Claude SEO: How Claude Finds and Cites Sources

Claude SEO means getting your pages retrieved, read and cited by Anthropic's Claude. In practice that means two jobs: get your pages into the sources Claude's web search returns, and write passages that answer a prompt cleanly enough for Claude to quote and link them. Ranking helps with the first job. It does nothing for the second.

The mechanics differ from classic search. Claude's citations point to passages that support specific sentences in its answer, rather than to whole pages. A page can rank on page one, appear in Claude's search results and still lose the citation to a shorter, clearer competitor that states the answer in its first two lines. As of October 2026, more B2B buyers start research inside assistants, and the gap between ranking and being cited now decides which brands in a category those buyers see.

This guide explains how Claude acquires sources, how it decides which ones to cite, which page types it favors for each query intent, and how to test whether your changes work.

How Claude finds sources

Claude finds sources through three paths: knowledge absorbed during training, results returned by live web search, and content a user or connected tool supplies directly. Each path produces a different kind of visibility, and only two of them can produce a linked citation.

Why users see different citations

Two people asking the same question can get different sources. This is expected behavior, and several variables change what Claude cites:

• Search on or off. Without search, Claude answers from training priors and names brands with no links. With search, it retrieves and cites.
• Query reformulation. Claude rewrites the user's prompt into one or more search queries. "Best CSPM for a multi-cloud team" might become "CSPM tools comparison AWS Azure GCP", and that query decides which pages enter the candidate set.
• Developer configuration. Through the API web search tool, developers can restrict or block domains, so a Claude-powered product may never see your site.
• Model version and date. Newer models have later training cutoffs and different retrieval habits, and the web index itself changes week to week.

Crawler access

Anthropic documents separate user agents for separate jobs: ClaudeBot collects training data, Claude-SearchBot indexes content for search results, and Claude-User fetches pages when a person asks Claude to read them. Blocking ClaudeBot limits training exposure. Blocking Claude-SearchBot or Claude-User removes you from live retrieval, which is where citations come from. Many teams block all three by accident through a blanket AI-bot rule in robots.txt or a CDN bot-management setting. Check both.

How Claude decides what to cite

Claude cites the passage that most directly supports a sentence it is writing. Citations go to pages whose text answers a specific sub-question cleanly and carries evidence, from a source the answer can trust. Selection happens at the claim level. One answer can cite five domains, each for a different sentence.

The process runs roughly like this: the search step returns candidate pages, Claude reads them, drafts an answer, and attaches each citation to the text that backs a specific claim. A page loses the citation when its relevant passage is long, hedged or tangled with other topics, even if the page as a whole is strong. Other answer engines select citations at the claim level too, and our breakdown of how large language models decide what to cite covers it in more depth.

These signals raise the likelihood of being cited:

• Direct intent match: the passage answers the reformulated query, not a neighboring one.
• Explicit answer formatting: the claim comes first, followed by the qualification.
• Original data and first-party evidence: figures, benchmarks or product facts that no other candidate page holds.
• Quotable passages: 1–3 sentence statements that make sense when lifted out of the page.
• Crawlable HTML: content present in the server response, not injected by JavaScript.
• Clear entity associations: the brand, product and category named explicitly, not "our platform" or "this solution".
• Freshness where it matters: visible update dates on time-sensitive topics.
• Corroboration: the same claim appears on other trusted sources, which lowers the risk of stating it.

Hierarchy of evidence

When candidate sources conflict, answer engines lean toward the stronger form of evidence. Rank your own content against this order:

1. Primary studies and original research with stated methods
2. Official documentation, standards bodies and regulators
3. Original internal data published with a methodology
4. Named expert commentary with credentials
5. Secondary summaries and roundups
6. Forums and community posts, which are useful for edge cases and lived experience

Trust weighting matters more as the web fills with manipulated content. Security reporting shows threat actors abusing trusted AI platforms to host malicious content and poison search results, so retrieval systems have good reason to prefer sources with clear authorship, history and corroboration.

Mention, citation and presence are different wins

Track all three separately. A competitor named in every answer without citations has strong training priors. A competitor cited without being named owns the reference pages.

Which source types Claude cites for each query intent

Claude cites different page types depending on the intent of the query. Definitional prompts pull from reference explainers, comparisons pull from structured head-to-heads and reviews, and statistical prompts pull from original research. The table below maps intent to the sources that tend to win and to the structure worth copying.

Product pages rarely win informational citations. Claude uses them for factual specifics such as integrations, supported platforms and limits, and turns to neutral or expert sources for judgment calls. Community threads fill the gaps where no official page addresses a niche scenario.

When freshness matters

Freshness changes citation odds only on time-sensitive queries. Update these on a fixed schedule:

• Software comparisons and "best tools" pages, because features and pricing change
• Statistics pages, which lose credibility once newer data exists
• Policy and compliance explainers, whenever the rule changes
• Year-based queries such as "best X in 2026," where the reformulated search query often includes the year

Evergreen conceptual pages, such as a definition of CSPM, gain little from cosmetic date changes. Changing the date without changing the content risks signaling low-quality maintenance.

A step-by-step framework for earning Claude citations

Earning Claude citations is a repeatable loop. Find the prompts buyers ask, map who Claude cites today, build pages that beat those sources on extractability and evidence, then corroborate and monitor. Run it per prompt cluster, not per keyword.

