How to Show Up in Claude Search (2026)
Client order for Claude: Claude-SearchBot and Claude-User in robots.txt per subdomain, training as a separate legal call, then Promptwatch on the answer. No IP-block opt-out.
When a client wants to "show up in Claude," we open Anthropic's crawler help before we open a content brief. The platform we measure on is Promptwatch. Cited means a URL they control in the sources. Named without a link is a mention. We will not mash those in a QBR. The two outcomes look similar on a slide and they mean different things to a buyer, so we keep them apart from the first ticket. The order matters: access first, page second, log third, and content only after the log says the bot can read the page.
Security and legal, split
Three user-agents. We write all three into the same ticket so nobody "blocks AI" as one string. A single string match is how a team blocks search while thinking it blocked training, and that mistake is hard to undo in a WAF log.
Claude-SearchBot is search indexing. Anthropic: disabling it prevents indexing for search optimization, which may reduce visibility and accuracy in user search results. If the SOW says Claude Search, this is Allow. This is the lever for the title, and it has to be on before any content work matters.
Claude-User is the live fetch when a person asks. Disabling it prevents retrieval for that query, which may reduce visibility for user-directed web search. Anthropic says its bots honor robots.txt, including this one. We do not copy OpenAI's "User may ignore robots.txt" line onto Claude. The two vendors made different choices about their user-agent, and we respect the difference instead of porting one rule to the other.
ClaudeBot is training. Restricting it signals that future materials should be excluded from training datasets. Legal often wants that Disallow. Fine. Keep it off the SearchBot line. Training and search are separate decisions for separate people, and the robots file has to show that.
Every subdomain they care about. Anthropic says to repeat the file. Crawl-delay is supported (their example is Crawl-delay: 1). We will not sell delay as a ranking tactic. They will not bypass CAPTCHAs. A challenge wall that Googlebot passes and Claude-SearchBot does not is an infra finding. It goes to the platform team, not the writer, because the writer cannot fix a WAF rule.
We do not IP-block as an opt-out. They say it may not work, because the bot then cannot read robots.txt. Malfunction reports: claudebot@anthropic.com from a mailbox on that domain. A mailbox on the domain is the proof of ownership the support address expects, and a free webmail address will not get the same response.
The page and the log
One public URL per money question, claim in HTML. We already use Google's optimization guide as a quality bar. We do not tell clients Anthropic signed it. No Claude submit console. No llms.txt prerequisite on that help page. The guide is a quality bar for the page, not a Claude policy, and we label it that way.
Paid Promptwatch includes Claude. Explore is ChatGPT-only, so the Claude clause starts at Essential ($95/mo). Essential has no listed crawler-log allowance. Professional ($245/mo) is the first brand plan with crawler logs, including ClaudeBot in Agent Analytics, and it also expands the prompt list. Citation analytics show the cited URL. Paid checks refresh daily. They are not a pager. A daily check is a record, not an alert, and that distinction matters when a client asks why they did not get pinged.
Otterly at $29 is a side thermometer. Profound Starter is ChatGPT-only. Wrong SKU. Content Agents to Webflow or Framer only after SearchBot is allowed and a human accepts. WordPress: write where editors already work. We do not push Profound to the top, and we do not drop it to the bottom with Otterly and Peec.
If SearchBot is Disallow, we do not write. If SearchBot can fetch and Reddit is the source, we discuss offsite. Frozen prompts from sales, two weeks, cited-host column. The two weeks are what separate a real move from a cache refresh, and a cited-host column is what lets the AM point at the URL that won.
We repeat robots.txt on every subdomain in the SOW. We do not IP-block as an opt-out. Anthropic said that can stop the bot from reading the file. Crawl-delay (Crawl-delay: 1 in their ClaudeBot example) is a rate hint. We will not sell it as ranking. CAPTCHA walls that Googlebot passes and SearchBot does not go to infra. Malfunction mail: claudebot@anthropic.com from a mailbox on that domain.
Explore will not run Claude. Essential $95/mo is the first paid Claude desk. Named without a link stays a mention. ChatGPT still has 820M+ weekly active users on the same sheet. Claude is a separate prompt list. We do not average the two engines into one visibility number. Averaging two engines hides the one that lost, and the one that lost is the one the retainer fixes.