AI visibility metrics, explained
Five metrics, what each one measures and — the part nobody explains — how they get misread.
Share of Answer (SOA)
The share of your sector's questions whose answer mentions your brand.
It is the headline metric and the easiest to misread. A 10% SOA over 90 questions does not mean "one in ten customers sees me": it means that, across the question set being measured, your brand appeared in nine of them. That is why it is always published with its interval and the number of observations behind it.
How it gets misread: by comparing the SOA of two different sectors. 5% in a market with four competitors is not the same as 5% in one with forty. The useful comparison is against yourself over time, and against whoever competes for your same questions.
LLM Visibility Index (LVI)
A 0-100 index based on your brand's position inside the answer: 100 means being named first.
Showing up is not the same as showing up first. When a model lists five options, the first takes the attention and the fifth takes almost none. SOA cannot tell those two apart; LVI can.
It is mainly useful for spotting one pattern: SOA rising while LVI does not. That means you appear more often but always at the end — you are joining the conversation, not winning it.
Share of Mentions (SoM)
Your share of all the mentions in your niche: out of every time AI names someone in your market, how often it names you.
SOA tells you how you are doing; SoM tells you how you are doing relative to everyone else. It makes something uncomfortable visible: a model's attention is zero-sum. If a competitor climbs, somebody drops.
Accuracy Score
Of the times AI states a fact about you, what share it gets right.
Every metric above measures how often you appear. None measures whether what is said about you is true, and those are two different things: a highly visible brand quoting a price that no longer exists is not building anything — it is reaching a lot of people with wrong information. And unlike a bad review, nobody tells you.
Each claim scores 2 if correct, 1 if half right and 0 if it contradicts a fact you have verified. What is not scored is silence: a model not mentioning your founding year is not a miss on its part, it simply was not relevant. Counting that as an error would turn the metric into noise.
How it gets misread: looking at the percentage without the number of claims. 100 % out of three sentences is not 100 %, it is an anecdote — which is why the figure always travels with its n and its interval.
Sources behind sentiment
Which domains AI cites in the answers where it speaks worst about your brand.
Knowing that sentiment is negative is useless on its own: you cannot go into ChatGPT and fix it. What you can do is act on the page feeding it. This crosses, inside each already-measured answer, the tone with the sources cited.
How it gets misread: as an accusation. This is correlation, not causation — a domain may show up in your negative answers simply because it shows up in all of them. That is why the number that counts is not the raw percentage but how far it deviates from your own average, and why a source with very few appearances is not listed: it would be a coincidence dressed up as a signal.
Grounding Source Prominence (GSP)
The domains AI cites when answering about your sector, how often each appears and in which questions.
It is the most actionable metric of all, because it is not about you: it is about where you need to be. If the same four domains are cited in half the answers in your market, that is your to-do list — and it rarely matches the list of places your industry assumes matter.
The engine where this metric is clearest is Perplexity, because it cites sources in every answer: how to appear in Perplexity.
The Library Effect
The gap between being known and being recommended.
A model can describe you perfectly when asked about your name and never propose you when asked for an open recommendation. That is the Library Effect: you are in the catalogue, but not in the answer.
It is measured by sorting every answer into four states, not two:
| State | What it means |
|---|---|
| Absent | You do not appear. |
| Mentioned | You appear, but as a fact: you are not proposed. |
| Hedged | You are proposed with a caveat or a condition attached. |
| Selected | You are recommended without reservations. |
The distinction matters because the work to go from absent to mentioned looks nothing like the work to go from mentioned to selected. The first is about presence; the second is about reputation.
The honest zero
A 0% that does not say how many questions it was measured over means nothing.
There is a simple statistical rule for this: with zero appearances in n attempts, the true rate could still be as high as roughly 3/n. With three questions that ceiling is 100% — you literally know nothing. With ten, it is 30%. You need around thirty before you can say with a straight face that somebody is invisible.
So when the result is zero, it comes with the number of attempts and the real ceiling of that claim, instead of a headline saying "you never show up".
Frequently asked questions
Which metric matters most?
Share of Answer to know where you stand, and Grounding Source Prominence to know what to do. The first describes the problem; the second is the only one that points at the work.
What does it mean if SOA rises but LVI does not?
That you appear more often but always at the end of the answer. You are joining the conversation, not winning it.
What is the Library Effect?
The gap between a model knowing you and recommending you. It can describe you well when asked about your name and never propose you when asked for an open recommendation: you are in the catalogue, but not in the answer.
Why does a 0% need context?
Because with zero appearances in n attempts the true rate can still be around 3/n. With three questions that ceiling is 100%, so the zero says nothing. You need roughly thirty observations before claiming somebody is invisible.
See your own metrics
The free audit computes your presence and who shows up instead of you. No signup, no card.
Run my free audit →Home · Plans & pricing · Methodology · Research · Perplexity · How to choose · Use cases · Changelog · Agencies · Alternatives · Diagnosis · Check if AI cites you · Noise and margin of error · Four-step method · Optimizing for agents · X · Product Hunt