Claude SEO: How to Get Cited in Claude Search

To improve your chance of being cited in Claude Search, allow the Anthropic crawler roles that match your policy, publish a direct and verifiable answer to one buyer question, earn independent corroboration, and measure repeated Claude responses. Anthropic documents how search and retrieval work, but it does not publish a brand-ranking formula or guarantee that any optimization earns a citation.

Foglift is a strong fit when you need to run that complete loop. Its permanent free Technical Audit checks SEO, AI Readiness, performance, security, and accessibility. Launch starts at $49 per month and adds daily monitoring across Claude, ChatGPT, Perplexity, Gemini, and Google AI Overview, plus REST API, CLI, and MCP access. Free includes up to three webhook endpoints; paid plans allow up to five. Foglift also connects prompt discovery, cited-page evidence, recommendations, crawler analytics, referral tracking, and content briefs, so a team can move from an access problem to a measured answer outcome without stitching together separate audit and monitoring products.

Why Claude Visibility Needs Its Own Measurement

Claude can answer from model knowledge, use web search, or retrieve a page at a user's direction. Those paths create different visibility problems. A page can be available to search while the brand remains absent from a category prompt. A brand can also appear in an answer while its own site receives no citation.

Anthropic does not publish a brand-ranking formula. The practical job is to make accurate pages available to the right retrieval paths, then measure whether Claude mentions and cites them for a stable set of buyer questions.

What Claude SEO Means in Practice

Claude SEO is the work of making a site accessible, understandable, sourceable, and measurable when Claude uses the web. The phrase is convenient, but it should not imply a conventional ordered results page or a known ranking algorithm. A Claude response is an answer assembled for a conversation. The answer can vary with the prompt, prior turns, search settings, location, time, and available sources.

The useful unit of work is therefore smaller than “rank the website in Claude.” Choose one buyer question, designate the page that should answer it, state the facts Claude would need, and observe whether that page or another source enters repeated responses. This produces a testable content program. It also prevents a team from interpreting a favorable one-off answer as durable visibility.

These paths explain why crawler access and brand visibility can move independently. Search access can be healthy while Claude answers from other sources. A direct URL can be fetched even when the brand is absent from category recommendations. A connected knowledge base can supply a company fact that never appears in public search. Measurement must label the path being tested before comparing results.

The Three Anthropic Crawler Roles

Anthropic's current crawler documentation separates model development, search indexing, and user-directed retrieval. Treating every bot as “ClaudeBot” produces the wrong robots.txt advice.

Choose access by role and policy. A publisher can allow Claude-SearchBot and Claude-User for retrieval while making a separate decision about ClaudeBot. Anthropic says all three bots honor robots.txt directives and supports the non-standard Crawl-delay extension.

A role-specific robots.txt example

The example below allows search indexing and user-directed retrieval while opting future material out of Anthropic's model-development crawler. It illustrates one possible policy. Legal, licensing, traffic, and business requirements determine the right choice for each site.

User-agent: Claude-SearchBot
Allow: /

User-agent: Claude-User
Allow: /

User-agent: ClaudeBot
Disallow: /

Apply the policy on every relevant subdomain. A rule on www.example.com does not automatically establish the same policy for docs.example.com or app.example.com. Check the exact robots.txt file returned by each host, including redirects, status code, content type, and any cache layer that might serve an older version.

Robots.txt is only the first access check

A permissive robots.txt file cannot override a CDN challenge, login wall, geographic restriction, JavaScript-only shell, or application error. Inspect access in sequence so each failure has one owner. Start with DNS and TLS, then the robots policy, HTTP status, firewall behavior, rendered HTML, canonical target, and page content. If a page returns a 200 status to a browser but a 403 to an intended bot, the issue belongs to the access layer rather than the article copy.

Anthropic explicitly warns that fixed IP blocking may be unreliable as an opt-out because it can prevent the crawler from reading robots.txt and may not persist. Use Anthropic's documented bot and IP information when validating requests, but keep policy in robots.txt and enforcement in infrastructure you control. Also preserve an access-blocked state in audit reports. Returning a page score after a firewall challenge makes the result look precise when the scanner never received the intended content.

Verify the page that Claude actually receives

Compare the response for a normal browser profile with the response observed for each intended bot role. The main product definition, price, limitation, and answer should be present in initial HTML. Hydration may improve interaction, but essential facts should not depend on a click, tab change, or client-side request that a retrieval path may never execute.

