Claude is the engine your buyers use.

Chapter 01

Why Claude matters far more than its traffic share suggests

If you look at consumer usage, Claude is a rounding error next to ChatGPT. If you look at where B2B buying research actually happens, it's the second most important surface on the internet — and the gap between those two facts is the whole reason this page exists.

In Goodie's 2026 AI search traffic study, Claude's share of measurable B2B AI referrals rose from 1.4% to 18.5% in eight months, while ChatGPT's fell from 89.1% to 62.6%. In the same period Claude accounted for roughly 1.3% of AI platform visits. It produces about fourteen times more B2B referral share than its visit share implies.

The reason is who's on the other end. Anthropic reports 300,000+ business customers. Menlo Ventures puts Anthropic at roughly 40% of enterprise LLM API spend, ahead of OpenAI. Claude sits inside a large share of the Fortune 100 and holds an outsized slice of the enterprise-assistant market. When a director of finance researches vendors on a work laptop, there's a meaningful chance she's doing it in Claude.

The traffic is small. The traffic is also excellent.

Claude referrals arrive with unusually high engagement — around 65% engagement rate and roughly 55 seconds of engagement time in referral datasets. Ruler Analytics' 2026 benchmark puts AI referral traffic overall at a 5.8% conversion rate, above every established organic channel. A visitor who arrives from Claude has already asked a specific question, been given a short, hedged, carefully-sourced answer, and clicked one of three links. That's not a cold click. It's closer to a warm referral.

Chapter 02

The two doors into a Claude answer

Almost every guide about AI search assumes one mechanism: the model searches, retrieves pages, cites them. For Claude that's only half true, and the half people miss is the half that decides most answers.

Claude leans on its training corpus more than its rivals and invokes web search more selectively. For conceptual, definitional and general-knowledge questions, it often answers without searching at all — which means no citations, no links, and a recommendation shaped entirely by what the model absorbed during training. For recency-dependent questions, explicit search requests, and prompts that ask for sources, it searches and cites.

Both doors matter, and they need different work.

Test which door your category uses

Ask Claude ten of your buyers' real questions and watch for the search indicator. Log which prompts trigger a search and which get answered from memory. That split tells you where to spend:

• Mostly searched (pricing, comparisons, "best X in 2026") → prioritize retrievability, freshness and index visibility. Chapters 3–6.
• Mostly unsearched (what your category is, what your company does, who it's for) → prioritize corpus presence and third-party description. Chapter 7.

Chapter 03

Anthropic runs four crawlers. Most sites only think about one.

In early 2026 Anthropic split its crawler documentation into separate user agents with separate jobs and separate consequences for blocking. This is the single most common technical mistake I find on Claude audits: a site blocks "the AI bot," believes it has made one decision, and has actually made three.

Blocking ClaudeBot does not block Claude-SearchBot. Blocking Claude-SearchBot does not block Claude-User. Each needs its own directive, and each closes a different door.

Anthropic states that all of its bots respect robots.txt, including the user-initiated fetcher — a stricter position than OpenAI or Perplexity take on their equivalents. The legacy anthropic-ai and claude-web agents are deprecated but still worth a line for old traffic.

# Anthropic crawlers — allow all four, protect the same paths you always do
User-agent: ClaudeBot
User-agent: Claude-SearchBot
User-agent: Claude-User
User-agent: claude-code
User-agent: anthropic-ai
Allow: /
Disallow: /admin
Disallow: /api/

Sitemap: https://example.com/sitemap.xml

Render server-side, because Claude reads text

Claude's retrieval path works from indexed and cached text, not from a rendered browser session. Anything injected client-side — pricing tables, feature grids, tabbed FAQs — is effectively invisible. Fetch your own key pages with a plain HTTP client and read what comes back. If the number a buyer needs isn't in that response, no amount of content strategy will fix it.

Chapter 04

The Brave problem

Here is the fact that reorders most Claude strategies: the search index behind Claude's web results does not appear to be Google's or Bing's. Multiple independent lines of evidence point to Brave Search.

The evidence chain: Anthropic added Brave Search to Claude's published subprocessor list when web search launched; Claude's search implementation contains a Brave-specific parameter object; and independent testing has found Claude's citations overlapping Brave's top organic results roughly 87% of the time. Brave has said publicly that its Search API supplies most of the top-10 LLMs, and that for some it's the only index behind their answers.

What follows from it

• Your Google rank is not your Claude eligibility. Two different indexes with different crawl coverage and different ranking logic. Sites that dominate Google can be thin or absent in Brave.
• Check Brave directly. Run your twenty highest-intent buyer queries in Brave Search and record whether you appear in the top ten. That's a free, five-minute proxy for Claude retrieval eligibility.
• There's no submission tool. Brave offers no webmaster console. Its crawler broadly follows the same rules Googlebot does, so the fix is standard hygiene: crawlable, fast, linked-to, not blocked.
• Inbound links do double duty. Independent indexes lean harder on discovery through links. Digital PR that earns you coverage on well-crawled industry sites improves your Brave coverage as a side effect.

