How to Check Your Own AI Search Visibility Score (By Hand, No Tools)

Your page can rank on page one of Google and still never get said out loud by ChatGPT, Claude, or Perplexity when someone asks the exact question it's supposed to answer.

Those are two different systems measuring two different things, and almost nobody checks the second one. A search engine indexes a page and returns a list of links. An answer engine does something stranger: it synthesizes a response from whichever sources it judges worth quoting, and then either says your name or doesn't.

You can hold a perfectly good ranking and still be invisible to the second system, because ranking well and being citable are not the same skill. Here's the method, and you don't need any tooling to run it.

The method

• Write down the actual questions a buyer would type, not keywords. "best X for Y," "where to buy X," "X vs Y," "is X worth it." Five to ten is plenty to start.
• Type each one into ChatGPT, Claude, and Perplexity separately, fresh conversation each time, no leading context.
• Read the answer and note two things: did your name appear in the text (not a link in a sources list - the actual prose), and if not, who got named instead.
• Score it strict, yes/no, per query per engine. Not "sort of mentioned," not "implied." A buyer either heard your name in the answer or didn't, and that's the bar.
• Repeat it monthly. Answer engines reshuffle sources as their indexes and retrieval change; a score from six months ago tells you nothing about today.

That's the whole technique. It costs you twenty minutes and some typing.

Why the gap exists at all

We ran this at scale against 111 agent and MCP tool vendors - ten buyer-intent queries each, across the major answer engines - and the spread was blunter than expected: 59 of the 111, more than half, never got named once. The median vendor scored zero.

These are not bad products. Several are the actual best answer to the query they were tested against. They're invisible to this one channel because nothing on their site gives an assistant something extractable to quote.

That's the actual mechanism, and it explains why ranking and citation diverge: a search index rewards a page that contains the right words. An answer engine needs a claim it can lift cleanly and attribute - which in practice means a direct answer stated plainly near the top of the page, FAQ-shaped Q&A pairs rather than paragraphs the model has to summarize on the fly, and (increasingly) an llms.txt file pointing a crawler at which pages are the canonical source for which questions.

A page that ranks fine can fail all three and still show up nowhere in a synthesized answer.

What to do with a zero

A zero isn't a verdict on the product - it's a measurement that the surrounding content has never handed an assistant anything quotable. The fix is structural, not promotional: restate your core claims as direct, standalone answers; add real FAQ pairs that match how people actually ask the question, not how your nav bar phrases it; and make sure something is explicitly scoped as the citable source so the model isn't guessing.

Run the five-step method above by hand first - it's free and it tells you exactly where you stand before you build anything. If you get to the point where you're tracking this weekly across a catalog of pages instead of checking one page once, that's the specific job we built the AI Search Visibility Kit for - but the method itself doesn't need it to work.