How to Get Cited by Claude in AI Search (2026)

Last updated: 9 August 2026

We’ve published per-engine citation playbooks for ChatGPT and Perplexity, and the honest reason Claude didn’t have one until now is that for a long time it barely sent anyone traffic. That has changed abruptly. Claude is now the fastest-growing AI referrer in the B2B data we can find, it runs on a different search index to every other engine, and almost nothing written about “AI SEO” applies to it specifically. This post covers what is actually different about getting cited by Claude — its crawlers, its search stack, its citation mechanics — with every claim sourced to Anthropic’s own documentation, third-party reporting, or our own server logs.

The short answer

Why Claude suddenly matters for B2B visibility

The clearest public numbers come from Goodie’s 2026 AI Search Traffic Report, which analysed GA4 referrer data from an anonymised brand panel and cross-checked it against SimilarWeb traffic across five AI platforms. Between its baseline window (May–August 2025) and March–April 2026, ChatGPT’s share of measurable AI referral traffic in the panel fell from 89.1% to 62.6% — a 26-point drop — while Claude’s share grew from 1.4% to 18.5%.

Read that carefully: it’s share of a growing pie, drawn from one brand panel, not a census of the whole web. But an engine going from a rounding error to nearly a fifth of measurable AI referrals in under a year is exactly the kind of shift that punishes single-engine strategies. Our own 5,051 first-party citation polls found that 74% of the prompts we were ever cited on were cited by only one engine — the engines simply don’t share source pools. A Claude-shaped hole in your visibility is not covered by your ChatGPT work.

We see the same shift from the server side. Our nginx bot monitoring on leadsnow.ai logged ClaudeBot more than doubling its crawl volume month-on-month into winter — 545 requests in June 2026 to 1,166 in July — alongside smaller volumes of Anthropic’s older user-agent strings. Anthropic is crawling harder, and the referral data says the answers built on those crawls are increasingly where B2B buyers are.

How Claude finds and cites pages (the verified mechanics)

The index: Brave, not Bing

Anthropic has never officially named its search provider. But when Claude’s web search launched in March 2025, Anthropic added Brave Search to its subprocessor list, researcher Simon Willison found a BraveSearchParams parameter in Claude’s search function and showed Claude’s citations matching Brave’s results for the same query, and TechCrunch reported the connection. Brave, for its part, states it serves results “from its own, built-from-scratch index” — it does not resell Google’s or Bing’s results. The practical consequence: your Google rankings and even your Bing rankings tell you nothing about whether Claude can find you. Brave’s index is a third, independent surface, and most B2B sites have never once checked their position in it.

The crawlers: three bots, three different jobs

Per Anthropic’s own help centre, three distinct user agents matter:

• ClaudeBot — collects web content “that could potentially contribute to their training”. This is the training-data crawler.
• Claude-SearchBot — “navigates the web to improve search result quality for users”. This is the search-quality crawler.
• Claude-User — fetches your page live when a user’s question triggers a web search. This is the one that fires at the moment of citation.

All three respect robots.txt, and Anthropic publishes source IPs for verification. The distinction matters because plenty of sites blocked “AI bots” wholesale in 2024–2025 to protest training-data scraping — and in doing so also blocked the agents that would have cited and linked them. If AI visibility is a goal, at minimum Claude-User and Claude-SearchBot need to be allowed, whatever you decide about ClaudeBot and training.

The citation contract: always cited, 150 characters at a time

Anthropic’s developer documentation is unusually explicit about citations: for web search, “citations are always enabled”, and every cited passage carries the source URL, title, and up to 150 characters of quoted text. Search results also carry a page_age field — when the page was last updated — which is a direct, documented freshness signal. And as of the February 2026 version of the web-search tool, Claude uses “dynamic filtering”: it writes and runs code that filters fetched search results before they reach its context window, keeping only content relevant to the question and discarding the rest.

Each of those three documented facts converts directly into a writing rule, which is what the playbook below is built on.

Claude vs ChatGPT vs Perplexity: citation behaviour compared

The overlap between engines is the part most teams underestimate — we measured it directly in our ChatGPT vs Perplexity citation-overlap study, and the short version is that each engine assembles answers from largely different source pools. Claude adds a third pool with its own index underneath. Treat it as a separate channel, not a checkbox.

The Claude citation playbook

1. Audit your robots.txt and CDN before anything else

Check that ClaudeBot, Claude-SearchBot and Claude-User are not disallowed in robots.txt — and then check your CDN and bot-management layer, because several popular presets block AI crawlers by default and never surface it to marketing. Then confirm the bots actually arrive: grep your access logs for the three user agents. If Claude-User never appears, Claude is never reading your pages at answer time, and nothing downstream matters. Anthropic honours robots.txt, so unblocking is entirely in your control.