1. Prompt research. Build the questions buyers ask assistants, phrased conversationally: "Which CSPM tool handles multi-cloud best for a 300-person security team?" Pull them from sales call notes, support tickets, Reddit threads and People Also Ask boxes.
2. Current citation mapping. Run each prompt with search enabled. Record every cited URL, every brand named and the order in which they appear.
3. Source-type analysis. Label each cited source: your site, a competitor, a review site, a community thread, a publication or documentation. The mix tells you whether to build a page, earn a mention or both.
4. Content-gap extraction. For each prompt, list the sub-claims Claude made and which source backed each one. Sub-claims backed by weak or outdated sources are your openings.
5. Article production. Write one page per cluster in the intent structure from the table above. Put a quotable answer under every heading, and add one piece of evidence the cited sources lack.
6. Publishing and indexing. Ship static HTML, submit the page through Search Console and IndexNow, and link to it from relevant pages on your site so crawlers find it quickly.
7. Corroborative mentions. Get the same facts stated on review profiles, analyst pages and community discussions. Claude is more willing to repeat a claim that appears on several independent sources.
8. Monitoring. Re-run the prompt set on a fixed cadence and compare results against the baseline, using the measurement protocol below.

Pre-publish checklist

• The title, H1 and opening sentence all target the same prompt cluster
• The first sentence under each H2 answers that heading on its own
• All core content appears in the raw HTML response (check with curl or by viewing the page source)
• The page shows visible publication and update dates
• A named author with relevant credentials appears on the page
• Every statistic names and links its primary source
• Tables include source notes
• Article, FAQPage or Product schema is added where it fits the page type
• Headings are descriptive, with stable anchor IDs that do not change on update
• The brand and product are named explicitly in key passages

Why pages rank but don't get cited

A page ranks but loses the citation when Claude cannot lift a clean, supported statement from it, even though search engines judge the page relevant overall. The two teardowns below use illustrative passages to show the difference.

Teardown: "What is CSPM?"

The cited page wins because its first sentence works as a complete answer when quoted alone. The other page makes Claude dig, and its statistic has no source, which makes it risky to repeat.

Teardown: "Wiz vs Orca for a mid-size team"

The losing page alternates paragraphs about each product and never states a verdict. The winning page opens with a two-line summary ("Choose X for agentless coverage across many accounts; choose Y if you need…"), follows with a feature table that has a source note, and ends with "choose X if" and "choose Y if" lists. Each row and each bullet can be quoted as an answer to a narrower follow-up question, so the winning page can earn several citations in a single answer.

Common failure modes

• Weak answer extraction: paragraphs that build up to a point instead of starting with it.
• Buried claims: the key fact sits halfway down a 300-word block.
• JavaScript-rendered content: many AI fetchers read the raw HTML and do not reliably execute client-side scripts.
• No unique evidence: the page restates what ten other pages already say, so Claude has no reason to pick it.
• Thin authorship and context: no author, no dates, no "about" signals.
• Vague headings: "Key considerations" matches no query.
• Unsupported statistics: numbers with no named source get skipped or cited elsewhere.
• Manipulation attempts: hidden text or instructions aimed at AI systems. Researchers have documented indirect prompt injection in the wild, including SEO poisoning to promote phishing domains, and pages that look like that pattern lose trust and do not earn citations.

How Tellr helps brands get cited by Claude

Tellr runs the citation loop as a managed program for enterprise marketing teams. Its senior team writes comparison pages, reviews and answer-shaped articles built to be quoted by Claude, ChatGPT, Perplexity, Gemini and Google AI Overviews, working from the brand's knowledge base and real user reviews. It places guideline-checked Reddit replies that supply the community corroboration answer engines draw on, and every week it tracks which domains Google and its AI Overviews cite for the category's queries. Tellr is built for teams spending $10k+ a month on marketing, so smaller teams will get more from a self-serve tracker.

• Articles published directly to the client's WordPress site
• Weekly citation share labelled by source: your site, competitors, Reddit, review sites and references
• Reddit replies placed behind an approval gate, with live status tracked
• "Tellr for Claude", an MCP server that lets teams query their program in plain language

How to measure Claude citations

Measure Claude citations with a fixed prompt set, a recorded baseline, and repeated runs on a set cadence. Single checks are unreliable when answers vary between sessions. Use this protocol:

1. Create the prompt set. For example, 60 prompts across your category, written the way buyers phrase questions.
2. Cluster by intent. Group prompts into the intent types from the table above so results map to page types.
3. Capture a baseline. Use fresh sessions with search enabled. Log the full answer, every cited URL and every brand named.
4. Analyze source overlap. Count which domains recur across prompts within a cluster. Domains that recur often are the reference sources for that cluster.
5. Compare answer shapes. Note whether Claude answers with a list, a table or prose, and match your page structure to that shape.
6. Track citation frequency. Citation rate = prompts citing your domain ÷ prompts run in the cluster. Run each prompt three times and average the results, because answers vary between runs.
7. Re-test on a fixed cadence. Re-run monthly. After publishing, allow a few weeks for crawling and indexing before you judge a page.

What is Claude-specific and what applies everywhere

Most of the editorial work transfers across engines. Compare these results with how ChatGPT chooses which sites to cite, and reuse the same prompt set when you work on showing up in ChatGPT answers, so results stay comparable.

Methodology note: The mechanisms described here come from Anthropic's public documentation of its crawlers and search citations, and from retrieval patterns shared by answer engines. Citation behavior shifts with prompt wording, configuration, model version and date, so treat each pattern as a hypothesis and confirm it on your own prompt set using the protocol above.

Claude SEO rewards the same discipline on every page: make the content reachable, state the answer first, back it with evidence other pages lack, and get it confirmed by sources Claude already trusts. Teams that measure citations prompt by prompt will see which of those changes is working for them.

FAQ