Chapter 05

What Claude actually rewards

Claude cites fewer sources than any other major engine, and it puts them in the body of the answer rather than a long footnote list. That scarcity is the defining constraint. You are not competing for a slot in a list of fifteen. You are competing to be one of three.

It cross-checks before it recommends

The behaviour practitioners report most consistently: Claude is reluctant to repeat a company's self-description if it can't find corroboration elsewhere. Across engines, most AI citations already come from third-party sources rather than brand-owned sites, and Claude sits at the strict end of that distribution. A perfect landing page with no independent coverage behind it produces a hedged answer, or someone else's name.

It prefers the primary source

If you summarize a study and the study itself is reachable, Claude tends to cite the study. The implication is uncomfortable but useful: aggregating other people's numbers is a weak citation strategy. Producing numbers nobody else has is the strongest one. Your own product data, a customer survey, an aggregate benchmark from your user base — published with a real methodology — makes you the primary source that gets cited instead.

Its source mix is unusual

Claude has the lowest social citation rate of the major engines — under 4%, against roughly 15% for Google AI Overviews and 11% for Perplexity. It leans instead toward documentation, practitioner explainers and research-style writing. Two practical consequences:

The bar, stated plainly

Write the page a careful analyst would be comfortable forwarding to their CFO: a direct answer up top, depth underneath, a named author, primary data, clean markup, and independent coverage backing the claims. Every engine likes that page. Claude requires it.

Chapter 06

The pages Claude actually cites

Analyses of cited-URL sets show clear patterns. Directional rather than definitive, but they line up with what I see in client tracking.

• Blog paths dominate. In one 2026 analysis of Claude-cited URLs, over half sat under a /blog/ path. Product pages get cited far less often than the content that explains the category.
• Listicle and comparison paths overperform. Close to half of cited URLs used /best-…, /top-…, /alternatives-… or /vs-… structures.
• Year tokens signal freshness. Roughly a quarter of cited URLs contained a year. Research on generative engines has found recency filtering produces a 2–3x higher citation rate for content published in the last twelve months.

The answer-capsule pattern

Every section should open with a complete, self-contained answer in two or three sentences, with the subject named rather than implied. Retrieval works on chunks; a chunk that depends on the paragraph above it can't be quoted safely.

Add the evidence Claude looks for

1. Date your facts in the text. "As of July 2026" inside the sentence, not only in schema. It gives a recency-sensitive model something to trust.
2. Name your sources. Pages that cite credible sources get cited more themselves — the original generative-engine research found citation and statistic density measurably lifted source visibility.
3. Use real HTML tables for comparisons. They survive chunking intact; screenshots of tables do not.
4. State the tradeoff. A comparison page that admits where a competitor wins reads as analysis rather than marketing, and analysis is what gets quoted.

Chapter 07

The training-corpus play

This is the door nobody optimizes for, because it's slow, unattributed and impossible to A/B test. It also determines what Claude says about you when it doesn't search — which, for category and vendor questions, is often.

You can't inject yourself into a training corpus. What you can do is make sure that when the corpus was assembled, the internet described you clearly, consistently and in more than one place.

One sentence, everywhere

Models reconcile a picture of your company from everything they've read. Conflicting descriptions across your homepage, LinkedIn, G2 and Crunchbase produce a blurry entity, and blurry entities don't get recommended. Write one 25-word description and make it identical on every surface you control.

Get described by other people

The corpus contains far more third-party writing about companies than first-party writing by them. Bylined features, podcast transcripts, conference listings, partner directories, integration marketplaces, review profiles — each one is another place the model can learn what you do from a source with no stake in the answer.

Own your terminology

If you've named a method, a framework or a product concept, use the name consistently and define it in one canonical, well-linked place. Distinctive terminology is easier for a model to attach to an entity than generic category language, and it gives you a query where you're the only credible answer.

Publish for the long horizon

Content published now can influence answers for years. That reframes evergreen explainers from a traffic play into a positioning play: you're not writing for this quarter's sessions, you're writing the description of your category that a model will absorb and repeat.

Chapter 08

Measuring Claude specifically

Claude needs its own tracking, because it behaves differently enough from other engines that a blended "AI traffic" number will hide what's happening.

Chapter 09

The 60-day Claude plan

Sequenced so that nothing you build gets wasted on a door that's still shut.

Chapter 10

Seven mistakes specific to Claude

1. Blocking one agent and assuming you blocked them all. Four crawlers, four decisions, four separate directives.
2. Auditing robots.txt but not the WAF. Bot-management defaults override your file, silently.
3. Using Google rank as the proxy. The retrieval index behind Claude is very likely not Google's. Check Brave.
4. Optimizing only the searched door. Half your category's questions never trigger a search, and those answers come from the corpus.
5. Publishing self-description with no corroboration. Claude is the engine most likely to decline to repeat a claim only you have made.
6. Summarizing other people's data. The primary source gets the citation. Produce your own numbers.
7. Judging the channel on volume. Small traffic, high intent, enterprise buyers. Measure pipeline.

Chapter 11

Proof this works

Numbers from client analytics and citation tracking. The approach above is what produced them.

Questions

FAQ