How to Rank on Claude in 2026: What Anthropic Actually Documents

By Minel Gunesoglu, founder of Is My Brand in AI · Published September 6, 2026. Every claim on this page links to the primary source it came from, and every source was read on September 6, 2026.

TL;DR: Anthropic publishes no ranking factors for Claude, so every ranking-factor table in this niche is invented. What it does document is checkable: three crawlers, citations always on in web search, a 150-character quote budget, and a web fetch tool that does not render JavaScript. Start there. The widely repeated 86.7% Brave overlap figure comes from 15 results on three queries.

This page exists because the published advice on ranking in Claude is mostly unsourced. Search for it and you get ranking-factor tables with percentage weights next to each factor, a crawler that Anthropic does not document, and one overlap statistic repeated across a dozen sites with no link back to where it came from. I opened those pages and I opened Anthropic's own documentation, and the two do not describe the same system.

So this is not a tactics listicle. It is the documented record: what Anthropic states in its own docs, what it has never stated, and which popular claims fall apart when you check them. Where the honest answer is "Anthropic has not published this," I say that instead of filling the gap. The sibling guides for Perplexity, Gemini and ChatGPT have far more published evidence behind them. Claude has less, and pretending otherwise is how the bad tables got written.

Does Claude render JavaScript?

No, and this is the single most actionable fact available about Claude and your website, so it goes first.

Anthropic's documentation for the web fetch tool states it plainly: "The web fetch tool currently does not support websites dynamically rendered with JavaScript." If your main content only appears after a client-side render, the tool retrieves a shell. There is no queue and no second pass, the way Googlebot handles rendering. Anthropic's suggested workaround is a separate browser tool that your own application drives, which is not something a publisher can trigger from the outside.

Two more limits from the same page shape what survives a fetch. Content types are restricted: an unsupported_content_type error fires for anything that is not "text, HTML, and PDF". And long pages get cut, because the max_content_tokens parameter means "if the fetched content exceeds this limit, the tool truncates it," with an average 10 kB web page costing roughly 2,500 tokens. A very long page can be fetched and still have its later sections never reach the model.

The practical version: serve your answer in the HTML that comes back from the server, and put it early. That is not a Claude-specific growth hack, it is the same server-rendered-HTML discipline that governs whether any crawler can read you, and it is testable in one command with curl.

Which crawlers does Anthropic run, and what happens if you block one?

Anthropic documents exactly three robots on its crawler page, last dated April 7, 2026, and it states a consequence for blocking each one. The consequences are the useful part, because they tell you which bot governs live citation and which governs training.

Source: Anthropic support documentation, page dated April 7, 2026, read September 6, 2026. Quotes are verbatim.

The two bots most robots.txt files miss

ClaudeBot is the one everybody knows, and it is the training crawler. Blocking it has no stated effect on whether Claude cites you today. The two that Anthropic ties directly to visibility are Claude-User and Claude-SearchBot, and those are the ones that go missing from hand-written allow lists. If you copied an allow list from anywhere in the last two years, check it for all three strings; the full multi-engine reference is in our guide to allowing AI crawlers in robots.txt.

Two strings you will find in older robots.txt files and older guides, Claude-Web and anthropic-ai, do not appear anywhere on Anthropic's current crawler page. I searched the page text for both on September 6, 2026 and found zero occurrences. Any guide that assigns one of them a live job today is describing a bot Anthropic does not document.

Robots.txt, crawl-delay, and why IP blocking backfires

Anthropic's stated rule is that "Anthropic's Bots respect 'do not crawl' signals by honoring industry standard directives in robots.txt." The same page says it supports "the non-standard Crawl-delay extension to robots.txt" and that it will not attempt to bypass CAPTCHAs.

It also warns against the blunt instrument. Blocking by IP address "may not work correctly or persistently guarantee an opt-out, as doing so impedes our ability to read your robots.txt file." Block the address range and you also block the file that carries your instructions.

Robots.txt matters on the other side of the pipeline too. The web fetch tool's url_not_allowed error is documented as a URL "blocked by domain filtering rules (including your organization's settings) or by Anthropic-side restrictions, such as private addresses and robots.txt". The file is enforced at fetch time, not only at crawl time.

