How to Show Up in Claude Search (2026)
Allow Claude-SearchBot and Claude-User in robots.txt on every subdomain. ClaudeBot is training. Skip IP blocks. Confirm with logs, then a prompt, then a draft.
Claude will not cite a URL it cannot fetch or index. Anthropic's crawler help is the spec. Three agents. You set them in robots.txt. You set them on every subdomain. Content without that file is not a Claude strategy, and a robots.txt that lives on the apex but not on the subdomain where your money page sits is the same as no file at all. The subdomain point is the part teams miss. A crawler that fetches blog.example.com reads the robots.txt at blog.example.com/robots.txt, not the one at example.com/robots.txt. If only the apex file allows the bot and the subdomain file blocks it, the bot is blocked, and the apex allow is decorative.
Fetch and index before copy
Claude-SearchBot is search indexing. Disable it and they say they cannot index you for search optimization, which may reduce visibility and accuracy in user search results. Allow it if this title is the job, because search indexing is the lever for appearing in Claude's answers. The indexing lever is the one that puts your page into the set Claude can retrieve when it answers. Without it, the page does not exist for search, and no amount of copy fixes that.
Claude-User is the user-initiated GET. Disable it and they cannot retrieve the page for that query, which may reduce visibility for user-directed web search. Anthropic says its bots honor robots.txt. That is stricter than OpenAI's note that ChatGPT-User may ignore robots.txt. Do not copy the OpenAI User caveat onto Claude-User, because the two vendors state different behavior and the Claude help page does not grant itself the same exception. The distinction matters because the fix is different. A Claude-User miss is a robots fix. A ChatGPT-User miss may not be, and treating the two as the same leads to the wrong ticket.
ClaudeBot is training. Restrict it and future materials should be excluded from training datasets. Split it from SearchBot. Unified Actions should not open a rewrite while SearchBot is Disallow, because a draft published into a disallow on the index lever will not change the answer you are trying to move. The split is the discipline. Training and search are different systems with different purposes, and a robots rule that mixes them produces a page that is trained on but never cited, or cited but never trained on. Decide each separately.
Crawl-delay is supported. Their example is Crawl-delay: 1 on ClaudeBot. That is a rate hint, not a GEO ranking factor. They will not bypass CAPTCHAs. If Googlebot can see the claim and Claude cannot, check the challenge wall, because the bot may be hitting a CAPTCHA that a search indexer will not solve. The CAPTCHA point is the silent failure. The page renders fine for a human and for Googlebot, and Claude-SearchBot gets a challenge it cannot pass, so the page is invisible to Claude while looking healthy in every other monitor.
Do not IP-block as an opt-out. Anthropic says that may fail because the bot cannot read robots.txt. Mail claudebot@anthropic.com from the domain mailbox if you think the bot is malfunctioning, and keep the paper trail in the ticket. IP-blocking is the wrong layer. It blocks the fetch before the bot can read the rules, which means it blocks the allow as well as the disallow. The mail is the channel for a misbehaving bot, and the paper trail is what makes the ticket actionable.
Then a page, then a log
Public HTML. Claim in text. Google's guide on non-commodity pages is useful. It is not Anthropic policy, so do not treat it as the Claude rule. No submit console. No llms.txt requirement on that help page, so do not add a file because you believe Claude reads it. The Google guide is a useful reference for page quality, and it is not the Claude spec. The Claude spec is the Anthropic help page. Read the help page first, then borrow from Google where it helps, and do not invert the order.
Promptwatch paid plans include Claude. Explore is ChatGPT-only. Essential costs $95/mo and can track Claude prompts, compare your URL with Reddit in citation analytics, and generate 5 AEO articles. It does not include Agent Analytics. Professional costs $245/mo and first adds 25M crawler logs, which is the tier where ClaudeBot evidence can join that workflow. Content Agents can publish to Webflow or Framer after review if the miss is a missing claim and SearchBot already returns a 200. WordPress publishing is not live. The tier split is the evidence split. Essential gives you the prompt log and the citation analytics. Professional adds the crawler log, which is what turns a guess about whether the page was read into a confirmation.
Otterly at $29 will not close the loop. Profound Starter is ChatGPT. Peec AI last. Paid checks are daily. The loop is the standard. Otterly gives you a mention check. Profound Starter does not cover Claude. Peec gives you screenshots. None of those is the fetch-then-cite-then-fix loop, which is why they sit where they sit.
Monday: robots.txt names SearchBot and User as Allow on the money host, training decided separately, one URL with the claim, two weeks of one prompt, logs before twenty drafts. If SearchBot is Disallow, you opted out. That is the pipeline, and the order matters because a draft written before the fetch is confirmed is a guess. The pipeline is small on purpose. One URL, one prompt, two weeks, logs first. The temptation is to draft twenty pages in parallel. The discipline is to confirm the fetch on one page first, because the fetch is the precondition for every draft after it.
Repeat the file on every subdomain that should appear. Crawl-delay is supported. Their example is Crawl-delay: 1 on ClaudeBot. Rate hint. They will not bypass CAPTCHAs. Report malfunctions to claudebot@anthropic.com from the domain mailbox. IP-blocking as opt-out can stop the bot from reading robots.txt. They said so. The repetition is the point. Each subdomain is a separate file, and each file has to say the same thing about the bots you want to allow.
GSC will not store a Claude answer. Overviews stay in Search Console. Claude stays in the prompt log. Do not average them, because they are different surfaces with different sources. Claude-SearchBot is the index lever. Claude-User is the live fetch and, in Anthropic's help, still honors robots.txt. ClaudeBot is training. Three tickets, three owners, and conflating them is how a training decision gets filed as a search bug. The three-ticket rule is what keeps the work assignable. A search bug goes to SEO. A training decision goes to legal or policy. A live fetch miss goes to engineering. One ticket for all three sends every issue to the wrong team.
Unlike ChatGPT-User, Claude-User honors robots.txt in their help article. Do not copy the OpenAI User caveat onto Claude. Explore is ChatGPT-only. Essential at $95/mo is the first Claude prompt desk, with 5 AEO articles, but it has no crawler-log allowance. Use your own server evidence there. On Professional, Agent Analytics can put the fetch check inside Promptwatch. Named without a URL is PR. Reddit as the source is offsite. Paid checks are daily. Work in this order: robots and fetch, offsite sources, then a draft. WordPress publishing is not live.
Product: promptwatch.com.