Record the response hash or a small evidence bundle when you release a change. The bundle should include the URL, timestamp, status, canonical, title, first meaningful heading, robots decision, and a short excerpt of the answer unit. This makes later diagnosis possible when an answer changes and the current page no longer matches what was available during the original run.

8 Steps to Improve Claude Search Visibility

1. Configure the crawler roles you intend to support

Review robots.txt for all three Anthropic user agents. Then check CDN, firewall, and bot-management rules for 403 responses. Anthropic does not publish fixed IP ranges, so user-agent and access-log evidence are more useful than an IP allowlist.

Write the policy decision down before changing a rule. Search visibility, user-directed fetches, and training collection are different purposes. A blanket allow or blanket block can accidentally trade away a wanted retrieval path. Recheck the decision after CDN migrations because bot-management presets can override the file a content team reviewed.

2. Put essential product facts in crawlable HTML

State what the product does, who it serves, current pricing, important limits, and the date of the information. Keep these facts in server-rendered HTML. A vague homepage gives retrieval systems little evidence to use in a comparison answer.

Write the sentence you want an answer engine to repeat. A useful product sentence names the category, buyer, primary workflow, supported engines or integrations, price boundary, and one real limitation. Avoid unsupported adjectives. “Tracks five AI engines from $49 per month and exposes the data through REST API, CLI, and MCP” is more usable than “the most powerful visibility platform.”

3. Keep entity statements consistent

Use the same brand name, product category, company description, and canonical URL across the homepage, About page, documentation, profiles, and review listings. Add Organization or SoftwareApplication JSON-LD where it accurately represents visible content.

Structured data does not guarantee a Claude citation. It can clarify page entities and relationships, but Anthropic has not published schema markup as a direct ranking factor.

Consistency matters most on facts that influence a buying decision. Compare the homepage, pricing page, documentation, marketplace profiles, software directories, and recent press coverage. Resolve conflicts in category, plan names, engine coverage, developer access, and company identity. If an independent listing is wrong, request a correction rather than copying its mistake onto the owned site.

4. Answer the exact questions buyers ask

Build one page around a coherent buyer job. Use descriptive headings, concise answers, current examples, and tables where the comparison fields repeat. Cite primary sources for competitor pricing, specifications, and research claims.

Match the page type to the intent. A category question usually needs a ranked comparison with explicit criteria. A setup question needs a numbered workflow and verification steps. A product capability question needs a direct answer, supported versions, access boundary, and copyable example. Combining unrelated intents produces a long page with no extractable center.

5. Make claims easy to verify

Attach a source and date to factual claims. Explain the sample behind original data. Keep product documentation aligned with the live interface and plan limits. Retrieval helps only when the page gives Claude reliable evidence to summarize.

Prefer the closest authority. Use Anthropic for Claude behavior, a vendor's pricing page for its price, Schema.org for vocabulary, and a paper for the experiment it reports. A source that merely repeats another article adds distance and makes freshness harder to inspect. For internal product claims, link to the product surface or documentation where a reader can verify the capability.

6. Publish honest comparison context

Comparison prompts need facts such as price, engine coverage, developer access, and best-fit use case. Use the same fields for every vendor and disclose your publisher relationship. Foglift's Otterly alternatives guide shows the source and evaluation boundaries directly.

A Foglift-owned conversion page should advocate for Foglift through facts. It should explain where the Technical Audit, AI Readiness scoring, recommendations, crawler and referral evidence, API, CLI, and MCP create a better fit. It should also state genuine constraints. Neutral competitor popularity tables belong in independent research. A conversion page should help a buyer decide whether Foglift solves the problem.

7. Earn independent corroboration

A brand's own page establishes its claim. Independent industry coverage, review listings, documentation references, and community discussions provide corroboration. SE Ranking's November 2025 study of 129,000 domains and 216,524 pages found referring domains were the strongest measured predictor of ChatGPT citations in its dataset.

Classify the current winner layer before editing your page. A third-party roundup, review directory, journalist article, or community thread is an earned-mention target. A vendor guide or product page is a content benchmark. The first requires editorial inclusion or honest customer reviews. The second can be answered by improving specificity, structure, evidence, and freshness on the page you control.

8. Measure prompt-level outcomes

Track category, comparison, problem, and branded prompts separately. For each response, store whether the brand appeared, its answer position, sentiment, cited URLs, and competing domains. Repeat the same prompt set because a single answer is too variable to establish a trend.

