How the AI search visibility check works

Every check asks the default model behind the three assistants a buyer would actually open: ChatGPT, Gemini and Claude. Six questions each, so eighteen answers per run.

Why the full models and not the cheap tiers

We benchmarked both on a live check in August 2026. The mini and haiku tiers answered only 4 of 18 questions with any real business name, against 10 of 18 for the full models, and they were slower doing it (30 seconds against 16). A model that declines to name anybody tells you nothing about your visibility, so the cheaper panel would have reported a problem that was ours rather than yours.

The questions are written for your business, not pulled from a list. They are full first-person questions the way someone actually types them, not search keywords, and they are constrained in three ways that matter:

• No company names. Not yours, not a competitor’s. A question that names you would guarantee you appear in the answer and measure nothing.
• Your city, for a local business. Most of the six name the place, because that is how buyers ask.
• A spread of intents. General discovery, an urgent need, one specific service, a “who should I hire” comparison, and one from someone researching before they choose.

They are then reused for a week. A repeat check is comparable to the last one instead of moving the goalposts with a fresh set of questions.

The denominator is the part worth understanding. Assistants sometimes refuse to recommend any specific company and answer with general advice instead. Those answers are dropped from the denominator rather than counted against you, because they say nothing about whether you are visible. Counting them would make every business in every industry look worse than it is.

The four bands

Two choices that make the number repeatable

Sampling is pinned low. At the default setting, the same business scored 1, 3 and 3 mentions across three back-to-back runs. A number that moves that much between identical runs looks invented, so the randomness is turned down to make week-to-week comparison mean something.

Name matching is deliberately strict. Legal suffixes and common words like “the” and “company” are stripped before comparing, and a near-miss is treated as a miss. A false “you were mentioned” is worse than a missed one, because the entire report hangs off that single comparison.

When a model names nobody at all, it gets exactly one follow-up, the way a person would push back once before giving up.

The panel tells you if assistants name you. These checks tell you what is in the way, and every one of them is a plain page fetch you can verify in a browser yourself. No API keys, no estimates.

1. Can the crawlers get in

Your robots.txt is read and checked against ten named AI crawlers. Blocking one of these is the most concrete reason a business can be missing from an AI answer, and plenty of sites block them without anyone having decided to.

2. Is there an llms.txt

A plain-text file that states what your site is and what matters on it. The check requires it to be genuine plain text, because a styled 404 page served at that path would otherwise read as a real one.

3. Is there a machine-readable statement of who you are

Structured data on the homepage is read for the facts an assistant needs to recommend you by name: an Organization or LocalBusiness type, a postal address, a phone number, opening hours, linked profiles elsewhere, an aggregate rating, and FAQ markup.

4. Does the page itself say the basics

Reading the visible text, not the markup: does the homepage name the city you work in, does it state a phone number, is there any copy about the business beyond a headline and a form, and how many words are on the page. A homepage that is one hero line and a form gives a model nothing to learn.

• Volume. It measures whether you get named, not how many people are asking. Nobody has that number, including the assistants.
• What your buyers specifically asked. The six questions are representative, not a transcript.
• Why a model chose someone else. You can see who it named instead. The reasoning is not visible to anyone outside the model.
• Tomorrow’s answer. These systems change weekly. A check is a reading on a date, which is why the date is on the report and why the questions are held steady between runs.
• Whether a mention sent you anything. Being named in an answer is not a visit. Those are separate measurements and we keep them separate.