Verifying that a crawler in your logs is really Anthropic

Anthropic publishes its crawler IP ranges as a live file at claude.com/crawling/bots.json, and says that "if a crawler has a source IP address on this list, it indicates that the crawler is coming from Anthropic." I fetched it on September 6, 2026: HTTP 200, a creationTime of 2026-08-18T23:56:36Z, and 26 IPv4 prefixes. If you are auditing log entries claiming to be Claude, that file is the check, and it takes one request.

How does Claude decide to search, and what does it get back?

Anthropic's web search tool documentation describes the trigger and the payload in more detail than any competing guide reflects. Four details change how you should write.

Claude does not search every time

"Claude determines when to search based on the prompt." It searches when the request "depends on information that is current, changing, or outside its training data," which the doc lists as recent events, current prices and statistics, information about specific organizations or products that might have changed, and explicit requests to look something up. It answers directly, with no retrieval at all, on "established facts, math, science fundamentals, or coding concepts," on creative work, and on content already in the conversation.

That splits your target questions in two. Questions about your category as it stands today trigger retrieval and can cite you. Questions about settled concepts may never touch the web, in which case what the model already absorbed during training is what gets said. Volume varies with the question: "simple factual queries typically use 1 to 3 searches; comparative or multientity research can use 10 or more."

Citations are not optional

"Citations are always enabled for web search." Unlike the Perplexity and Gemini surfaces, where citation display is a product decision that has shifted over time, this one is stated as a property of the tool. The consumer help page says the same thing to users: "Every response includes citations, so you can easily verify sources yourself."

The inversion catches people out: for the web fetch tool, "citations are optional for web fetch and disabled by default." Search and fetch are two different tools with opposite defaults, and guides that talk about "Claude's citations" as one behavior are flattening that.

The quote budget is 150 characters

Each citation returned by web search carries a url, a title, an encrypted_index, and a cited_text field documented as "up to 150 characters of the cited content."

That is a hard, published constraint and I have not seen another guide mention it. It means the unit that gets lifted out of your page is roughly one or two sentences, not a paragraph and not a section. A claim that only makes sense across four sentences cannot fit in the field. State the finding, the number and the qualifier inside a single sentence that survives being cut out of the page.

Anthropic's own agent prompts point the same way. The citations_agent.md prompt in the claude-cookbooks repository instructs the model to "cite meaningful semantic units", meaning "citations should span complete thoughts, findings, or claims that make sense as standalone assertions." That is a description of what a citable sentence looks like, written by the vendor, for its own system.

Freshness is a field, not a theory

Every search result carries a page_age, documented as "when the site was last updated." Anthropic does not say what the model does with it. It says the value is there. That is the entire published basis for "freshness matters on Claude": a field exists on the payload. Anyone who tells you what weight it carries is inventing the weight.

One more mechanical change worth knowing. From tool version web_search_20260209, search runs with dynamic filtering: rather than loading every result into context, "Claude instead writes and runs code that filters the results first, so only relevant content reaches the context window." A result being retrieved no longer guarantees the model reads it.

Does Claude run on Brave Search?

Partly, and the popular version of this claim is both better sourced and weaker than it looks.

Anthropic names its search vendors in its Trust Center subprocessor list. Read live on September 6, 2026, two entries carry the category Web Search: Brave Search, scoped to "All products", and TurboPuffer, scoped to "All Products except Claude for Government". Both rows link to Anthropic's web search help article. No Bing or Google entry appears under Web Search. The page renders client-side, so a plain fetch returns an empty shell; you have to load it in a browser to see the table.

Anthropic names Brave in prose in exactly one place I could find: the help article for web search in Claude for Government, dated April 9, 2026. It says the connector "calls the Brave Search API," and that once you approve a search, "the query goes to Brave, results come back, and Claude continues." Read the scope carefully before you generalize it. The same page states that "in commercial Claude, web search is a built-in capability. In Claude for Government, native web search is disabled," and the connector replaces it. That is a documented Brave dependency for a surface where the normal web search is switched off.

So "Claude is powered by Brave" is directionally right and imprecise. Brave Search is a named web search subprocessor across all products. A second web search subprocessor sits beside it. Anthropic has never published how the two combine, whether either result set is re-ranked, or what the model does with what comes back.

