The method
How we measure AI visibility, and what we do about it
Long version, for people who want to know what they are buying before they talk to anyone. Nothing here is a secret: if you wanted to do it yourself, this is what you would have to build.
- 6
- kinds of question, sorted by what the customer wanted
- 4
- assistants, each reported on its own
- 11
- pages behind a single answer, on average
01
Where the questions come from
A customer looking for a supplier does not type your name. They describe the situation they are in: what they need done, where they are, how soon, and how much they want to spend. Those are the questions we measure.
We write them from your services and your city, in the language your customers actually use — which in many cities means more than one. Then we sort them by what the person wanted: a service, a price, someone nearby, someone they can trust, or help right now. That split is what turns a single visibility figure into something you can act on.
Questions that name a company are kept apart and marked as brand questions. Someone who already knows your name is not choosing any more, so those answers tell you nothing about whether you get found.
How the questions are sorted
- Generalwho should I use for this in my city
- Servicea named job: implants, payroll, a leaking roof
- Pricehow much, and where it is cheaper
- Localnear me, in this district, on this street
- Trustreviews, guarantees, who to believe
- Urgenttoday, tonight, this week
02
Why we ask the same question more than once
Assistants do not answer identically twice. Ask the same question in the morning and in the afternoon and the list can change, even though nothing about your business has.
So we ask each question several times over and report a share — how often you were named — instead of a yes or a no. A share can be compared between measurements; a single answer cannot.
It also gives us a sense of the noise. Between two measurements with nothing changed, the figure drifts on its own. That is why a small movement we report as no change rather than dress it up as progress.
Two measurements, nothing changed in between
03
Why we never average the assistants
ChatGPT, Gemini, Claude and Perplexity read different sources and answer differently. In our own measurement, four assistants asked the same questions named 63 businesses between them — and only four of those were named by all four. Forty-six appeared in one assistant and in none of the others.
An average across four such numbers describes nobody. It hides the assistant you are absent from, which is exactly the one worth working on. Every figure we report belongs to one assistant.
63 businesses named across four assistants
- 4 named by all four
- 13 named by two or three
- 46 named by one, and by none of the others
04
What an answer is built from
An answer is assembled, not quoted. The assistant reads what it can find about your market — your own pages among them, and alongside them directories, maps, review sites and round-up articles with titles like “best accountants in Manchester” — and writes a short list out of all of it.
In the answers we measured, a single answer rested on eleven pages on average. Your own site is one of them, and it is the one source you control completely: it is where an assistant checks what you actually do, for whom, in which city, and at what price.
So the work runs on both sides at once. Your pages are made to answer the questions people ask, in plain words a machine can read without guessing. The pages other people keep about you are corrected so that they agree. Where the two disagree, the assistant usually believes the others — which is exactly why both halves are the job.
What one answer was built from
- Your own pagesservices, prices, hours, the plain facts about the company
- Directories and mapswhere your details are repeated, right or wrong
- Review siteswhat customers said, and how recently
- Round-ups“best accountants in Manchester” and its like
- Local mediaanything written about the market you work in
05
What we change
Your details, made identical everywhere. Name, address, hours, services and phone number differ across directories more often than not, and an assistant reading three versions of you will often use none of them.
Presence in the lists that get read. We find the round-ups and directories the assistants actually cite for your questions, and get you into the ones you are missing from — by the ordinary route: contacting whoever maintains them, with the correct information.
The pages on your own site that answer the questions people ask, marked up so a machine can read them without guessing, plus a page of plain facts about the company for the same reason.
What we do not do is invent reviews, generate filler articles, or place your name on pages built to be read by machines. Assistants are already learning to discount that, and you would be paying for a problem.
Three kinds of work
- Your details, identical everywhereName, address, hours and services differ across directories more often than not.
- Into the lists that get readThe round-ups and directories the assistants actually cite for your questions.
- Pages and markup on your own siteThe questions people ask, answered plainly, in a form a machine can read.
06
What you get, and when
The dashboard opens with the first measurement: every question, every answer word for word, who was named instead of you, and the pages those answers came from. Nothing is summarised away and nothing is estimated.
After each measurement there is a report — the same document, page for page, that you can read in full on the sample report page before you decide anything. Next to it the work log: what we changed, on what date, and what happened after.
First movement usually shows after six to eight weeks, because the pages the assistants read have to change first, and then be read again. If nothing moves, that will be in the numbers too, and we will say so rather than dress up a small swing as progress.
What lands, and when
- Day onethe dashboard, with every answer we collected
- After each measurementa report you can send on, and the work log with dates
- Six to eight weeksthe first measurement where a change could show
See where you stand
The free check takes one email and tells you whether the assistants name you at all.