Track your visibility in Claude

本文目前只有英文版。站点其余部分已汉化，这篇内容尚未翻译。

Claude is the engine most AI-visibility tools leave out, and the reason is not that it is unimportant. It is that measuring it honestly takes more work than measuring the others, and the work is invisible to anyone comparing feature lists.

This page is about that work, because the design decisions are also the reasons you should or should not trust a number from any tool that reports on this engine.

How the probe works

Each tracked query goes to the Anthropic Messages API with the web_search server tool available. The response comes back with structured citations attached to the parts of the answer that came from retrieved pages, and those citations are the only thing our verdict reads.

Two details in that sentence are doing more work than they look like.

The search tool cannot be forced. With OpenAI's API you can require the model to use web search before it answers. Anthropic's server tools do not work that way — the model decides. That creates a specific and dangerous failure mode: ask about a brand, get a fluent answer assembled from training data, see no citations, and record "not cited" for a query the engine never actually searched. That would be a fabricated negative, and it would be systematically biased toward whichever brands the model happens to know from training.

We handle it in two ways. The prompt explicitly instructs a web search, which also lands on the engine's own documented triggers for searching. And if the response comes back with no successful search result in it at all, the probe raises rather than returns: the caller records unknown, and unknown is what shows on screen. A cell that says unknown is telling you something true. A cell that says "not cited" because nobody looked is telling you something false.

The tool version is pinned. Newer variants of the search tool default to a different response shape, nesting results inside additional wrapper blocks. Pinning the basic version keeps the flat, documented structure the extractor walks — an unglamorous decision that stops a silent extraction failure from looking like a run of honest zeroes.

What "cited" means here

A query counts as cited when one of the structured search citations resolves to your domain: host reduced, lowercased, leading www. dropped, and matched only when it is your domain exactly or a subdomain of it.

As with every engine we probe, the answer text is never read for brand mentions. Claude naming your product in a sentence with no retrieved page behind it is not a citation, and we do not score it as one.

Who else tracks Claude

As of 2026-08-12, the models Peekaboo's own FAQ and footer list as tracked are: ChatGPT (OpenAI), Gemini (Google), Google AI Overviews, Google AI Mode, and Perplexity. Claude is not among them.

That is the entire comparison, and it is stated here rather than on the other four pages in this section for a reason: on ChatGPT, Perplexity, Gemini and Google AI Overviews there is no gap to point at, and dressing up an overlap as a difference is how comparison pages lose their usefulness. Where a competitor covers the same surface we do, saying so adds nothing you can act on.

What the gap means in practice is narrower than it sounds, and worth being precise about: it means one tool's coverage did not include this engine on that date. It does not mean Claude matters more than the engines it does cover, and it is not an argument that Claude should outweigh ChatGPT in anyone's scoring — it should not, and in ours it does not.

What this cannot tell you

It is not the consumer app. The API's search tool is a proxy for Claude-with-web-search, and the label carries that qualifier everywhere it appears. The consumer product has no API and personalises per account; nobody measures it at scale without driving an automated client through a session that was not built to be automated.

Unknown is common here, and that is correct. Because the search step cannot be compelled, some runs legitimately come back with nothing to measure. Those land as unknown rather than being folded into either the cited or the not-cited side. It makes the number less tidy and it makes it true.

One reading is noise. Live retrieval varies run to run in every engine. The weekly series is the measurement.

There is no position number here, and there is not one to have. A results page has ten slots, so a rank is a row you either occupy or you do not. A synthesized answer is prose that names a few brands and links a few pages, and there is rarely a numbered position to track inside it. What this check reports is therefore presence and citation, not a row number, and nothing on this page converts into one. Where you would want a rank, the measurement doing that job is the trend — how often this engine cited you across weeks of the same questions, rather than where you placed inside any single answer. The word "rank" is still settling in this market: we use it in the sense set out under AI search ranking, and AI search rank tracking covers what tracking it involves.

There is no published cross-brand benchmark for this engine. What runs on your account is your own sweep: your tracked queries, through this engine, on the same weekly cadence as the other four. What does not exist is a public study across many brands that would let you ask how this engine behaves in general — how often it cites anyone, or which kinds of sites it tends to reach for. No number on this page comes from one, and we are not putting a date on when there will be one. The distinction is worth stating plainly, because "we do not measure this engine" and "we have not published a study about this engine" are different claims and only the second is true.

What to do with the result

Read the source list for the queries where you are absent. It names the pages the engine actually retrieved, which is a far more useful competitive set than whatever is ranking in blue links for the same words.

If your category is technical, one pattern is worth checking first: Claude retrieves documentation, changelogs and primary sources at a noticeably higher rate than marketing pages. If your docs live on a subdomain and your marketing pages are what you have been optimising, you may already be cited through the half of your site you were not measuring — which is exactly why the subdomain rule above counts them as you.

The free audit runs this probe against your domain and shows you the sources that came back. If you would rather query it from your own tooling, checking AI visibility with Claude covers the MCP route.