How to Show Up in Claude Search (2026)

In our directory, "Claude SEO" packages rank below a robots.txt that names the right agents. Anthropic already split them. Official help.

The split is the whole game. A single "block Claude" rule that matches the string catches three different bots doing three different jobs. One is search indexing. One is a live fetch for a person asking. One is training. Blocking all three because you wanted to block one is how a site opts out of the answer it actually wanted. Naming the agents is free. The packages that sell "Claude optimization" without first checking the robots.txt lines are working on the wrong layer.

Show up means the stored Claude answer cites a URL you control, after SearchBot was allowed to index. A name without a link is a mention. Wrong KPI.

We repeat this because it is the most common mistake in a GEO brief. A mention is brand recognition. A citation is a link a person can click. Buyers click citations. Mentions do not send traffic. If the goal is traffic and conversions, the metric is the cited URL, not the name. If the goal is pure awareness, the mention counts. Most briefs want the first and report the second.

Rank 0: three user-agents

Claude-SearchBot: indexing for search quality. Disable it and Anthropic says they cannot index you for search optimization, which may reduce visibility and accuracy in user search results. This is the lever for this title.

If the page is not indexed, it cannot be surfaced in search. Disabling SearchBot is the single setting that removes you from the pool. Everything else in this guide assumes SearchBot is allowed on the money host.

Claude-User: live fetch when a person asks. Disable it and they cannot retrieve the page for that query, which may reduce visibility for user-directed web search. Unlike ChatGPT-User, Anthropic says these bots honor robots.txt. Crawl-delay is supported (their example is Crawl-delay: 1 on ClaudeBot).

Claude-User is the fetch that happens when a person types a question and Claude goes to read a page right now. It is not the indexer. It is the live reader. Honoring robots.txt is the difference from OpenAI's user agent, and it matters because the same Disallow line that blocks training also blocks the live fetch if you put it on the wrong agent.

ClaudeBot: training. Restrict it and future materials should be excluded from training datasets. Independent of SearchBot. A WAF that only matches "Claude" as one string is how teams block Search while thinking they blocked training.

ClaudeBot is the training crawler. Restricting it tells Anthropic to keep your pages out of future model training. That is a legal and policy call, separate from whether you show up in search. The independence is the point: you can allow SearchBot and restrict ClaudeBot. A WAF rule keyed on the substring "Claude" cannot tell them apart, so it blocks all three.

Repeat the directives on every subdomain. Anthropic says that out loud.

A robots.txt on the marketing host does not cover the docs host or the app host. Each subdomain reads its own file. If you only publish on www, you leave the apex and the docs subdomain as disallow-by-accident. Repeat the file.

Rank 0.5: stop IP-blocking as an opt-out

They say blocking IPs may not work correctly or persistently, because it can stop the crawler from reading robots.txt. Opt out in robots.txt. Report issues to claudebot@anthropic.com from a domain mailbox. They will not bypass CAPTCHAs. A challenge wall is a hide, not a ranking tactic.

IP blocking is the wrong tool for two reasons. First, it can prevent the bot from ever reading robots.txt, which means your allow or disallow never gets seen. Second, IP ranges change. The correct opt-out is the robots.txt directive on the right agent. If the bot misbehaves, the report address is claudebot@anthropic.com, sent from a mailbox on the domain so they can verify ownership. CAPTCHAs that block the bot also block indexing, so they hide the page rather than rank it.

Rank 1: software after the allow

Promptwatch first in the rankings: Claude on paid UI checks and page-level citations on Essential at $95/mo. ClaudeBot appears in Agent Analytics on Professional at $245/mo or an agency plan. Explore skips Claude (ChatGPT only). Daily paid checks, not instant alerts.

After the allow-list is in place, you need a record of the answer. Promptwatch stores the answer per prompt per check, daily, from the real UI. Essential is the first plan that includes Claude. Explore does not run Claude, so it cannot be the Claude desk. Professional is where the crawler logs show ClaudeBot, which is how you confirm the fetch happened before the citation.

Otterly $29: mentions. Profound Starter: ChatGPT-only. Peec AI last.

These three are not Claude programs. Otterly at $29 is a mention thermometer. Profound Starter is ChatGPT-only, so it is the wrong engine for a Claude brief. Peec AI stays last in our ranking. None of them replaces the robots.txt allow-list.

Google's optimization guide is homework on unique pages. Not Claude's algorithm. No llms.txt requirement on Anthropic's help page.

The Google guide is good homework for page quality. It is not Claude's ranking algorithm. Anthropic's help page does not require an llms.txt file, so adding one is optional, not a prerequisite for being indexed.

Ranked checklist

User-agent: Claude-SearchBot allowed on money host and www.
Claude-User allowed if live fetches should work.
ClaudeBot decided with legal, separately.
Public HTML, claim in text.
Two weeks of frozen prompts on a paid plan. Cited URL column.
If never cited, read robots first. On Professional or an agency plan, check logs before you rewrite.

The checklist is ordered. The allow-list comes before the content, because content that cannot be fetched cannot be cited. The two-week freeze gives the tracker time to collect daily checks and show whether the citation appeared. Reading robots before rewriting prevents the common error of editing a page that was simply never indexed.

Show up is a stored answer plus an allow-list you can screenshot. Anthropic wrote the allow-list. The tracker stores the answer.

Repeat SearchBot and User on www and the apex, plus any app subdomain that should appear. A robots.txt only on the marketing host leaves the docs host as Disallow-by-accident. Crawl-delay is their rate hint (Crawl-delay: 1 on ClaudeBot in the example). Not a ranking factor. They will not bypass CAPTCHAs. A challenge wall is a hide.

Mail claudebot@anthropic.com from a domain mailbox if you think the bot is malfunctioning. Do not IP-block as the opt-out. They said that can stop robots.txt from being read.

Named without a source URL is a mention. Freeze prompts two weeks on Essential or above. Explore will not run Claude. If SearchBot is Disallow, you opted out. Paid checks are daily. Otterly $29 will not replace robots.txt. Profound Starter is ChatGPT-only. Wrong engine. Product: promptwatch.com.