What is an AI visibility index?

We publish one, so read this knowing that. What follows is what running an index has actually taught us, including the parts that were expensive to learn. The distinction below is not marketing: it changes what your number is capable of telling you.

What makes a measurement an index rather than a dashboard?

Four conditions, and all four have to hold.

A fixed field. Every brand competing in the category is measured, not just the one paying. Without that there is no denominator, and without a denominator there is no share of anything.

One exam. Every brand answers the identical questions, sampled the same number of times, judged by the same logic. Two brands measured on different questions cannot be compared, however similar the questions look.

Published results. The ranking is standing and citable. A number nobody else can see is not a position, it is an opinion you paid for.

A versioned method. The question set and the judging logic carry version numbers, so when they change everyone can see that they changed.

Drop any one and you still have something useful. You just do not have an index.

Why can't you build a ranking from your own numbers?

Because a rate is not a position. If an assistant names you in 11% of relevant answers, that is genuinely useful — but it tells you nothing about whether 11% is excellent or dismal in your category. It might be the best figure in the field. It might be last.

Share of voice is the metric that answers that, and it is a fraction: your mentions over everyone's mentions. The denominator has to be measured, not estimated. That is the whole cost of running an index, and it is why most products do not.

What is the hardest part of running one?

Deciding who is in the field. It sounds like admin. It is the measurement.

When we audited our own cigar-retail cohort against what the assistants actually said, the brands on our list accounted for fewer than half of all the brand mentions in the answers. Every rival we had missed was inflating our subject's share of voice, because the denominator was too small.

The failure modes are duller than you would hope. One retailer appeared under three different names across the answers — the shop, the street it trades on, and its formal company name — and counted as three separate competitors, each with a third of its real weight, until the aliases were resolved. Another was simply absent from the list because nobody had thought of it.

An index has to keep finding these. A dashboard never has to look.

What happens when a category is not really one field?

We ran a category that turned out to be two. It contained cigar retailers and whisky retailers, and while both are specialty retail, no buyer question spanned both — nobody asks an assistant for a shop that sells cigars and single malt.

The result was a share-of-voice figure divided by a denominator half of which was irrelevant to the question being asked. Splitting the category and re-judging the same stored answers, with no new measurement at all, moved brands' standing materially. Nothing about the brands had changed. The field definition had been wrong.

This is the failure a dashboard is immune to, because it never claims to know what the field is.

Why does an index have to version its own method?

Because otherwise it cannot tell you what a change means.

If the question set changes and your score moves, the honest reading is that the exam moved, not your brand. We refuse to draw a trend line between two readings unless the question set, the judging logic and the set of assistants all match — and where they do not, we say the comparison is void rather than plotting it.

That rule costs us. It means a brand measured before we added a fourth assistant cannot be compared with one measured after, and it means publishing a new version restarts the trend for everyone in the category. An index that quietly re-scores history looks smoother and tells you less.

What must an index never do?

Publish a number it did not measure.

We know because we did it. Early on, seeded demonstration data — invented scores for real, named companies — reached the public leaderboard and was ranked alongside genuine readings. It was labelled in one place and nowhere else. Two of the businesses ranked above a real, honestly-measured brand on the strength of numbers that were fiction.

The rule that replaced it is absolute: a simulation may be displayed where it is labelled as one, and can never feed a published number. The immediate effect was that our leaderboards got much thinner, because most of what was on them had not really been measured. Thin and true beats full and false, and the only remedy for thinness is to go and measure.

Does the measurement even agree with itself?

Worth knowing before you trust any single figure: assistants are non-deterministic, so a rate has a range around it. One of our recent readings put a brand's mention rate at 10.7% — with the true value somewhere between 7.5% and 15.1%, and that was across 252 sampled answers. The interval is wide because the underlying thing is noisy, not because the sample was small.

Anyone reporting a bare percentage is reporting the midpoint of a range they have not shown you. Month-to-month movement inside that range is dice, and it will be sold to you as progress.

So do you need an index or a dashboard?

Honestly, most operational work is better served by a dashboard. If you want to know what changed this week, get alerts, and watch dozens of prompts across markets, buy a monitor — we are the wrong shape for that, and we say so in our guide to choosing.

An index earns its place when you need a number that survives contact with other people: a board paper, an investor deck, a claim on your own website, a comparison with a named competitor you would have to defend. That number has to come from outside you, cover your rivals as well as you, and rest on a method anyone can read.

You can see what that looks like in practice — the index itself, the method behind it, and what we mean by the term.