Claude SEO: How B2B Companies Can Win Visibility in Claude and AI Search

Claude SEO is the practice of structuring content, technical infrastructure, and brand signals so that Claude AI cites a company in its responses. 

Claude.ai has grown from about 89 million monthly visits in January 2025 to about 288 million by February 2026, according to Semrush traffic estimates. A year ago, it was primarily a productivity tool for writing and coding, but this has changed. According to Omnius datasets covering 11,000+ domains, mostly AI, SaaS, and fintech websites, Claude is now the second-largest AI traffic source for B2B companies:

• It captures over 21% of AI-referred sessions.
• Its QoQ growth exceeds 76.2%.
• It has the highest engagement of all LLMs, with an average session of 3:27.
• Its user base skews heavily toward enterprise buyers, engineers, and technical decision-makers rather than casual browsers.

In the same data, Claude sessions show a 16.8% conversion rate and $4.56 average session value, the highest of any AI search engine and roughly 6× above Google organic.

In TrustRadius’ 2026 B2B Buying Disconnect report, 63% of 1,862 technology buyers used AI during their purchase journey (TrustRadius, 2026), and many of them will meet Claude while comparing vendors. This guide uses Omnius research data and Anthropic’s documentation to show what is known about how Claude finds sources and how a company can end up in its citations.

Key Takeaways

• Claude distributes visibility differently from the other major AI engines. It spreads citations across mid-market domains. Our model comparison data shows gaps exceeding 40 percentage points on the same domain. 
• Anthropic documents Claude’s search inputs, not its ranking. It lists Brave Search, TurboPuffer, and Parallel Web Systems as web search providers. Its developer guide to hybrid retrieval shows why pages should match both meaning and exact terms, the two signals semantic and BM25 search score. 
• Being cited on Claude starts with four layers. Fix crawler access and check indexing in Brave Search first. Restructure content into self-contained passages, name the author on every article and support it with schema markup, and build third-party corroboration across independent sources.
• Measure visibility to understand a brand's position and optimize accordingly. Track crawler access, citation flow rate, prompt coverage, share of voice, and citation sentiment. 

Claude Distributes Visibility Differently From Other LLMs

We’ve found that Claude and other engines, such as ChatGPT, Gemini, and Perplexity, approach visibility in distinct ways across every segment in our dataset:

‍

ChatGPT and other models concentrate visibility among a few high-authority domains, while Claude distributes exposure across mid-market domains. The two pools rarely overlap. A domain can receive a 5–10% visibility score with Claude while scoring zero with another model. Omnius datasets spanning 11,000+ domains show gaps between Claude and ChatGPT on the same domain that often exceed 40 percentage points.

One likely reason is the search stack, although Anthropic does not explain how Claude chooses sources. Its first listed web search provider, Brave Search (since March 2025), runs its own independent index, separate from Google and Bing. TurboPuffer (May 2026) and Parallel Web Systems (October 2026) have since joined it. 

Which domains Claude cites varies by prompt. Tracking a fixed set of customer prompts shows which page types it cites for a B2B SaaS company, such as documentation or technical articles.

B2B companies that can't compete with enterprise domains on Google aren't locked out on Claude. In Omnius datasets, mid-market domains earn Claude visibility that they rarely reach on Google, which makes Claude a genuine opportunity for mid-market brands willing to invest in structured, evidence-backed content.

Why Most Pages Are Invisible To Claude

In July 2025, Anthropic’s crawlers requested about 38,000 pages for every visitor Claude referred to websites on Cloudflare’s network, down from about 287,000 in January (Cloudflare). The ratio includes training crawls, yet the gap is plain: most crawled pages never send a reader back.

The phrase "ranking on Claude" is misleading. There’s no global position, no page 1, and no slot to hold. Claude decides when a question needs a web search, reads the results, and cites the passages it used, so the sources change with every query. The durable advantage is content that reliably answers the exact question better than the alternatives.

A useful working model, Omnius’s rather than Anthropic’s, has five checks. A page can be reached by Claude’s crawlers, appears in the results Claude gets, matches the question, holds a passage worth quoting, and looks trustworthy next to other sources. Anthropic documents only the controls behind the first check and the citation format behind the fourth.