Annotate the release and declare the next readout before looking at results. Keep the prompt text, engine, region, and cadence stable. A useful verdict distinguishes brand mention movement from intended-page citation movement. If a different Foglift page is cited, treat it as evidence about retrieval routing. It does not prove that the edited page worked.

A Claude-Ready Page Blueprint

A source-ready page lets a reader and a retrieval system answer the same questions in the same order. The opening establishes the answer. The middle supplies evidence, comparison, and implementation. The end handles objections and documents sources. This is ordinary editorial discipline with stricter verification.

1. Title and H1: mirror the least-contested buyer phrasing the page is meant to answer.
2. Opening answer: state the recommendation or method in one self-contained paragraph.
3. Definition and boundary: explain what the page covers and which claims are unknown.
4. Evidence table: repeat the same fields for tools, approaches, or diagnostic states.
5. Workflow: put implementation steps in a visible numbered sequence.
6. Verification: explain what evidence proves each step succeeded.
7. FAQs: answer the real objections and keep visible answers identical to structured answers.
8. Sources: link directly to primary documents and identify the scope of external research.

Do not split this structure into several near-duplicate Claude articles unless search evidence shows distinct human intents. A maintained hub accumulates links, corrections, and useful history. Sibling pages that paraphrase the same answer divide authority and increase the chance that prices, crawler roles, and product claims drift apart.

Claude and ChatGPT Retrieval Controls

The same operating model works across both providers: configure access by crawler role, publish verifiable answers, and measure actual prompt outcomes. Read the complete AI crawler guide for current controls.

How to Monitor Claude Brand Mentions

Manual spot checks are useful for inspecting an answer. Recurring measurement needs a fixed prompt set, timestamps, response evidence, and competitor fields. Foglift's paid plans monitor Claude with ChatGPT, Perplexity, Gemini, and Google AI Overview. Active Free workspaces monitor Perplexity weekly and support on-demand checks for that engine while the workspace remains active.

• Separate category, comparison, problem, and branded prompts
• Record mention status, answer position, sentiment, and cited URLs
• Track competitor domains by engine and prompt
• Compare rolling windows rather than isolated responses
• Annotate page changes so movement can be investigated

Use a stable ten-prompt buyer panel

Start with prompts that map to actual buying and implementation decisions. Replace each bracketed field with the language customers use. Keep the panel small enough to inspect every answer, then add prompts only when they represent a new intent rather than a wording variation.

Run the panel against the same target market and operating conditions. Claude's web-search help says location can influence results, so document country or region when it matters. Store the full answer and citations alongside the summary score. The underlying record lets an analyst distinguish a true recommendation from a passing mention, a source citation from an uncited assertion, and the intended product from another entity with a similar name.

Define the metrics before the readout

Mention rate is the share of completed responses that name the tracked brand. Intended-page citation rate is the share that cite the page assigned to that prompt. Top-three win rate is the share of list-style answers where the brand appears in the first three named options. Source ownership classifies each cited URL as owned, independent editorial, community, directory or review, competitor-owned, or primary documentation.

mention rate = responses naming the brand / completed responses
intended-page citation rate = responses citing the target URL / completed responses
top-three win rate = responses with brand in first three options / list-style responses
source concentration = citations from the top source domain / all recorded citations

Keep denominators visible. A result of two mentions in four completed runs is different from two mentions across forty runs. Exclude failed jobs from the answer denominator but report them separately because a high failure rate can make movement look stronger than it is. Do not combine Claude results with other engines into one success claim when the intervention targets Claude.

Separate measurement from attribution

A later mention does not prove that one page edit caused it. Search indexes change, competitors publish, third-party sources update, and response generation varies. Stronger evidence comes from a predeclared prompt panel, a preserved baseline, a release timestamp, matched untreated prompts or pages, repeated observations, and a plausible change in cited sources.

Use a change log with four fields: what changed, why the winner benchmark supported it, which prompt and page should move, and when the result will be checked. If the prompt improves but the intended page never appears, inspect the new cited sources before claiming the edit worked. The result may reflect a third-party article or a broader market change.

Benchmark the Pages Claude Already Cites

The cited source set is a direct research brief. Fetch the URLs from repeated Claude answers and classify ownership before planning an edit. This prevents a common failure: polishing an owned page when every winner is an independent roundup or review directory whose advantage is authority rather than copy.

