How to track your brand in Claude

To track your brand in Claude, send your buyers' questions to a Claude model with the web_search server tool enabled and check each answer plus its structured search-result blocks for your brand and your competitors — on a fixed prompt set, daily. Claude's audience skews technical and professional, so it punches above its raw traffic share for B2B, developer and infrastructure categories.

Last updated 2026-07-28

How do you query Claude for measurement?

With the web_search server tool enabled on the Messages API. Claude then issues real search queries, reads results, and returns structured web-search result blocks alongside the answer — which is what you read citations from.

Without that tool, Claude answers from training data, and the resulting citations are unreliable in the usual way. AI Visibility Tracker enables the search tool on every Claude call and flags in Settings any selected model that is not grounded, rather than silently collecting data that cannot be compared with the rest of your run.

What does Claude cite?

Documentation, comparison content, technical writing and community discussion, with a noticeably conservative and explicit attribution style — answers frequently name where each claim came from, which makes citation-based measurement unusually clean on this engine.

That same carefulness produces a specific measurement quirk: Claude often discusses a product substantively without linking to it, particularly when the discussion is drawn from general knowledge rather than a retrieved page. Checking only cited URLs will undercount your Claude visibility, which is why brand detection has to read the answer text as well. The mechanics are in how do AI assistants choose their sources.

When does Claude deserve a full weight in your score?

When your buyers are technical. In developer tools, infrastructure, security, data and professional-services categories, Claude's share of the actual purchasing conversation is far higher than its share of consumer chat traffic, and treating it as a minor engine will hide real losses.

For consumer and local categories the reverse holds, and a lower weight is honest. Either way, keep the per-engine numbers separate rather than blending them — the point of per-engine citation rate is that the spread between engines is the actionable part.

What is Claude-specific about reading the results?

Strong in developer, technical and professional categories — weight accordingly rather than assuming traffic share.
Web search must be explicitly enabled; ungrounded Claude reflects training data, not the live answer surface.
Careful attribution style means a brand is often named without being linked — check answer text as well as cited URLs, or you will systematically undercount.
Claude tends to hedge and qualify recommendations more than other engines, so a mention is sometimes a caveat rather than an endorsement. The raw response is worth reading, not just the citation flag.
Two consumer tiers exist (a default Sonnet-class model and a heavier Opus-class one) and their answers can differ. Pick one and hold it, or your trend is measuring the model change.

Step by step

1Write down the questions your buyers actually ask — discovery ("best X for Y"), comparisons ("A vs B"), alternatives ("A alternatives"), and brand checks ("is A worth it"). Ten to fifteen core prompts is enough to start.
2Build a brand profile: every name your brand goes by, plus the domains that belong to you. This is what a citation is matched against, so spelling variants matter.
3Send each prompt to the engine with live web search enabled — an ungrounded answer reflects stale training data and can cite pages that do not exist.
4Check the answer text AND the structured citation list for your brand and your competitors. A mention without a link and a link without a mention are both visibility.
5Repeat daily on the same prompt set, and read the trend rather than single answers.
6Record which domains are cited on the prompts you lose. That list, ranked by frequency, is your action plan.

Frequently asked questions

Can I just ask Claude whether it recommends my brand?

Not reliably. Answers vary run to run, and rephrasing the question changes the result — so a single check tells you about one sample, not about your visibility. Meaningful tracking asks the same fixed prompt set every day with web search on, then reads the rate over weeks. See what is Share of AI Answer for the formula.

Do I need API access, or can I do this in the chat app?

You need API access with web grounding enabled. The chat apps personalise answers with memory, history and location, which is the opposite of what a measurement baseline needs — and manual checking stops scaling past a handful of prompts. One OpenRouter key can carry every engine, or you can use each vendor's own API key directly.

Does Claude cite fewer sources than Perplexity?

Generally yes. Perplexity is retrieval-first and attaches many citations per answer; Claude synthesizes from a smaller retrieved set and attributes carefully within it. Neither is better for measurement, but it means a Claude answer offers fewer competitor data points per prompt than a Perplexity one.