Ask an AI assistant for “a good invoicing app for freelancers” and it answers in a paragraph, naming three or four products. More and more people start exactly there — where a Google search used to be. Which means there’s a new question anyone who makes anything should be able to answer: when someone asks an assistant about the problem you solve, do you come up?
This is what I work on. I’m part of the team behind Suparanku, an AI-visibility tracker built by Supasaito, the company I co-founded in Tokyo — and I want it to be that rare thing: a tool that’s powerful and simple and clear to use. That job starts with the words. This field is young, its vocabulary is younger, and every term sounds more solid than it actually is. A number you can’t define is a number you can’t act on.
So here they are: the eleven words behind every AI-visibility number, in plain language, in the order the numbers are made. No dashboard required — by the end you’ll be able to run the whole measurement by hand, and I’ll show you how.
Part one: how the number is made
Prompt
A prompt is a question, written the way a real person would ask it: “What’s a good invoicing app for freelancers in Japan?” Not a keyword — a full sentence. Nobody types “invoicing app japan” at an assistant; they ask, and the answer is shaped by the whole question.
Every AI-visibility measurement works the same way: someone writes a list of prompts, asks the assistants, and counts how often your brand appears in the answers. Every number in the rest of this article is built on top of that list.
Which means the list decides everything. Change the questions and the numbers change — with nothing about your business having changed. So before you trust any figure, look at the questions behind it. If they’re not the questions your customers would actually ask, the number is precise and useless.
Run
Ask an assistant the same question twice and you’ll get two different answers. That’s not a bug: an assistant doesn’t look up an answer, it writes one, from scratch, a little differently every time. Some versions of the answer will name you; some won’t.
That’s why a single answer proves nothing. To measure anything, you ask the same question several times and count. Each attempt is called a run. Ask one prompt six times and get named four times: that’s “4 out of 6”, and that little fraction is the honest unit this entire field is built on. The more runs behind a number, the more it means.
Visibility
Visibility is the share of runs in which your brand was named at all. Named — that’s the entire definition. Not praised, not recommended, not linked. Just: your name appeared in the answer.
Suppose you track three questions, six runs each (these figures are invented, for illustration):
| The question | Named in |
|---|---|
| ”best invoicing app for freelancers” | 5 of 6 runs |
| ”how do I invoice a client in Japan?“ | 3 of 6 runs |
| ”which invoicing apps handle Japanese tax IDs?“ | 0 of 6 runs |
Turn each row into a percentage: 5 of 6 is 83%, 3 of 6 is 50%, 0 of 6 is 0%. Your visibility is their average:
(83 + 50 + 0) ÷ 3 ≈ 44%.
Now you can see exactly what the headline number is made of — and two things follow. First, with only three questions, each one moves the total by a third: one bad question and the “score” collapses, one lucky one and it soars. Small lists make jumpy numbers. Second, that 0% row isn’t shameful. It’s the most useful row in the table: a specific question, asked by real customers, where assistants never think of you. That’s not a grade. That’s a to-do list.
Part two: what the numbers say about you
Mention and citation — the two that get confused
These are the two words in the title, and mixing them up is the most expensive confusion in this field.
A mention is your name appearing in the answer’s text. A citation is a link — the assistant pointing at your page as a source someone can click.
They feel like the same thing. They are not even close. Before announcing my knowledge base, I measured where I stood: ten questions, four assistants, scored by hand — 31 points out of a possible 80. The pattern behind the score was consistent: the assistants knew my work, described my work, even recommended my work. They just sent people somewhere else — or nowhere. Mentioned, not cited. And the brand-new knowledge base itself scored zero, for the least mysterious reason imaginable: nothing on the internet linked to it yet. That’s the full story, with every score on the table.
So when someone tells you “the AI knows your brand” — that’s a mention, and it’s worth something. But look at what each one leaves the reader with. A mention leaves them a name they may or may not remember to search for later. A citation gives them a door: a link, right there in the answer, that lands on your page. A customer is someone who visits, reads and decides — and to get the visit, a mention is almost never enough. If you want the customer, you need the citation. Getting mentioned is the milestone; getting cited is the goal.
Crawl and live fetch
A crawl is a program — a crawler — downloading your pages ahead of time: to learn from them, or to file them into a search index for later. A live fetch is the opposite of “ahead of time”: a person asked an assistant something seconds ago, and the assistant went to your page right then to answer them.
The reason you should care about the difference: on 29 July, Anthropic’s crawler read all 42 pages of my knowledge base in about an hour. Being read, cover to cover, by the company behind Claude — that felt like winning. On 30 July I asked Claude a question my knowledge base answers better than any other page online, because it’s about a component I built myself. Claude answered superbly… and linked a domain I no longer publish on.
Being read and being credited are not the same event — and they’re not even done by the same program. AI companies run different crawlers for different jobs: one reads to train the model, one reads to build a search index, one fetches live when a user asks. Which of them visited you changes what the visit means — and whether it can ever turn into a link. I mapped all three, logs in hand, in a separate article about the three kinds of AI crawler; if this section leaves you wanting the mechanics, that’s where they live. The short version: a training crawler turns your pages into knowledge without links — that’s exactly what happened to my 42 pages. The visits that can become citations are the other two: the search-index crawl and the live fetch. Those are the ones to watch for.