There is a third piece of vendor language that no Claude SEO guide I read has quoted, and it is the most practically useful sentence Anthropic has published for site owners. In its page on blocking and removing content from Claude, it explains what noindex does: "the noindex robots meta tag is a rule that tells our partners not to index your content so that they don't send it to us in response to your web search query." The same page adds that "our search partners only index images and videos that their bots are allowed to crawl."

Anthropic is confirming, in its own words, that Claude's web search is served by third-party search partners whose index honors ordinary indexability rules. A noindex tag keeps you out of Claude. That is the one lever Anthropic explicitly documents, and it only works in the negative direction. There is no published counterpart telling you how to get in.

What has Anthropic published about ranking factors?

Nothing. Not a factor, not a weight, not an ordering rule.

I read the web search tool docs, the web fetch tool docs, the crawler page, the consumer help pages and the launch post. They describe when Claude searches, what the results contain, and how to keep content out. Not one of them describes how candidate results are ordered or which of them gets cited. Every "Claude ranking factors" table with percentages beside each factor was written by someone who did not get those percentages from Anthropic, because Anthropic has not published any.

What Anthropic has published instead is more interesting, and it points the opposite way from most of the advice in this niche. In its engineering write-up on building a multi-agent research system, dated June 13, 2025, the team describes a problem it had to fix:

"Human testers noticed that our early agents consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources like academic PDFs or personal blogs."

The next sentence: "Adding source quality heuristics to our prompts helped resolve this issue." Anthropic treated a preference for SEO-optimized content as a defect and corrected it deliberately.

Some of those heuristics are public. The research subagent prompt in the claude-cookbooks repository tells the model to watch for problematic sources, listing "news aggregators rather than original sources of the information, false authority, pairing of passive voice with nameless sources, general qualifiers without specifics, unconfirmed reports, marketing language for a product, spin language, speculation, or misleading and cherry-picked data."

Read that list as a specification, because it is the closest thing to a published quality signal Anthropic has shipped. It is not a ranking factor, and it belongs to agentic research prompts rather than to the general web search path, so do not overstate what it governs. But it names, precisely, the things a source can do that make a model distrust it: repeat someone else's reporting, claim authority it lacks, attribute to nobody, qualify without specifics, sell, speculate, and pick cherries. Fixing those is available to anyone today and requires no theory about weights.

Which widely repeated Claude SEO claims do not survive checking?

Each row below is a claim I found on pages ranking for Claude visibility queries. I traced each one back and read the source myself on September 6, 2026.

The 86.7% figure deserves the extra sentence, because it is the load-bearing claim of the entire niche. It is not fabricated. Profound ran a real comparison and published it. The problem is what happened next: a 15-result observation from the day after launch got repeated for eighteen months as a settled property of the system, usually without the sample size, often without a link, and always without noting that Anthropic now lists a second web search subprocessor alongside Brave. That is how a small honest test becomes folklore.

The same publisher has since run a larger one, and it is worth knowing about even though it is not published in a checkable form. At Zero Click New York on June 11, 2026, Josh Blyskal and Jasman Singh presented a figure of 79.2% of Claude citations coming from Brave's top 10 results. That points the same way as the 2025 number and rests on more than fifteen results. It also states no sample size, names no query set and offers no data to download, so it cannot be recounted by anyone outside Profound. Read it as a stronger signal than the launch-week test and still as one vendor's unverifiable measurement. We have written before about how GEO statistics decay in exactly this way.

Has anyone measured what Claude actually cites?

Barely, and I want to be exact about our own position first.

We have not measured Claude. Our Cross-Engine Citation Study put the same buyer questions to four engines, and those four were ChatGPT, Perplexity, Gemini and Google AI Overview. Claude was not one of them, in either the June or the August 2026 round. The full dataset is published and you can confirm that yourself in the engine column. Nothing on this page comes from a Claude measurement of ours, because there is not one.

The largest public count I could find is Otterly's Claude citation study, published June 30, 2026.

Read September 6, 2026. Otterly discloses its scope and date range but publishes neither the prompt list nor the raw data, so none of these figures can be independently recounted. Treat them as one vendor's unverified measurement of one sector in one month.

If those numbers hold, two things follow that are unusual against the other engines. Citation is spread very thin, with the top ten domains taking under a tenth of the total, so there is no small club of incumbents to displace. And the Reddit playbook that works on Perplexity has no visible purchase here. I am reporting that, not endorsing it, because a study whose data is not published cannot be checked.