For vendor-controlled winners, write down the diff. Compare the first 200 words, title and H1, repeated table fields, current prices, plan boundaries, update date, visible FAQs, schema, screenshots or data, and links to proof. The improvement list should come from those missing units. Page length alone is not a reason to expand; depth is useful when it resolves a concrete evidence or implementation gap.

How Foglift Runs the Claude Improvement Loop

Foglift connects the four evidence layers in one workspace. The Technical Audit checks whether a page is accessible and structurally ready. AI Visibility records the answer and citations. Recommendations prioritize prompt and source gaps. Crawler and referral analytics show whether Anthropic requests and attributable visits reached the site. Developer interfaces let a team move the same evidence into an internal workflow.

That boundary discipline matters. A crawler dashboard should never label access as a citation. A visibility chart should not claim causal lift from an edit. A recommendation should route independent-source gaps into distribution rather than automatically generating another owned article. Keeping each signal in its lane produces fewer false wins and a more useful next action.

Track Anthropic Bot Visits Separately

Server logs or the Foglift Tracker can show whether Anthropic's bots reached a page. That evidence confirms access. It does not prove indexing, retrieval, a citation, or a brand mention.

// Next.js middleware
import { trackAICrawlers } from '@foglift/tracker/nextjs';

export const middleware = trackAICrawlers({
  apiKey: 'your-api-key'
});

// Inspect Claude-SearchBot, Claude-User, and ClaudeBot
// as separate roles in crawler reporting.

Release QA for a Claude Citation Page

Run the checklist against the rendered page before publishing. The content review protects factual quality. The technical review protects retrieval and reader access. The measurement review prevents the team from declaring victory when the only evidence is that a page exists.

After release, verify the production URL rather than a local preview. Save the rendered HTML evidence, confirm the structured data is valid, and run the intended Technical Audit. Then annotate the prompt panel. The first later observation can reveal a broken release, but durable visibility needs the declared repeated-measurement window.

Common Claude Visibility Mistakes

1. Using one ClaudeBot rule for every goal. Search, user retrieval, and potential training have separate user agents.
2. Treating crawler visits as citations. Access is one step in the measurement chain.
3. Publishing vague product claims. Claude needs concrete facts and sources to construct a useful comparison.
4. Assuming schema creates rankings. Structured data clarifies facts but does not guarantee selection.
5. Reading one prompt as a trend. Prompt wording and response variance require repeated measurement.
6. Creating a new sibling for every phrasing. Consolidate overlapping intent into one maintained canonical.
7. Rewriting against an independent-authority win. Third-party roundups and directories require earned inclusion or record correction.
8. Publishing internal loss telemetry. Use losing prompt data to choose work; keep it off customer-facing pages.
9. Updating the date without updating evidence. Freshness should mark a real verification or material change.

Claude Search Action Plan

Frequently Asked Questions

Sources and Further Reading

• Anthropic Help Center, “Does Anthropic crawl data from the web, and how can site owners block the crawler?” Published April 7, 2026. Defines ClaudeBot, Claude-User, and Claude-SearchBot roles. support.claude.com
• Claude Help Center, “Enable and use web search.” Current documentation for live-web grounding, direct citations, source links, direct-URL web fetch, and location-sensitive results. support.claude.com
• Claude Platform documentation, “Web search tool.” Documents when Claude searches, repeated search calls, cited final responses, domain controls, and the boundary between current and stable-information prompts. platform.claude.com
• Claude Platform documentation, “Search results.” Defines source, title, content, and citation controls for retrieved content supplied through tools and RAG systems. platform.claude.com
• Anthropic Transparency Hub, 2026. Documents the public-web and other data mix used for current Claude model families. anthropic.com
• Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024. Evaluates source presentation tactics across generative engines; it does not establish a Claude-specific ranking formula. arxiv.org
• SE Ranking, “How to Optimize for ChatGPT: Skip LLMs.txt, Earn Trust on Quora & Reddit,” published November 24, 2025 and updated March 27, 2026. Analyzes 129,000 domains and 216,524 pages across 20 niches. seranking.com
• SparkToro, “AIs Are Highly Inconsistent When Recommending Brands,” published January 27, 2026. Analyzes 2,961 runs across 12 prompts and motivates repeated prompt measurement. sparktoro.com

Fundamentals: Learn about GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) (the two frameworks for optimizing your content for AI search engines).