How Claude’s Retrieval System Works

When Claude decides a question needs current information, it searches the web and answers with citations. Anthropic’s web search documentation covers what goes into and comes out of that search, not how Claude ranks results. Anthropic’s developer guides show how a retrieval pipeline scores text, with building blocks that other AI search engines, such as Perplexity, also describe. The steps below keep the two apart.

1. Retrieval

Anthropic’s Trust Center lists three web search subprocessors: Brave Search (March 2025), TurboPuffer (May 2026), and Parallel Web Systems (October 2026). Its help center adds that image results are powered by Bing. Brave also has outside evidence behind it: in a study of about 400 queries presented in June 2026, Profound, an AI-visibility software company, found that 79.2% of Claude’s cited URLs also appeared in Brave’s top 10. Profound calls this an observed match rather than proof and has not published its full method. 

A page ranking #1 on Google can still be missing from Claude’s answers if the providers Claude searches have not indexed it, so a search on Brave for key commercial queries is a quick first test.

Three Crawlers

Anthropic operates three distinct bots: ClaudeBot, Claude-SearchBot, and Claude-User.

‍

Each block has a different effect. For visibility in Claude’s answers, Claude-SearchBot and Claude-User matter most, because Anthropic says blocking them may reduce a website’s visibility in Claude’s search results and stops Claude fetching its pages when a person asks.

2. Scoring

Anthropic does not publish how Claude scores web results. Its Contextual Retrieval guide, written for developers who search their own documents, does publish a scoring recipe that runs two searches side by side. Semantic search compares embeddings, numeric representations of meaning, so it finds passages that say the same thing in other words. BM25 keyword search scores exact term matches and weights rare terms more, so it finds the precise words, names, and codes in a query. In Anthropic’s example, a support query for error code TS-999 gets general error-code content from embeddings and the exact TS-999 match from BM25.

The two ranked lists are merged with rank fusion and deduplicated. The companion cookbook scores each chunk by 1 divided by its rank in each list and weights semantic search at 80% and BM25 at 20%, as defaults to tune. A reranker then scores the top 150 chunks against the query and keeps 20 for the model. 

Perplexity describes the same mix in its own engine, which does not force a choice between lexical and semantic retrieval. The lesson for a B2B page holds for any engine built this way: explain the concept plainly for semantic matching, and name the exact product, category, and feature terms a buyer types, since a passage that writes around a term gives a keyword index nothing to match.

3. Reranking

In the same recipe, reranking is a separate second stage, built differently from first-pass retrieval: 

• First-pass retrieval scores the query and each chunk separately, by comparing embeddings or by counting terms with BM25. This is fast enough to scan a whole index, but it only captures general relevance.
• Reranking reads the query and each shortlisted chunk together, which lets it judge how well a specific passage answers the specific question. It is more accurate but too slow and costly to run on every document, so it scores only the shortlist. That is one reason two pages on the same subject can fare differently on the same query.

Note: Anthropic documents this retrieve-then-rerank pattern for developers, and its guide uses Cohere’s reranker. It does not say whether or how Claude’s web search reranks results.

Passage Design Matters

Claude’s citations point to passages, each quoting up to 150 characters of the page. Pipelines like Anthropic’s recipe also split documents into chunks and score each one on its own, so every section of a page competes independently, and a passage that leans on surrounding paragraphs for meaning is harder to match. 

Standard retrieval-augmented generation (RAG) has a documented failure mode that exacerbates the issue: chunks extracted from documents lose their surrounding context. Anthropic's Contextual Retrieval approach addresses this by prepending a short, chunk-specific explanation of where the chunk sits in its document before embedding it and before building the BM25 index.

On Anthropic’s test sets (codebases, fiction, arXiv papers, and science papers), contextual embeddings cut the share of relevant chunks missing from the top 20 results by 35%. Adding contextual BM25 raised the cut to 49%, and reranking raised it to 67%, from 5.7% to 1.9%. For a website the lesson is simple: a passage that names its subject, product, and timeframe carries its own context.

