Measuring AI search visibility

If you came from SEO, your instinct is to open a rank tracker and check your position. In AI search that instinct fails on contact, because there is no position to check. The engine read the sources for you and returned one answer. The only questions left are whether you are in it, how often, how you are framed, and where you sit.

This guide defines the metrics that answer those questions, names the traps that make them easy to measure wrong, and shows how to turn them into a loop. No borrowed numbers: where a figure would matter, we point to our own ongoing measurement rather than cite one we cannot stand behind.

Why the old metric does not transfer

Ranking position measured a list. A person scanned ten links, and being third instead of eighth changed whether they clicked. The metric existed because the surface was a list and attention dropped down it.

An AI assistant removes the list. It composes one answer and names a few sources inside it. There is no eighth place to climb out of, so position on a results page stops meaning anything. The unit of visibility moved from a rank on a page to a presence in a paragraph, and the metrics have to move with it.

That is not a tooling gap you can patch with a better rank tracker. It is a different surface, and it needs metrics built for the answer, not for the list the answer replaced.

The five metrics that track AI visibility

Five numbers describe whether an engine names, trusts and recommends you. Read in order, they go from "are you there at all" to "did it produce anything".

Mention rate is the floor. For a defined set of prompts your buyers actually ask, how often does the engine say your name? If it is zero, nothing below matters yet.

Citation rate is the next rung. A mention is your name in prose; a citation links you as the source behind a claim. Citations are stronger, harder to fake, and more durable, because they reflect the model trusting you as a reference rather than recalling you in passing.

AI share of voice makes it relative. Visibility is a competition for a finite answer, so the right comparison is not "are we mentioned" but "what fraction of the answers in our category name us versus a competitor". You can improve in absolute terms and still lose share if a rival improves faster.

Position in the answer captures framing. Being named first, as the recommended option, is worth more than being listed fifth as an also-ran. The reading order the model composed is the closest thing the answer layer has to a rank, and it is qualitative as much as ordinal.

AI referral traffic is the proof. The visits that arrive from an assistant's answer are the downstream signal that presence became action. It is the smallest and noisiest of the five today, because attribution from inside an answer is still maturing, so weigh it as confirmation rather than as your primary dial.

The two traps: per engine, per prompt

The metrics are simple. Measuring them honestly is where most efforts go wrong, in two specific ways.

First, measure per engine. Presence on ChatGPT tells you nothing about Perplexity, Claude or Gemini. Each pulls from different sources and composes differently, so a brand strong on one can be absent on another. An average across engines hides exactly the gap you need to see. Keep the score per engine, then aggregate, never the reverse.

Second, measure per prompt set. "Are we visible in AI" is not a measurable question. "How often is our brand named when someone asks for the best tool in our category, across these forty prompts, on each engine" is. The prompts have to be stable so movement is comparable week to week, and representative so you are tracking the buying questions that matter rather than vanity queries you happen to win.

Get either wrong and the number lies. A flattering average across engines, or a prompt set quietly tuned to your strengths, produces a score that goes up while your real visibility does not.

Rolling it into one score

Five metrics across many engines and prompts is the right level of detail for the work and the wrong level for the owner who needs to know, in one glance, whether things are improving. So they roll into a single GEO Score from 0 to 100, and Surface reads all five across ChatGPT, Perplexity, Claude, Gemini and Google AI Overviews and acts on the gaps behind the number.

The score is the headline. Its value is that you can drill from it: a drop is not "we went from 62 to 54", it is "Perplexity stopped citing us on three buying prompts because a competitor published a comparison". The number gets attention; the breakdown drives the decision. Kept that way, the score is a steering instrument, not a vanity metric.

Turn the numbers into a loop

A measurement you take once is a report. A measurement you take on a cadence is a system. AI answers shift as engines recrawl, as competitors publish, and as the models update, so a single snapshot ages within weeks.

So run it as a loop: measure the prompt set per engine, see what moved, change one thing, and measure again. A weekly cadence on a stable prompt set is enough to catch real movement and react when a competitor takes your slot, without chasing daily noise. That continuous loop, run across every engine that matters, is the whole discipline.

Surface Agent runs that loop for you. It watches your visibility across ChatGPT, Perplexity, Claude, Gemini and Google AI Overviews around the clock, along with the Reddit threads where your buyers ask for recommendations. It drafts the content and comparison pages that move the metrics, and drafts the Reddit replies in your voice for you to post from your own account. You review and approve, and the score climbs.

Start your 3-day free trial