What is measured, and what the number does not mean.

What we read

Your site’s own public pages, in one pass, with nothing installed and nothing connected: /robots.txt, the XML sitemap, /llms.txt, and the rendered HTML of the home page for its structured data, headings, prose and outbound identity links.

Nothing is authenticated. There is no admin to connect, no analytics to link and nothing to disconnect afterwards, because nothing was connected. Every source is one an AI search fetcher would read.

Any website. There is no platform requirement and no feed to publish: the engine reads what a crawler reads, so a Shopify store, a WordPress site and a hand-written static site are all measured the same way.

Complete report additionally crawls. 5 of the 36 checks need more than one page (pages are not duplicates of each other, link profile, dead links on the page, pages nothing links to, every page, crawled and listed one row each), so they run only on that tier.

How the score is built

36 checks sit in five parts of one score. Each check returns a ratio between zero and one rather than a pass or a fail, so a site with most of its structured data in place is not scored the same as one with none. Checks are weighted within their layer, layers are weighted against each other, and the result is rounded to a single number out of 100.

Layer weights are “Can they reach you” 28, “Can they read the facts” 26, “Is there anything to quote” 22, “Why you, not a competitor” 14, “Does the site load well” 10. The ordering is deliberate. A site AI search cannot reach gets no benefit from immaculate copy, so whether it can reach you at all dominates. Reasons to pick you over a competitor matter but are the easiest to add, so they carry least. Every finding in the report arrives with what it measures, the number it scored and how to move it.

A site that cannot be reached does not get a score. The run is gated before anything is scored: if fewer than 60% of the checks complete, the report says the scan failed rather than publishing a number. A site that was merely down once scored a confident 22 out of 100 with entirely plausible findings underneath it, which is a worse failure than no answer.

Where the questions come from

The buying questions in the report are derived from your own pages, not invented and not chosen by you. The scan builds a profile of what the site says it does, and one language-model call turns that into the questions a buyer would type before they had heard of you.

A question may never contain your brand name. Naming you guarantees a mention and measures nothing, so the rule is enforced in code rather than left to the model, which breaks it about one time in ten.

They are then ranked twice: whether the question is winnable at all, and how much of it your wording already covers. Broad category questions rank last because listicles hold those answers and rarely give them up. Deriving nothing is a valid answer; a wrong question is worse than no question.

How we count mentions

Complete report asks the derived questions of Claude, ChatGPT and Gemini, with web search on, and counts how often the answer names you. Each question is asked up to twice of Claude, twice of ChatGPT and five times of Gemini, 9 answers at most. The first 5 answers come first; when all of them agree, named or not, the question stops there. The reported rate carries the confidence interval that goes with the number of answers actually taken.

Asking once is not cheap measurement, it is not measurement. With two AI search engines and one sample each, every rate you can possibly get back is 0%, 50% or 100%: three coin flips wearing a percentage sign. These models are stochastic, so a single-shot check can tell you that you are absent from an answer that names you two times in five. That is a wrong answer you paid for.

This is most of why Complete report costs $25 rather than $5: the tokens are spent every time someone buys one. We also do not run the cheapest available model. The engine stands in for AI search as a consumer uses it, and consumers are not using the cheapest model, so that saving would buy margin by making the number less true.

Known limits

Most of it is one page. The scan reads the home page for structure, prose and trust signals. It is a fair proxy on a consistent template, but a site whose home page is a splash screen and whose substance lives three clicks deep will score worse than it reads.

It measures legibility, not demand. A perfect score says AI search can read and quote you. It does not say anyone is asking the question, and it is not a prediction of traffic, ranking or revenue.

Answers move. A measured rate is a measurement of a day. These systems re-index, change models and change retrieval behaviour, so the useful comparison is the same questions measured again later, not the number on its own.

No AI search engine publishes its ranking function. The checks are built from schema.org, published crawler documentation and the eligibility requirements for machine reading. They are a well-grounded proxy for machine legibility, not a reconstruction of any specific system’s scoring. The mention rate, by contrast, is not a proxy at all: it is the thing itself, counted.