears in the answer.
How much traffic does Claude actually send?
It depends heavily on the audience. Across broad web panels Claude sends a small single-digit share of AI referrals — and sometimes less. In B2B panels it sends far more. Any single "Claude share" figure is only meaningful alongside the panel and the month it was measured in.
The figures, with their methodology:
| Source | Claude share of AI referrals | Panel | Period |
|---|---|---|---|
| StatCounter | 0.95% (peaked at 3.74% in June) | 1M+ sites, worldwide | September 2026 |
| SE Ranking | 2.62% | 101,574 sites | 2026 to April |
| Goodie | 18.5% | B2B brand sites | March–April 2026 |
The broad panels and the B2B panel are not contradicting each other; they are measuring different audiences. Developers, technical buyers and professional services clients use Claude far more than the general population. A B2B SaaS company and a consumer retailer should not use the same benchmark.
Growth is real but volatile. SE Ranking's dedicated Claude study measured a 386% rise in Claude referrals between January and April 2026. StatCounter's monthly data then shows Claude's share falling back from its June peak through September. Treat Claude as a meaningful channel for specific audiences, not a mass one.
What makes a page more likely to be cited?
Statistics, quotations and cited sources. The most rigorous published test of this, the GEO study from Princeton and partners, found those three changes lifted visibility in generative answers by 30–40%, while keyword stuffing produced "little to no improvement".
That study used GPT-3.5 and predates Claude's search entirely, so treat the magnitudes as indicative rather than transferable. Its most useful finding for smaller sites is that lower-ranked sources gained most: adding citations lifted visibility by 115% for sites ranked fifth, while the top-ranked site lost ground.
Combining that with how Claude retrieves gives a short, mechanical checklist:
- Be in Brave's index. Search for your key pages on Brave; if they are absent, nothing downstream matters.
- Allow Claude-User and Claude-SearchBot in robots.txt, on every subdomain.
- Serve content as HTML, not client-rendered JavaScript.
- Answer in the first sentence under question-shaped headings, inside 150 characters where you can.
- Carry specific, sourced numbers. Claude searches precisely when a question involves current figures.
- Keep pages fresh.
page_ageis in every result; a refresh with new data is a signal, a date change is not.
The broader discipline behind that list is what we mean by generative engine optimization, and we set out the definitions in what GEO actually is.
How do you know whether Claude cites you now?
Ask it. Run the twenty to fifty questions your buyers actually ask, in Claude with web search on, and record which domains are cited and in what position. Repeat monthly. A baseline built this way is the only reliable measure, because no analytics tool reports citations that did not produce a click.
Referral traffic undercounts the effect badly. Goodie's own report warns that "every number in this report understates AI's true footprint", because many citations never produce a visit. Our method for building a repeatable baseline across engines is set out in full in running an AI citation baseline.
The short version
Claude searches when a question involves current or specific facts, retrieves through Brave and its own infrastructure, reads only server-rendered HTML, and quotes about one sentence per source. Each of those steps is something a site can check and fix in an afternoon.
Whether Claude matters to your traffic depends on who your buyers are, not on any headline market share. If you want to know where you stand across Claude, ChatGPT, Gemini and Perplexity before deciding what to fix, that baseline is the first deliverable of our GEO engagements.