Position within the page matters too. In Kevin Indig’s analysis of 18,012 verified ChatGPT citations (Growth Memo, February 2026), 44.2% pointed to the first 30% of a page’s text, 31.1% to the middle, and 24.7% to the last third. ChatGPT is not Claude, but the pattern argues against spending the top of a section or article on setup, transitions, or filler.

4. Answering

Anthropic trains Claude with a published constitution that orders four values: broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. The constitution calls this prioritization holistic rather than strict. It governs how Claude behaves, and Anthropic publishes no rule that scores web pages by tone. 

The closest published text on sources is about skepticism. The constitution asks for appropriate skepticism toward content from low-quality or unreliable websites, and the example prompts Anthropic open-sourced for its Research subagents list marketing language for a product and spin language as signs of a problematic source. On tone itself, Wan, Wallace, and Klein (ACL 2024) found that models judged evidence mainly by relevance and largely ignored stylistic features such as a neutral tone. Benefits and superlatives without evidence give a skeptical reader, human or model, little to check.

‍

The favorable example wins on substance: it names the audience, the functions, and the limits, so every claim can be checked.

Claude Research: A Higher Bar

Anthropic’s Research feature uses an orchestrator-worker architecture. A LeadResearcher agent plans a strategy, spawns parallel subagents, and refines its searches based on what each returns, so Research adapts to new findings instead of running a single search.

Citations in Research mode work differently. A distinct CitationAgent runs after the LeadResearcher synthesizes findings, auditing the completed report and mapping specific claims back to source locations. Citations are a post-synthesis verification pass, so a source cited in Research has been read during a longer search and matched to a specific claim in the finished report.

Anthropic reports that multi-agent systems like Research use about 15× more tokens than chats. Its prompts scale the effort to the task: simple fact-finding runs one agent with 3–10 tool calls, while complex research deploys more than 10 subagents. In B2B, that depth suits vendor shortlists that need many sources. Our guide to agentic SEO follows one B2B software research task from its requirements through evidence checks to a shortlist.

7 Factors That Influence Claude Visibility

Seven properties make a passage easier to find, trust, and cite across the five checks above. Anthropic does not publish ranking factors, so each section names the evidence it rests on:

1. Factual density
2. Author credibility
3. Primary source attribution
4. Content freshness
5. Corroborating source density
6. Technical parseability
7. Domain authority

1. Factual Density

In the GEO study (Aggarwal et al., KDD 2024), adding quotations raised a source’s share of a generated answer by about 41%, adding statistics by about 31%, and citing sources by about 27%, measured by position-adjusted word count on a test engine built on GPT-3.5. A later benchmark, C-SEO Bench (NeurIPS 2025), found most such edits largely ineffective and often harmful to ranking, so verifiable facts earn their place through accuracy first. 

What to do: Make sure every substantive claim in a passage can be verified by a third party without reading the surrounding context. Replace subjective benefit statements with specific, verifiable claims. 

2. Author Credibility

Anthropic does not list authorship as a signal, and the widely shared figures on author credentials and AI citations have no published method. A named author still tells any reader, human or model, who stands behind a claim. Google’s helpful-content guidance asks whether it is self-evident who wrote a page, and an anonymous byline fails that test.

What to do: Attribute every piece of content to an individual with relevant expertise. Create a dedicated author bio page for each contributor, listing their role and credentials. Include institutional affiliation, such as company name, title, and relevant certifications. Link to the bio from every article they write using Article + Person schema.

3. Primary Source Attribution

Anthropic’s open-sourced Research prompts list news aggregators used in place of the original source as a warning sign, and one example task says to prioritize original sources over news aggregators. A named, dated, linked source lets a reader or a model follow a claim to its origin, while "a recent study" leads nowhere. 

What to do: Provide a named source, a date, and, where possible, a direct link to the primary document for every data point.

4. Content Freshness

Each result from Claude’s web search tool carries a page_age field, the date the website was last updated, so the date reaches the model; Anthropic does not say how much it counts. Across seven AI search platforms, Ahrefs’ analysis of 17 million citations (July 2025) found that cited URLs averaged 1,064 days old, against 1,432 days for organic results, which makes them 25.7% fresher.

What to do: Review key pages at least every three months, update them when facts change, and make sure the page’s updatedAt metadata and visible date reflect the real change.