Position
Position is where you appear in an answer that names you, counted from the top. Position 1 means you’re the first brand named; position 5 means four others came before you. Since it counts places from the top — like finishing places in a race — a lower number is better.
It’s measured only across the answers that mention you, and it adds a nuance visibility can’t see: being consistently named and consistently last is its own finding. You’re in the conversation — as the footnote.
Sentiment
Sentiment is how the answers that name you talk about you — recommended, described neutrally, or warned against.
One consequence hides in that definition: sentiment only exists where mentions exist. If you were named in three answers out of sixty, your sentiment is built on three sentences — and a glowing verdict on three sentences is not a reputation, it’s a tiny sample. So a “positive” sentiment score means very little while your mentions are low, and it starts meaning something only once there are enough mentions to average. Mentions first. Sentiment second.
Part three: what the numbers say about everyone else
Source
A source is where the assistant got the answer that time — the pages it leaned on, sometimes shown as links, sometimes invisible. Sources answer a question most people never think to ask: not “does the AI know me?” but “who does the AI trust about my subject?”
Share of voice
Share of voice is how often each brand — you and your competitors — gets named across the same set of questions. It’s the “compared to whom?” that visibility alone doesn’t answer.
And here is the sentence to read twice before panicking at any competitor chart: a competitor being ahead of you means more has been published about them in the places assistants read. That’s all it means. Not a better product, not a bigger company, not a lost war — more text, in the right places. Which is bad news with a silver lining, because “publish more, where it counts” is an action, and “be a better company than them” is a mood.
Where does it count? Look at the sources again. List the domains assistants cite most in any field and most of them belong to nobody in the market — they’re the comparison sites, directories, forums and publications where the category gets discussed. Those aren’t competitors to worry about. They’re places to show up.
One example from my own logs, because it shows how much a source can tell you — and how the story ends. On 29 July I counted the visits AI crawlers had made to my knowledge base: 117, from six different companies. From OpenAI — the company behind ChatGPT — zero. Not “a few”: zero. Whatever ChatGPT was saying about my subject those days, it was saying without having read a single one of my pages.
Writing more pages would not have fixed that. A crawler that never comes doesn’t care how good the new page is — that’s a distribution problem, not a content one. What it needs is links from places that crawler already reads, so it finds the way in.
Which is exactly what happened next. The announcement went out, and the first links to those pages appeared on the internet. On 1 August — three days later — one of OpenAI’s crawlers finally showed up, and fetched two pages. Two, out of forty-two: a glance, not a read. And I can only see a 24-hour window of my own logs, so I can’t claim that was the very first time. But the way in had been found, and not one word of my pages had changed. Content problem or distribution problem: the sources tell you which one you have, and knowing that is worth more than any score.
What none of these numbers can tell you
The honest paragraph, because every measurement has edges:
- A percentage is not a ranking. Visibility 60% doesn’t mean you’re “top three in your market”; it means you were named in 60% of the answers to your chosen questions, that week.
- Small sets are anecdotes. Numbers built on a few questions rest on a handful of answers. Two questions asked six times each is twelve answers — enough to notice something, nowhere near enough to conclude anything. Treat those numbers as hints, not verdicts.
- The numbers describe the questions, not the market. A visibility score is not a fact about your market; it’s a fact about your list of questions. Two companies could measure the same market with two different lists and get two different winners — and both would be right.
- “Nothing left to fix” is not “you won.” When a tool has no more warnings for you, it means nothing is in the way — not that assistants now recommend you. Silence from the referee isn’t applause.
And none of it — none — can tell you whether your product is good. It measures what gets said, in a medium where what gets said is downstream of what got published.
Try it yourself, this week, for free
You don’t need a tool to start. You need an hour:
- Write ten questions your real customers would ask an assistant. Not your keywords — their sentences, in their language.
- Ask two assistants all ten, in a logged-out window. For each answer write down: were you named? were you linked? who was?
- Note who gets cited instead of you. Those domains are your actual to-do list — the directories, communities and publications where you should exist.
- Check that your pages can be read at all — that your site isn’t quietly blocking the crawlers of the assistants you care about. (Here’s how the crawlers work.)
- Come back in a month and ask the same ten questions. Same questions — that’s the whole trick. Change the questions and you’ve changed the ruler.
Do that twice and you will understand your own AI visibility better than most dashboards can tell you, because you’ll know exactly what stands behind every number: which questions, how many runs, named or linked.
And if you’d rather not spend that hour every month — this is the job Suparanku does. It asks your questions on a schedule, counts the mentions, the citations and the sources, and shows you what moved — and what to do about it. One honest limit, and it applies to every tool in this field, mine included: no tracker can make your work worth recommending. What it can do is make sure the work you put in lands where it counts — and that you find out what’s working from a number, not from a feeling.
My own numbers — the 31 out of 80, the 117 crawls, the zero that later stopped being one — live dated and public on one page that grows every measurement round. Round 1 lands on 28 August. If the numbers moved, I’ll say so. If they didn’t, I’ll say that instead.
Have an awesome journey,
Francesco