What Do Marketing Teams Use to Measure How Often Their Brand Gets Recommended by AI Tools?

Marketing teams measure AI brand recommendation with a small set of concrete numbers, not a single score: mention rate, average position in the answer, share of voice against every company named, citation share, and downstream AI sessions tied to leads and revenue. Each one comes from running a fixed set of buyer questions through ChatGPT, Claude, Gemini, and Perplexity on a repeating schedule and parsing what comes back. The teams that get real answers hold the sampling conditions constant and report the metrics separately. The teams that get fooled buy a blended "AI visibility score" and never learn which input moved.

What teams actually run this on

Three setups show up in the wild, and they are not equally honest.

The first is a person with a spreadsheet who opens ChatGPT on the first Monday of the month, asks nine questions, and pastes the answers into a doc. Better than nothing. Not a measurement, for the reason the next section gets into.

The second is a homegrown script hitting the model APIs on a cron. It fixes the cadence problem and introduces a subtler one: a bare API call carries no persona, no memory, and no location, so it behaves differently from the consumer apps your buyers are actually typing into. You end up measuring a model, not a market.

The third is a dedicated monitoring product that runs the question set on a schedule across all four engines, detects every company named in every answer, and stores the raw text. That is the setup that produces numbers you can defend in a meeting.

The instrument matters less than the numbers it produces. Here is the working stack, in the order it earns its place.

Mention rate: the base number

Mention rate is the percentage of tracked prompt runs where an engine names your brand in the answer body. Named in 30 of 100 runs is 30 percent. That is it.

The discipline is in the sampling. A single ask is a sample, not a measurement, because these models are nondeterministic: run the same question twice and you can get two different shortlists. So mention rate only means something across repeated runs on a stable prompt set. On the paid plans Clark runs monitoring every day and rotates through your question pool, so each individual question comes back around roughly weekly and the trend is built from dozens of real answers instead of one screenshot.

You should not be staring at that chart daily. Set thresholds and let the number come find you. The three worth an email are a first appearance in AI answers, a 15-point week-over-week swing in mention rate, and a 2-place move in average position. Anything smaller is usually nondeterminism breathing.

This is the first number to instrument, and the only one that directly answers "how often do we get recommended."

Average position: are you first or fifth?

Position is your ordinal placement when the engine lists vendors. Being named first in a shortlist of five is a different commercial event from being named fifth. In practice the first two names are the shortlist and the rest are furniture.

Mention rate without position flatters you. A brand mentioned in 60 percent of answers, always last, is losing to a brand mentioned in 35 percent and always first. Track them side by side or you will misread your own progress as momentum.

Share of voice: the number your competitors set

Share of voice is your mentions as a proportion of all brand mentions across the prompt set. It is only valid if the tool detects every company named, not just yours. Otherwise the denominator is invented.

The useful output is a leaderboard of who keeps taking the ground you want. One practical wrinkle: the assistants themselves show up as "brands" in answers about AI tooling, which quietly distorts the chart, so being able to hide specific companies from the leaderboard matters more than it sounds. Hide a competitor and the change should carry everywhere the leaderboard feeds, including the companies-seen count.

This is the metric a Head of Growth puts on the board slide, because it is relative and it names names.

Citation share: separate from mentions, always

A mention is your name in the answer. A citation is your domain in the source list. They move independently. You can be recommended in an answer that cites only third-party roundups, and your own post can be cited as the source for an answer that recommends a competitor.

Plenty of brand mentions carry no citation to that brand's domain at all, because models name companies from parametric knowledge, listicles, review sites, and community threads. Report citation share separately or you will either undercount how often you are recommended or overcount how much your owned content is doing. Clark records the distinction explicitly: if your website appears only as a cited source link and your name never appears in the answer itself, that is logged as a citation, not a mention. Most AI-visibility tools over-report here, and the inflated number is always the flattering one.

Prompt set design: the method that decides whether any of it is real

The measurement method matters more than the tool.

• Bottom-of-funnel prompts. "Best X for Y" and head-to-head comparisons correlate with buying intent. Broad informational prompts inflate mention counts and predict nothing.
• Persona and location context. Memory, history, and inferred location shift which brands an assistant names. Prefixing each ask with a short buyer persona, plus a location line where geography matters, gets you closer to what a real prospect sees. Keep the persona from naming your own company, or you have biased the exact number you are trying to measure.
• Stored answers. Keep the full answer text, which model produced it, and the source links behind it. Those source lists double as a target list of the third-party pages the engines keep leaning on, which is often where a competitor's advantage actually lives. It is also the only way to answer "how do I know this is real?" without asking anyone to take your word for it.

AI sessions, leads, and revenue: the metric almost nobody connects

Visibility metrics stop one step short of the question your CFO asks. The connecting metric is sessions referred by each engine, tied to the conversion events they produced and the pages they landed on.

This is harder than it looks, because engines increasingly strip or obscure the referrer, so a genuine ChatGPT visit can land in analytics looking like direct traffic. That gap is why so much AI traffic hides in direct, and closing it takes explicit source tagging plus conservative handling of the sessions that cannot be resolved.

The loop closes when the visibility number and the revenue number describe the same question. When Clark writes a post against a question the engines never named you in, it reuses that question word for word, so the next monitoring run measures the change that post produced, and the sessions arriving on that post carry through to attribution. Gap, post, mention rate moves, session, lead, closed sale. That chain is the whole reason to measure any of this.

It is also the only metric on this list that defends a budget line.

The one to skip: the composite AI visibility score

We deliberately do not ship one. A blended score mixes mention rate, position, sentiment, and citations under a proprietary weighting nobody discloses, which makes it incomparable across vendors, unexplainable when it moves, and impossible to tie to a dollar. When it drops four points, it cannot tell you which question you lost or who took your place.

Concrete and traceable beats tidy. What share of answers mention you, where you land when they do, how many answers were analyzed, how many companies showed up, and what any of it earned. Everyone shows you rank. We show you revenue.

If you want a baseline before you build the stack, the free AI visibility check will show you which buyer questions currently return your competitors' names and not yours. It is one snapshot, on ChatGPT, on a handful of questions. That is enough to tell you whether you have a problem. Measuring it properly means running it on a schedule.