2. Check — and win — your rankings in Brave Search

Search Brave (search.brave.com) for the 20–40 buyer prompts you track elsewhere. Brave’s independent index means your positions there can differ sharply from Google’s: pages that rank via Google-specific authority signals may be nowhere, and clean, fast, well-structured pages on smaller domains often do comparatively better. There’s no submission console to game; Brave builds its index from its own crawling. The levers are the boring ones — crawlable HTML, working sitemaps, real inbound links, and pages that actually answer the query — applied to an index most of your competitors have never looked at.

3. Write claims that survive a 150-character lift

Claude cites by quoting short passages — up to 150 characters per citation, per Anthropic’s docs. That’s roughly the length of the sentence you’re reading right now. So every fact you want repeated needs to be complete inside one sentence: subject, number, date, source. “We increased leads by 43%” survives a lift; a claim spread across three sentences of build-up doesn’t. Lead every page and every section with the extractable version, then elaborate. This is the same answer-capsule discipline from our breakdown of what ChatGPT actually cites, but Claude’s documented snippet limit makes the sentence-level version of it measurable: read your key claim, count the characters, and if the claim doesn’t stand alone under 150, rewrite it.

4. Make freshness visible and real

Claude’s search results carry a documented page_age field — when the page was last updated. Put a visible, honest “last updated” date on evergreen commercial pages, keep it current by actually updating the content, and make sure your sitemap lastmod dates match. A 2024-dated page competing for a 2026 prompt starts behind regardless of quality.

5. Write sections that survive dynamic filtering

Since February 2026, Claude filters fetched content programmatically before reading it — only passages relevant to the question make it into the model’s context. Pages that interleave the answer with digressions, sprawling preambles and off-topic asides risk having the connective tissue stripped out. The defence is structural: question-shaped headings, one topic per section, the answer at the top of each section, and no burying a key fact inside an anecdote. Write so that any single section, extracted alone, still makes your case.

6. Cover the prompts where Claude has to search

Anthropic’s docs are clear that Claude only searches when a question depends on current, changing, or niche information — established general knowledge gets answered from training, with no citations and no visibility for you. So target the prompt classes that force retrieval: pricing (“what does X cost in 2026”), comparisons, “best X for Y” shortlists, recent data, and anything localised. These are also, conveniently, the buying-intent prompts. Our agency-focused ChatGPT playbook covers how third-party surfaces win the category-level versions of these prompts; the same logic applies here, filtered through what ranks in Brave rather than what ranks in Bing.

7. Measure it like a channel

Three cheap instruments: a GA4 referrer segment for claude.ai (this is exactly the data Goodie’s report is built on), a monthly grep of your server logs for the three Anthropic user agents, and your existing fixed prompt panel run against Claude with web search on. Baseline before you change anything. We run this loop on our own domain across every engine we track — the full method is in our first-party citation data write-up — and the single biggest lesson from 5,051 polls is that per-engine results diverge enough that unmeasured engines are effectively unmanaged engines.

FAQ

Does Claude really use Brave Search?

Anthropic has never officially named its provider, but the public evidence is strong: when web search launched in March 2025, Anthropic added Brave Search to its subprocessor list, and researcher Simon Willison documented a BraveSearchParams parameter in Claude’s search function and showed Claude’s citations exactly matching Brave’s results for the same query. TechCrunch reported the same connection. Treat “rank in Brave” as the working assumption until Anthropic says otherwise.

Should I block ClaudeBot in robots.txt?

That’s a policy decision about training data, but know what each bot does before blocking by pattern. Per Anthropic’s help centre, ClaudeBot collects content for model training, while Claude-SearchBot improves search quality and Claude-User fetches pages live when a user’s question triggers a search. Blocking Claude-User or Claude-SearchBot removes you from Claude’s cited answers directly. If you want AI visibility with training-data reservations, the granular play is to allow the two search-related agents and decide on ClaudeBot separately.

How do I know if Claude is already citing or crawling my site?

Server-side, search your access logs for the user agents ClaudeBot, Claude-SearchBot and Claude-User — on our own site these are logged daily, and July 2026 alone saw over a thousand ClaudeBot requests. Demand-side, build a GA4 segment for referral traffic from claude.ai and watch its trend. And directly: run your buyer prompts in Claude with web search enabled and record which domains it cites, the same panel discipline we use for every other engine.

How long does it take to get cited by Claude?

There’s no reliable public benchmark for Claude specifically yet, and we won’t invent one. What we can say from our own cross-engine tracking: purpose-built pages targeting one specific buyer prompt have earned first citations on search-led engines within weeks, while head-term category prompts take months of compounding third-party presence. Claude’s pipeline — Brave ranking plus live fetch — most resembles the search-led engines, so the sensible expectation is: long-tail first, category terms later, and nothing at all until its crawlers can reach you.

Where this fits

Claude is the newest surface in the same game we’ve been publishing data on all year: separate engines, separate indexes, separate playbooks, one measurement discipline. We run AI lead generation end to end — 50,769+ AI-booked sales appointments since 2017 and 1M+ leads generated — and we track our own AI citations daily across every engine that matters, now including Claude. If you want your brand showing up in AI answers while your competitors are still optimising for ten blue links, book a call.