5. Corroborating Source Density

Anthropic does not document a corroboration check in Claude’s web search, but research points that way: Xie et al. (ICLR 2024) found that when retrieved evidence conflicts, language models generally side with the majority of it. Ahrefs’ study of 75,000 brands also found branded web mentions correlated with visibility in Google’s AI Overviews at 0.664, far above backlinks at 0.218.

What to do: Build corroboration for key claims across independent sources. Contribute proprietary data to industry reports and analyst briefings. Co-publish case studies with customers so the results appear on both domains. Write for trade publications that cover the category. Pursue reviews on G2, Capterra, and relevant aggregators. 

6. Technical Parseability

Anthropic’s documentation says Claude’s web fetch tool does not support websites dynamically rendered with JavaScript. Controlled tests agree. In Vercel’s 190 fetch probes (August 2026), a client without JavaScript could not retrieve an answer available only through JavaScript, and Writesonic’s test page (March 2026) showed ChatGPT, Claude, and Gemini fetching pages without running JavaScript. Content rendered in the browser may reach such a fetcher as an empty page. 

What to do: Load key pages in a browser with JavaScript disabled. Content that disappears is likely invisible to fetchers that do not render JavaScript. Ensure that all high-value pages, such as landing pages, product documentation, comparison pages, and FAQ sections, are served as static HTML or use server-side rendering.

7. Domain Authority

Two SE Ranking studies measured authority on other engines; neither covers Claude. In Google AI Mode (December 2025, 2.3 million cited pages), domain traffic was the strongest predictor of a domain’s citations, with a SHAP value of 0.63. In ChatGPT (November 2025, 129,000 domains), referring domains were the top predictor: domains with more than 350,000 referring domains averaged 8.4 citations, against 1.6 to 1.8 for domains with up to 2,500. Authority is a floor, not a guarantee.

What to do: Build domain authority through consistent earned media, backlink acquisition from industry-relevant publications, and digital PR campaigns that generate coverage on high-authority news websites. Publish original research, proprietary benchmarks, and data studies that naturally attract inbound links. Track the backlink profile and focus on link quality over volume.

How To “Rank” on Claude as a B2B Business: A Four-Layer Approach

The goal: be easy to retrieve, easy to match, easy to cite, and hard to distrust. 

Claude SEO operates upstream and downstream simultaneously. In the early stages, content needs to be discoverable by the search providers Claude uses, starting with Brave Search, and technically accessible to Claude's crawlers. Later on, content needs to hold up against the other sources for the exact question asked by being precise, attributable, and directly useful. 

Anthropic documents the controls behind layer 1 (its crawlers) and the passage citations behind layer 2, though not how Claude selects passages. It does not document layer 3 as a Claude signal. Layer 4 rests on outside studies and on Anthropic’s account of building Research, where early agents “consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources like academic PDFs or personal blogs,” until the team added source-quality heuristics to the prompts.

Layer 1: Fix the Silent Failures First

Access problems are the first thing to rule out, because a page Claude cannot fetch cannot be cited.

Ahrefs’ May 2025 robots.txt analysis of about 140 million websites found that ClaudeBot had the fastest growth in block rates of any AI bot, up 32.67% over the prior year. Among the top 50 US and top 50 UK news websites, a BuzzStream study found that 79% block at least one AI training bot and 71% block at least one live-retrieval bot. 

Blocks can also be accidental. A security team may write a robots.txt wildcard or Cloudflare rule to block GPTBot or ClaudeBot and prevent AI training data collection. If that rule also catches Claude-SearchBot and Claude-User, it can keep the website’s pages out of Claude’s answers. No error message reports it, so a team may not notice. 

In robots.txt, explicitly allow all three bots:

User-agent: ClaudeBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: Claude-User
Allow: /

Audit Cloudflare separately. On 1 July 2025, Cloudflare changed its default to block AI crawlers unless they pay for content, and its bot settings, such as Bot Fight Mode, can stop crawlers that robots.txt allows. Check server logs 48–72 hours after any configuration change to confirm the crawlers are getting through.