Everything else in this niche presented as measurement is a vendor's private dataset described in prose. The open questions are still open: nobody has published a same-query Claude-versus-Brave overlap test with the raw data attached, so none of the overlap figures in circulation can be recounted independently, nobody has shown whether headings, tables or JSON-LD survive the fetch tool's extraction step, and nobody has tested whether page_age affects which result gets cited.

What should you actually do?

Every item here traces to something Anthropic documents. Nothing on this list depends on a ranking factor, because there are none to depend on.

• Your main content is in the HTML the server returns, not injected by JavaScript, because web fetch does not render it.
• Claude-User and Claude-SearchBot are allowed in robots.txt, not just ClaudeBot. Those are the two Anthropic ties to visibility.
• You are not blocking Anthropic by IP, which Anthropic warns also blocks its ability to read your robots.txt.
• No noindex tag sits on a page you want cited. Anthropic states this keeps you out of Claude via its search partners.
• The answer to each target question fits inside one or two sentences that stand alone, because cited_text carries at most 150 characters.
• Key content sits early in the page, since long fetched pages are truncated at a token limit.
• Your page names its author, dates its claims, links its sources and avoids marketing language, speculation and nameless attribution, which is what Anthropic's own research prompts screen against.
• You have checked claude.com/crawling/bots.json against any crawler in your logs claiming to be Claude.

Notice what is absent. No schema recommendation, because nobody has shown JSON-LD survives fetch extraction. No llms.txt, because Anthropic has never confirmed reading it. No refresh cadence tied to a number, because the only published freshness signal is a field on a payload with no documented weight.

Frequently asked questions

Does Anthropic publish Claude ranking factors?

No. Its web search and web fetch documentation describes when Claude searches and what a result contains, never how results are ordered or selected for citation. Every ranking-factor table with weights beside each factor was invented by its author. What Anthropic does publish is the opposite of a ranking guide: a note that its early research agents wrongly preferred SEO-optimized content farms over authoritative sources, and that it corrected this.

Which Claude crawler should I allow if I want to be cited?

Claude-User and Claude-SearchBot. Anthropic states that blocking the first "prevents our system from retrieving your content in response to a user query" and blocking the second "prevents our system from indexing your content for search optimization." ClaudeBot is the training crawler and Anthropic ties no citation consequence to it. The syntax and the full multi-engine reference are in how to allow AI crawlers in robots.txt.

Does Claude use Brave Search?

Anthropic lists Brave Search as a Web Search subprocessor for all products in its Trust Center, and names the Brave Search API in its Claude for Government web search article. It also lists a second Web Search subprocessor, TurboPuffer, and has never published how the two combine or whether results are re-ranked. So "Claude runs on Brave" is partly documented and not the whole system.

Where does the 86.7% Claude and Brave overlap figure come from?

A Profound post dated March 21, 2025, the day after Claude web search launched. The stated basis is 13 of 15 results across three commercial queries. It is a real test with a very small sample from launch week, and it is quoted across this niche in 2026 as a settled property of the system, usually without the sample size.

Has Is My Brand in AI measured Claude?

No. Our Cross-Engine Citation Study covered ChatGPT, Perplexity, Gemini and Google AI Overview across two rounds in June and August 2026. Claude was not an engine in either round, which you can verify in the published dataset. Every figure on this page comes from Anthropic's documentation or from a named third party, and each is labeled as such.

Should I publish an llms.txt file for Claude?

There is no published reason to expect it to help. Anthropic has never confirmed reading llms.txt, and the two publisher-facing controls it does document are robots.txt and the noindex meta tag. Anyone telling you Claude reads the file is not quoting Anthropic.

---

Ranking on Claude comes down to a shorter list than the guides suggest, because the vendor has published less than the guides pretend. Serve your content as HTML, let the two visibility crawlers in, keep noindex off the pages you want cited, and write claims that survive being lifted 150 characters at a time. Everything past that is somebody's estimate, and the honest move is to label it as one.

This guide is part of our series on AI search visibility, anchored by how to rank on ChatGPT. For the answer-engine cousins, see how to rank on Perplexity and how to rank on Gemini. Written by Minel Gunesoglu, founder of Is My Brand in AI. More about us.