Before you measure, make sure they can read you.

Which tool to start with

Start with the free AI citation check. It answers the smallest useful question: when an assistant is asked what one of your pages is about, does that page show up in the sources, and who gets cited instead? You paste one page URL. Its heading becomes the question. It asks Perplexity with web search on, from London, and tells you the result. There's no account and nothing is stored. You get five checks a day per connection, so pick your pages with some care.

Next, run the AI readiness check. This one is about whether a machine can read you at all. It fetches your homepage, your robots.txt and your llms.txt once, as VisibilityToAI-ReadinessCheck, then runs six deterministic checks. No LLM is involved, so the same site gives the same result every time. If a check fails here, fix it before you read anything into a citation result.

Then the AI-crawler robots.txt check. This is the one people skip, and it's often where the real problem sits. It evaluates every AI crawler token the way that crawler reads the file: its own group if it has one, otherwise the * group, with the longest rule winning. So you see what each crawler actually sees, not what you assumed you'd written.

Last, the llms.txt generator. It runs in your browser, so nothing leaves your machine. Use it once the earlier checks say your basics are sound.

Also worth a look: the industry benchmarks. These are public runs of the same method, one page per category. Each uses a set of buyer questions, four engines, every answer read and every claim quoted. Each states its run date, its market and its prompt count. Vendor share is simply the share of answers that name the vendor. It's a good way to see what a full read looks like before you commission one.

What the free tools cannot tell you

Here's the honest gap. The free tools tell you whether you're readable. They can't tell you whether you're recommended.

A citation check on one page is one question, asked once, of one engine. That's a sample, not a census. Ask the same question tomorrow and you can get a different answer. And API answers sit close to what the consumer apps show, but they aren't identical. So a single clean result proves very little, and a single miss proves just as little.

Being recommended is a different matter. It depends on the questions your buyers ask before they know who to call, mostly unbranded, and on whether an assistant names you across a spread of them. That takes more than one URL. The Snapshot reads your site and proposes 15 prompts across 3 to 5 verticals. You edit and approve every one before anything runs, and you're charged only after you approve. Two of the prompts are branded, to check the engines know you exist. One is framed as a comparison. Run across ChatGPT, Claude, Gemini and Perplexity, that's 15 questions x 4 engines = 60 answers. The Deep Dive goes to 100 questions and 400 answers.

Each answer records three things: Mentioned, Cited and Rank. Mention and citation are never blended into one score, because they mean different things. Named but rarely cited is a page problem. One rival in the majority of answers means the market is Entrenched. No vendor named in more than about a quarter of answers means it's Open.

Even then, some questions stay unanswered. A citation is a source link, not a click, so none of this tells you about traffic. It can't say why a model named you. And it can't say whether the result will hold next month. Treat any report as a dated reading, and ask for the model ids behind it.