Claude

Claude AI Optimization Tool

Claude answers from training data unless web search is on, and in the Claude app search is a setting the user controls. 99Visibility asks Claude your buyers' prompts with search on and checks Anthropic's three crawlers on your site.

Last reviewed: 2026-10-16

9%of the coverage weight in the 99Visibility score. The weights are our model of reach, not a measured market share.

Start for Free

How Claude works for AI visibility

Claude searches only when search is on; otherwise it answers from training data. Our audit asks it with search on. Source: Anthropic crawler and web search tool documentation; how we ask, from our audit configuration.

Training data first

With search off, Claude answers from what the model learned in training. Anthropic lists June 2026 as the reliable knowledge cutoff for claude-haiku-5-5, the model we ask, so a brand founded after that cannot be in its answer.

Search when it is on

Through Anthropic's API, web search is a tool the developer adds, and Claude decides when to use it. In the Claude app, search is a setting the user turns on or off, and Anthropic does not publish a default. Which search index backs it is not documented in the sources we checked.

Three crawlers, three jobs

Anthropic runs ClaudeBot for training, Claude-SearchBot for search and Claude-User for pages a user asks Claude to fetch. Anthropic describes Claude-SearchBot as improving search result quality but does not state what blocking it does; our reading is that a block keeps your pages out of what Claude search draws on.

How it cites

With search on, Claude cites the passages it used, as quoted text of up to 150 characters through the API, and each search result carries a page_age field for when the page was last updated. Anthropic does not say whether page_age affects which pages are used.

JavaScript

Vercel's crawler study found Anthropic's crawler fetched JavaScript files on 23.84% of requests and executed none of them. Facts that appear only after scripts run were invisible to it. The study is a one-month snapshot from December 2024, run by a company that sells server rendering.

The crawlers that reach Claude

ClaudeBot (Claude model training)
Your content is excluded from Anthropic training crawls. Anthropic runs search (Claude-SearchBot) and user fetches (Claude-User) as separate agents.
Vendor-documented: support.claude.com →
Claude-SearchBot (Claude search)
Anthropic says this agent improves search result quality; blocking it likely keeps your pages out of the results Claude search draws from. Anthropic does not state the consequence outright.
Our reading of: support.claude.com →
Claude-User (Claude on-demand browsing)
Claude cannot fetch your pages live when a user asks about them. Anthropic documents the agent; the consequence of blocking it is our reading.
Our reading of: support.claude.com →

What is not documented

How a brand gets into Claude's training data. Anthropic does not publish what its training set contains or how sources are weighted.
Which search index Claude's web search uses, and how it ranks results. Neither is in the sources we checked.
How often Claude app users have search on. The setting has no published default.
Whether allowing ClaudeBot changes what a future Claude model knows about you. Nobody has measured it.

What 99Visibility does for Claude

How we ask it: Anthropic API, claude-haiku-5-5, with web search
Web search in our audit: On. Answers come from what it finds on the web when it answers
Weight in the score: 9% of platform coverage
Plans: Starter ($49/mo) and up
Crawlers checked: ClaudeBot, Claude-SearchBot, Claude-User: robots.txt, plus a firewall probe for each one that crawls

Asks Claude what it knows

Each keyword goes to the Anthropic API as claude-haiku-5-5, exactly as you typed it, with web search on. Claude is on Starter ($49/mo) and up.

Checks Anthropic's three crawlers

The access check reads robots.txt for ClaudeBot, Claude-SearchBot and Claude-User and requests your page as each one. A Claude-SearchBot block is ranked high; a ClaudeBot-only block is ranked low and labelled as training-only.

Flags wrong facts

Every Claude answer is read against the price, founding year, location and features in your brand profile. Each flag puts the claim beside your fact, with a severity of critical, moderate or minor.

Honest about citations

The passages Claude cites are stored with the answer, and the citation rate counts the share from your site.

A platform note in the fix list

When the Claude column scores more than 20 points under your overall score, the fix list adds a note for it. It tells you to check that the page is reachable by Claude-SearchBot, readable without JavaScript and free of snippet controls, then to look at what Claude cited instead.

Agent V for Claude

Agent V is the assistant in your dashboard, on every plan. It reads your audits, your accuracy flags and pages on your own site, so you can ask what an engine said for a prompt and which fix applies.

Ask for robots.txt lines and V drafts groups for ClaudeBot, Claude-SearchBot and Claude-User, with notes saying that search and training are separate agents. Ask what to write and it drafts page copy that states the corrected fact in its first sentence.

V can also check a page on your site or run a new audit. Each change it proposes shows up as a "V wants to" card and does nothing until you confirm, and V never edits your site: drafts are text you paste yourself.

How Agent V works

Claude optimization: what to fix first

Allow Claude-SearchBot in robots.txt if you want to be eligible for Claude search. Anthropic describes it as improving search result quality and does not say what a block does, so treat a block as a risk.

Decide on ClaudeBot separately. Your content is excluded from Anthropic training crawls. Anthropic runs search (Claude-SearchBot) and user fetches (Claude-User) as separate agents.

Put load-bearing facts in the HTML your server sends. Anthropic's crawler did not run JavaScript in Vercel's study.

Check your firewall as well as robots.txt. A bot-protection rule can refuse a crawler that robots.txt allows, and robots.txt will not show it.

Fix wrong facts on your own pages first. The search-on answer can change once Claude-SearchBot fetches the page again; the training-data answer changes only when Anthropic ships a newer model.

Claude optimization FAQ

How does Claude decide what to say about brands?

With search off, from training data, and Anthropic does not publish how brands enter it. With search on, from the pages its search tool retrieves, which Claude cites as quoted passages.

Can I change what Claude says about my brand?

With search on, yes, through access: Claude-SearchBot allowed and able to fetch your pages, and your facts in the HTML. With search off, only a future model's training data can change the answer, and no vendor publishes what goes into it.

Why do you ask Claude with search on?

To measure what Claude finds on the web, the way a user with search on sees it.

How is Claude different from ChatGPT in your audit?

In our audit both can read the web, so a fix to your site can reach either column once their crawlers fetch the page again. ChatGPT carries 36% of the coverage weight and Claude 9%.

Why is Claude not on the Free plan?

The Free plan audits ChatGPT and Google AI Overviews. Claude starts on Starter ($49/mo), with a 7-day trial that takes a card.

Sources

Anthropic documents three agents: ClaudeBot (training), Claude-SearchBot (search result quality) and Claude-User (fetches a user asks for).

Source: Anthropic crawler documentation (2026-04-07) →

ChatGPT and Claude crawlers fetch JavaScript files (ChatGPT: 11.50%, Claude: 23.84% of requests), but "they don't execute them. They can't read client-side rendered content."

Source: Vercel, The rise of the AI crawler (December 2024) →

Search results expose "page_age: When the site was last updated". Web search costs $10 per 1,000 searches plus tokens.

Source: Anthropic, web search tool documentation →

Anthropic lists a reliable knowledge cutoff of June 2026 for claude-haiku-5-5.

Source: Anthropic, models overview →

The other platforms we audit

See what Claude says about your brand

Claude is on Starter ($49/mo) and up, which starts with a 7-day trial. The Free plan audits ChatGPT and Google AI Overviews and checks AI crawler access, with no card.