Check Brave Search indexing by searching search.brave.com for key commercial queries. Serve high-value pages as static HTML or with server-side rendering, and keep them fast: no AI company publishes its crawler timeouts, so a slow page is an avoidable risk.

Layer 2: Compete With Every Passage

Claude’s citations quote up to 150 characters of a page, and chunked retrieval scores each section on its own. Engines that describe their pipelines ask for the same thing: Microsoft recommends sentences that make sense even when pulled out of context, and Perplexity decomposes documents into self-contained spans. A strong section can be cited even from an otherwise average page.

Open every section with a direct answer that names the entity, the claim, and the scope. Make every passage individually interpretable: 

• Repeat the subject noun rather than using pronouns.
• Include the entity name and date within the passage.
• Avoid cross-references to adjacent sections.
• Give each section at least one verifiable, attributed data point. 

In a retrieve-then-rerank pipeline like Anthropic’s recipe, the reranker scores passages against the exact query and BM25 rewards its exact terms. Mirror the phrasing of real purchase-intent queries instead of writing for a general audience.

Layer 3: Implement Schemas

Every article should name its author and link a biography that states the author’s role and expertise in the text. The markup then restates what each page shows, using the types Google documents for each kind of page:

• Article + Person for all editorial content
• Organization + sameAs links for company pages
• SoftwareApplication for product pages

In Ahrefs’ May 2026 study, pages that AI answers already cited showed no measurable citation gain after adding JSON-LD, against similar pages without it. The study tracked ChatGPT, AI Mode and AI Overviews, and our guide to schema markup for AI search explains where markup does its work.

Layer 4: Build Citation Authority

Earned media is a primary channel for AI visibility. Muck Rack’s Generative Pulse found that earned media made up 89% of the links cited by ChatGPT, Claude, and Gemini in July 2025, and 84% of more than 25 million cited links in its May 2026 edition. Across its editions, the share has ranged from 82% to 89%.

B2B brands can build third-party corroboration by publishing technical walkthroughs on YouTube, taking part in the communities where their buyers compare tools, and earning the independent coverage that Wikipedia’s notability guideline requires before a company qualifies for an article.

Original data is the strongest of these assets, because a summary can be rebuilt from its sources while unique data cannot. Benchmarks, proprietary datasets, and case studies with methodology and named outcomes offer more reasons for citation than aggregated content. 

Measuring Claude Visibility

There is no native Claude analytics tool. Measuring visibility requires assembling multiple signals manually.

To track Claude visibility, build a set of 20–30 prompts that reflect how buyers actually research solutions. Run them across Claude monthly and record whether the brand appears, how it's described, and which competitors show up alongside it. This works manually or with an automated solution like Atomic AGI.

These five signals give a working picture of where a brand stands with Claude, what needs fixing, and what to focus on first:

1. Crawler access
2. Citation flow rate
3. Prompt coverage
4. AI share of voice
5. Citation sentiment

1. Crawler Access

Crawler access shows whether Claude bots are actively crawling priority pages. It’s the signal to check first, before investing in any content changes. If the crawlers aren't reaching the pages, every other metric is irrelevant.

How to measure: Filter the web server logs for the ClaudeBot, Claude-SearchBot, and Claude-User agents. Server logs record every bot visit to a website, including which pages were requested, when, and by which crawler. 

2. Citation Flow Rate

The citation flow rate (CFR) is the ratio of AI citations to search impressions, calculated as AI Citations ÷ Search Impressions × 100. It shows how effectively existing search visibility converts into Claude citations.

A page can rank well on Google and receive thousands of impressions but produce zero Claude citations. CFR exposes that gap. 

How to measure: Calculate CFR per page and per query cluster, not as a website-wide average. A website-wide number will mask pages that perform well and pages that are completely invisible.

3. Prompt Coverage

Prompt coverage is the share of relevant purchase-intent queries that surface a brand. This is the closest equivalent to keyword tracking in traditional SEO, but instead of a Google position, it checks whether Claude mentions the brand at all.

How to measure: Include the following prompts in the tracking set, run them monthly, and note how often the brand appears:

• Category queries: best [category] tools for [use case]
• Direct comparisons: [product] vs [competitor]
• Problem-oriented queries: how to solve [problem the product addresses]

4. AI Share of Voice

Share of voice (SoV) reveals how much of the conversation Claude gives to a brand versus its competitors for a given topic or category. We calculate it as brand mentions ÷ total market mentions × 100.

