What the check measures
Two facts are read out of every answer, and they are never added together. If you read one thing in this section, read the second heading.
What happens after you type your address
Five steps, in this order, and you see the output of step two before any money is spent.
- Your homepage is fetched and read. No AI is involved in this part. It is a plain read of the page, to work out your name, what you sell and roughly where you are.
- Up to ten questions are written — the kind of thing a customer of yours would type into an assistant. They come off your own page, and they are shown to you on screen before anything is asked.
- ChatGPT and Gemini are each asked those questions. Each question is put to each assistant once.
- Every answer is read for two things: whether you were named, and whether a page of your site was fetched and used as a source. Those are separate facts. The next section is about why.
- Every planned run is accounted for, including the ones that never came back. Nothing is dropped quietly, and nothing that failed is scored as an absence.
A category is only ever read off your page. If the page does not say what the business does, the category questions are not written and a note says exactly that. And if what comes back is not really a homepage — a security challenge screen, a CDN error page, a “just a moment” interstitial, all of which happen to real sites — the check refuses and tells you what it got, instead of asking two AI assistants ten questions about an error message.
Mentioned and cited are two different facts
This is the one idea in the whole product worth five minutes of your time, and most people meet it here for the first time.
Mentioned — it said your name
The assistant named you in its answer without ever fetching your website. It is answering out of what it already holds about your business. Your name is in there. Your current pages are not necessarily in there at all.
Cited — it read a page of yours
The assistant fetched a page on your domain and used it as a source for that answer. This is the strong result, and it is the only one that shows a page of yours was read rather than remembered.
These two are never added together into one figure. A brand that is remembered but never read and a brand that is read every time would land on the same number, and they have nothing in common — not the problem, and not the fix.
An assistant can write a web address into an answer from memory, having fetched nothing. That is not a theory: it happened in the first real run this product ever made. An assistant with search switched off wrote a domain into its answer, and an early version of the scoring counted it as a citation. It does not any more. A web address only counts as evidence of a fetch when the assistant’s own answer shows it actually searched. When it does not, the answer is scored mentioned — the weaker state, and the true one.
Why the difference decides what you do next
Two states, two entirely different jobs:
- Named but never cited is a plumbing problem before it is a writing problem. Something knows who you are and is not reading your site. The questions to ask, in order, are whether an assistant is allowed to fetch your pages and whether your words are actually in the page when it arrives.
- Missing altogether is a different thing: nothing remembered you and nothing read you for that question. The useful half of that result is the list of who was named instead.
- Cited means the whole mechanism worked at least once on that question. Note which question, because that page is doing something the others are not.
What to do about each of them is its own page.
The four states
Every question, on every assistant, ends in exactly one of these. There is no fifth, and there is no in-between.
A page of yours was used as a source
Fetched and quoted. The strongest thing this check can tell you.
You were named, from memory
Your name appeared. Your site was not opened to produce the answer.
You were not in the answer
A real, measured absence. Somebody else was offered instead.
The question never got an answer
Nothing was established. Left out of every count, never scored against you.
Counts, with what they are out of
Every figure in a result carries its denominator. Named in 3 of the 10 questions we asked. Cited on 2 of 4 measured runs. A count printed without what it is out of is a number waiting to be misread, and the misreading always runs in the flattering direction.
There is no single blended figure and there will not be one. There is no position number either: an AI answer has no positions in it, so a rank would be something invented rather than something measured.
Who was named instead
When you are absent from an answer, the other businesses in it are the ones your customer was actually offered. Those get counted twice over: once as every competitor named anywhere in the scan, and separately as the ones named on the questions where you were missing. The second list is the one that costs you something, so it is kept apart from the first.
Where the check runs
On our servers, and it has to be. A web browser cannot ask ChatGPT a question on your behalf, and it cannot
read another website’s robots.txt either — the browser is not allowed to see the response.
So anything that takes a web address and goes and gets it happens on our side.
The four generators are the exception: the FAQ schema, Organization schema, robots.txt and
llms.txt builders fetch nothing, and what you type into them never leaves the tab. They say so on
their own pages, and nothing else here does.
If this page did not answer it
Write to the one inbox. There is no chat widget, no ticket queue and no phone number — a message arrives with the person who wrote the thing you are asking about, and gets answered by hand. Send the domain you checked, what came back, and what you expected instead. That difference is the whole message.
Any question that arrives twice becomes a page here, so writing in is also how this section gets longer.
The fourth state, what produces it, and why it is kept out of every count.