Tracking Claude's Answers About Your Brand

"Claude" is at least three different setups, and each one can answer the same question differently.

• Claude.ai. Web search is a user toggle. Anthropic publishes no default for it. Users can also switch models.
• The Claude API. Web search is a tool the developer enables. The API default does not describe Claude.ai consumer behaviour.
• Apps built on Claude. Claude decides whether to search, and the integrator's system prompt steers that decision.

So any answer you collect is tied to one search state, one model and one prompt context. That's why Claude answers differently in different apps. It also means no tracker can tell you what a real customer saw in Claude.ai yesterday.

The next move: decide which setup you are measuring, write it down, and keep it fixed. For the consumer toggle in detail, see whether Claude searches the web by default.

As of October 2026, the Anthropic web search tool is the only documented way to see which pages Claude pulled for an answer. Anthropic's docs list what each search result exposes:

• url and title for every page retrieved
• page_age, described as when the site was last updated
• span-level citations, each with a short cited_text excerpt of up to 150 characters

Pricing is $10 per 1,000 searches plus normal token costs. Errored searches are not billed, and the citation fields do not count toward tokens. The fee counts searches, not prompts, so log how many searches each request ran. Then your cost per prompt is a measured number.

Say you track 100 buyer questions once a week, and each triggers one search. Search fees come to $1 a week before tokens. The tokens usually decide the real bill, so run a small batch first and price from that.

Two cautions. page_age shows freshness is surfaced, not that Claude weights it; no vendor says it does. And Claude cites text spans while Perplexity cites whole sources, so citation counts across the two are not directly comparable. Our review of Claude's web search tool docs covers the fields in full.

When search is off, Claude answers from what it learned in training. That's the parametric stage: knowledge frozen at a cutoff. There are no retrieved pages, so there is nothing to cite and no citation rate to compute. Treating a search-off answer as "zero citations" would record a failure that never happened.

You can still measure two useful things in search-off answers:

• Mentions. Does Claude name your brand for a buyer question at all?
• Accuracy. Are the facts it states right? Wrong prices, old product names and confusion with a competitor are the errors worth catching.

Be plain about control here. Nothing a publisher changes this quarter reaches a search-off answer. Fixes land, if at all, in a future model. With search on, your pages can enter the answer this month, which is why the two states belong in separate columns.

The next move: run each prompt both ways and diff the answers. The gap shows what Claude believes about you versus what it finds when it looks.

Here is a setup that produces numbers you can defend.

1. Pick buyer questions. Write the questions a customer would ask, not your brand name alone. Decide whether to include prompts that name competitors.
2. Fix the model and record it. Model changes move answers. Log the exact model string with every run.
3. Run search on and search off. Keep two series. Never blend them into one score.
4. Log retrieved and cited sources separately. A page can be retrieved and dropped. Both sets matter; the retrieved vs cited logging playbook shows how.
5. Score facts against your own record. Keep a list of verified brand facts and flag answers that contradict them.
6. Keep the same cadence. Weekly runs on a fixed prompt set give you a trend you can read.

API results are a proxy for the consumer app, not a copy of it. The setup is the same idea as Perplexity prompt tracking with the Agent API, where web search is also opt-in. If you want a quick snapshot before building anything, start with a free check of what Claude says about your brand.

Anthropic documents three separate agents:

• ClaudeBot collects content for training.
• Claude-SearchBot crawls to improve Claude's search quality.
• Claude-User fetches pages when a user's request triggers it.

They are separate rules in robots.txt. A blanket AI block that catches Claude-SearchBot removes you from the search side, not just training. A site that wants to stay out of training but remain eligible for search could write:

User-agent: Claude-SearchBot
Allow: /

User-agent: ClaudeBot
Disallow: /

Allowing the bot buys eligibility, not citation. Two more gates sit behind robots.txt:

• Your firewall or CDN. It can return 403 to a crawler that robots.txt allows. Anthropic warns that IP blocking impedes its ability to read your robots.txt file at all.
• JavaScript. Vercel's crawler study found Claude's crawler fetched JavaScript files on 23.84% of requests but did not execute them. Prices or product details that appear only after JavaScript runs may be invisible.

The next move: request your key pages as each Claude agent and see what comes back. The AI Crawler Checker does this and shows which agents your firewall turns away.

Some things no tool can see, including ours:

• What a real user saw. The toggle state and model behind a consumer answer are not exposed.
• Why Claude cited one source over another. Anthropic publishes no ranking factors. Research on generative engines finds topical relevance and position in the context list matter most, and the critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) found no technique with a stable, cross-platform effect on being discovered.
• Whether a citation reflects real influence. Citation and use are not the same thing.

Claude can still pick up pages other engines ignore. In Otterly's 2026 test of two AI-generated sites, Claude cited the sites after Google had deindexed them. So measure Claude on its own, not as a stand-in for ChatGPT or Gemini.

How much does it cost to track Claude through the API? As of October 2026, Anthropic charges $10 per 1,000 web searches plus standard token costs. Errored searches are not billed and citation fields do not count toward tokens. Search fees are usually small next to tokens, so run a short batch of your real prompts, log the searches and tokens each one used, and price your schedule from those measured numbers rather than an estimate.

Can I score Claude's citation rate if search is off? No. With search off, Claude answers from training data and retrieves no pages, so there is nothing to cite. Counting those answers as zero citations records a failure that never happened. Score search-off answers on whether your brand is mentioned and whether the stated facts are correct, and keep them in a separate series from search-on results.

Does allowing Claude-SearchBot get my brand cited by Claude? It makes your pages eligible, nothing more. Retrieval and citation are decided per query by Claude, and Anthropic publishes no ranking factors. Allowing the bot removes one block. You still need your firewall to let it through and your key facts to appear in the HTML before JavaScript runs, since Claude's crawlers do not execute JavaScript.

Should I block ClaudeBot or Claude-SearchBot? They do different jobs. ClaudeBot collects training content, Claude-SearchBot serves search quality, and Claude-User fetches pages on a user's request. You can set each one separately in robots.txt. If you want out of training but still eligible for Claude's search answers, disallow ClaudeBot and allow Claude-SearchBot, then confirm your CDN does not block it anyway.

Are API results the same as what Claude.ai users see? Not necessarily. Claude.ai users choose whether search is on and which model to use, and Anthropic publishes no search default. Apps built on Claude add their own system prompts, which steer when it searches. API runs are a controlled proxy. They are useful for trends and fact errors, but they do not replay any specific customer's answer.