How to measure: Track two variants separately. Entity-based SoV measures brand appearance as a named recommendation and maps directly to business outcomes. Citation-based SoV measures brand content cited as a source. It’s a leading indicator of future entity SoV. Citation-based SoV is what gets built first. Entity SoV follows.

5. Citation Sentiment 

Citation sentiment shows how Claude frames a brand when mentioning it. There's a difference between Claude recommending a product as a first choice, describing it neutrally as one of the options, and listing it at the bottom of an alternatives list.

How to measure: Document how and where the brand appears in Claude’s responses to relevant queries, and track changes month-over-month. Track positioning categories like recommended, described, and listed, or sentiment categories like positive, neutral, and negative.

Measuring Issues To Consider

A common problem is undercounting. The session is registered as direct traffic in GA4 when users follow a Claude recommendation but arrive through Google search, by typing the URL, or via a bookmark. Referral tracking therefore understates Claude’s contribution. 

There’s also error exposure. A brand that does not monitor Claude’s answers learns about a factual error only when a buyer repeats it, and Claude can keep repeating the error until then.

Still, few brands track it. In a September 2025 McKinsey survey of about 30 Fortune 500 consumer-brand CMOs, 16% said their brand systematically tracks AI search performance. Without that tracking, a team can't see which pages LLMs cite, which queries trigger mentions, or whether answers carry inaccurate information about the brand, and content and infrastructure improvements become guesswork.

Solve for Claude. The Rest Follows.

The work that suits Claude, verifiable facts, named authors, primary sources, third-party corroboration, and pages any fetcher can read, is the same work the studies above link to citations on other AI platforms. 

Most B2B companies don't have the cross-functional infrastructure to run this alone. At Omnius, we handle the full B2B GEO stack for SaaS, Fintech, and AI companies, helping them become market leaders through organic positioning. Reach out to take a larger share of AI search.

Frequently Asked Questions

How does Claude get its data?

Claude’s baseline knowledge comes from training data, which includes public web content collected by ClaudeBot. For current information, Claude searches the web, and Anthropic lists Brave Search, TurboPuffer, and Parallel Web Systems as its web search subprocessors. Claude-SearchBot crawls pages to improve those results, and Claude-User fetches pages when a person asks Claude a question. Anthropic does not publish how Claude ranks what it retrieves. 

How long does it take to get cited by Claude?

There’s no published timeline, and it depends on the starting position and priorities. Two clocks run separately.

For live search, a page can be cited once the search providers Claude uses have indexed it and it answers the question better than the alternatives. For Claude’s built-in knowledge, a page can only appear after Anthropic trains a model on data that includes it, and each model’s documentation lists its training data cutoff.

Is Claude actually the best AI?

That depends on what is measured. When it comes to B2B search visibility, Claude achieves the highest session value and conversion rate among AI engines (Omnius, 2026). Its user base skews toward enterprise buyers and technical decision-makers, which makes it the most commercially valuable AI traffic source for SaaS and Fintech companies. Claude also distributes visibility more broadly than ChatGPT or Perplexity, giving mid-market domains a realistic chance of being cited. 

Whether it's the "best AI" in general is a different question. But for B2B companies focused on being found during vendor research, it is currently the highest-value channel. 

Author’s Disclaimers

This article separates what Anthropic documents about Claude’s web search, what Anthropic’s developer guides describe about retrieval in general, and what outside studies measure, mostly on other AI engines. It is not a map of Claude’s ranking system, because Anthropic does not publish one.

Anthropic has not published a ranking algorithm, signal weightings, or quality rater guidelines equivalent to Google’s. Where a section explains semantic search, BM25, rank fusion, or reranking, the source is Anthropic’s Contextual Retrieval guide for developers, and the section says so. Anthropic will keep updating Claude’s search without announcing every change. 

The signals that matter today, such as factual density, author credibility, indexability, and third-party corroboration, rest on evidence that is unlikely to reverse, even as the weights and thresholds shift. Reviewed in October 2026.