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 questions left are whether you are in it, how often, how you are framed, and where you sit.

This guide defines the numbers that answer those questions, names the two traps that make them easy to measure wrong, and shows how to turn them into a weekly loop. The work is AI search optimization, also called GEO or AEO (answer engine optimization); the numbers stay the same.

#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 page 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. 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. A better rank tracker cannot see inside that paragraph; the answer needs numbers of its own.

#The five numbers that track AI visibility

Read in order, they go from "are you there at all" to "did it bring anyone".

Mention rate is the floor. For a set of questions 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 the text; a citation links your page as the source behind a claim, which shows the engine used you as a reference.

AI share of voice makes it relative. The answer has room for a few names, so the right comparison is the fraction of answers in your category that name you against each competitor.

Position in the answer captures framing. Being put forward first, as the recommended option, is worth more than a mention at the end of a list. It is the closest thing an answer has to a rank, and it is as much about wording as about order. AirOps, analyzing answers on ChatGPT, Claude and Perplexity, reports: "When brands are cited, 80% appear in the top three positions" (AirOps).

AI referral traffic is the proof, and the hardest to see. Attribution from inside an answer is still incomplete, so weigh it as confirmation rather than as your main dial. It is getting easier: Google's Search Console now reports how your pages perform in AI Overviews and AI Mode (Google), and AI visibility tools such as Peec AI, Profound and Otterly track the visits that come from AI assistants.

#The two traps: per engine, per question

The metrics are simple. Measuring them honestly is where most efforts go wrong, in two 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 the gap you need to see. Keep the numbers per engine, then aggregate.

Second, measure per question set. "Are we visible in AI" cannot be measured. "How often is our brand named when a buyer asks for the best tool in our category, across these forty questions, on each engine" can. The questions have to stay stable, so this week compares with last week, and they have to be the questions your buyers ask, so you track the ones that bring revenue rather than the ones you happen to win.

Get either wrong and the number misleads. A flattering average across engines, or a question set quietly tuned to your strengths, produces a figure that rises while your real visibility does not.

#Read it per question

It is tempting to roll the five numbers into one score. A single figure gets attention, and it also hides what you need to act on. A drop from 62 to 54 tells you something moved. "Perplexity stopped citing us on three buying questions after a competitor published a comparison" tells you what to do next.

So keep the detail as the main view. For each question and each engine, the useful readout is short: are you named, where, with which recommendation, is your site cited, and which sites are cited instead. The questions where a competitor is named and you are not are your list of work.

#Turn the numbers into a loop

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

Run it as a loop: measure the question set on each engine, see what moved, change one thing, measure again. A weekly cadence on a stable set catches real movement and shows when a competitor takes your place, without chasing daily noise.

#Put the loop to work

Obi, an AI Search Employee, runs this loop for founders and GTM teams. It starts from your ideal customer profile and picks the buying questions those customers ask. On a paid pack, it reads the answers on ChatGPT, Gemini and Google AI Overviews every week, with Claude and Perplexity as options; the free plan reads ChatGPT and Gemini once a month. For each answer it scores, it records whether you are named, where, how you are recommended, whether your site is cited, and which domains are. Then it proposes strategies for those gaps and drafts the pages with you in your AI assistant, for you to review and publish.