Why does Claude score low in my AI visibility report?

A low Claude score usually reflects how the question was asked, not a verdict on your brand.

Claude answers in one of two states. With web search off, it answers from its training data, which is frozen at a knowledge cutoff. With web search on, it can fetch live pages and cite them. Those are different products as far as your brand is concerned.

Three things push a Claude score down:

• The tool asked Claude with search off. A tool that calls the Claude API without enabling the web search tool gets a training-data answer. If your brand is new, small, or changed recently, that answer can't know about it.
• Claude chose not to search. Even with search available, Claude decides per query whether to use it, and the integrator's system prompt steers that choice.
• Anthropic's crawlers can't reach you. If Claude-SearchBot is blocked by robots.txt or a firewall, your pages can't be in the results Claude draws on.

The first two are outside your control. The third is yours to check and fix. Our page on whether Claude searches the web by default covers the toggle in more detail.

As of October 2026, Anthropic describes web search on Claude.ai as a user toggle and publishes no default. The Claude API is a separate surface: an organization's setting there does not describe what consumer users of Claude.ai see.

That creates a measurement gap. Real users ask with search on, with search off, or in apps built on the API with their own system prompts. A visibility tool picks one setup and reports it as "Claude".

What each state can show

If your tool ran with search off, a missing citation is expected. Citation rate isn't scoreable in that column at all. This is one reason AI visibility tools disagree about the same brand: they ask in different states and average different things.

This is the part you control. Anthropic documents three agents, and each does a different job:

• ClaudeBot collects content that may be used for training.
• Claude-SearchBot supports search quality.
• Claude-User fetches pages when a user's request triggers it.

robots.txt rules are per bot. A blanket AI block written to stop training often removes search access too. If you want training excluded but search allowed, say so per agent:

User-agent: ClaudeBot
Disallow: /

User-agent: Claude-SearchBot
Allow: /

robots.txt is only half the gate. A firewall or CDN can block a crawler that robots.txt allows, often with a 403 response that no dashboard reports. Anthropic also warns that IP blocking impedes its ability to read your robots.txt file. From 15 September 2026, Cloudflare blocks its Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers, so access can change without anyone editing a file.

The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.

One more access check: Vercel's crawler study found that Claude's crawlers fetch JavaScript files but don't execute them. If your pricing or product facts only appear after JavaScript runs, serve them in the initial HTML. Our Claude SEO guide covers rendering in more depth.

Nothing published measures how often real Claude users see your brand. Any single AI visibility score, for any engine, averages answers given with and without live retrieval, in proportions nobody outside the vendor can observe.

The citation unit also differs by engine. Anthropic's web search tool cites text spans, with a cited_text field of up to 150 characters, and returns each result's url, title and page_age. Perplexity cites whole sources by index. A "Claude citation" and a "Perplexity citation" aren't the same thing, so blending them into one number hides more than it shows. See citation rate versus citation share for how those metrics behave.

Citations also undercount influence. A page can be retrieved and shape an answer without being cited. If you use the API directly, log retrieved and cited sources separately to see the gap.

What the evidence doesn't support: rewriting pages in an "AI-friendly" style, adding schema, or publishing an llms.txt to lift a Claude score. None has credible evidence behind it as a route into Claude's answers.

Work through these in order:

1. Ask your tool how it called Claude. Search on or off? Which model? Which system prompt? If it can't tell you, treat the Claude column as a training-data snapshot.
2. Check robots.txt per agent. Confirm Claude-SearchBot isn't caught by a blanket disallow.
3. Check the firewall and CDN. Request a page with each Claude user agent and look for 403s or challenge pages. Recheck after any Cloudflare plan or setting change.
4. Check rendering. Compare the raw HTML with the rendered page. Move load-bearing facts into the server response.
5. Check the facts Claude states. A wrong price or outdated product name is a fixable problem. Make sure your own site states each fact once, consistently, on a page that can be fetched.
6. Track the trend, not one run. Compare like with like: the same prompts, the same retrieval state. Our page on tracking Claude's answers about your brand sets up a repeatable method.

You can fix access, rendering and your own facts. You can't control whether a user flips the toggle or whether Claude chooses to search.