The most common reason GEO programs stall is that nobody can tell whether they are working. Classic SEO has rank trackers and Search Console. AI search has neither in a mature form, so teams either measure nothing or measure the wrong thing. This piece lays out a practical framework for measuring AI search visibility, and what we put in a monthly report. It pairs with the strategy in our generative engine optimization guide.
Why classic metrics fall short
The instinct is to open your analytics and look for a spike. The problem is that a large share of AI search value never shows up as a click. When ChatGPT answers a user with your information and your name attached, the user may be satisfied without visiting. That is a win with no session behind it. Measuring GEO purely by traffic will systematically undercount it and make good work look like nothing happened.
So we measure the thing that actually matters, which is whether you are present in the answers your buyers see. That takes four complementary signals.
Signal 1: Prompt sampling
You cannot yet pull a report of every AI answer that mentioned you, so you sample. Prompt sampling means building a fixed set of the questions your buyers actually ask, running them against the engines on a schedule, and recording what comes back.
- Build a prompt set. Twenty to fifty real buyer questions across the funnel, from broad category questions to specific comparisons. Keep it stable so results are comparable over time.
- Run it across engines. The same prompts against ChatGPT, Perplexity, Google's AI Overviews, and any engine your audience uses.
- Record the outcome. For each prompt, note whether you were cited, whether you were mentioned without a link, who else showed up, and the gist of the answer.
Run it on a fixed cadence, monthly is usually enough, so you are comparing like with like. The prompt set is the instrument. Change it constantly and you lose your baseline.
Signal 2: Citation share
Once you are sampling, citation share turns it into a number you can track. Citation share is the percentage of your prompt set where you appear as a cited source. If you are cited in twelve of forty prompts, that is a thirty percent citation share for that engine, that month.
The absolute number matters less than its direction and its context. Track it over time to see whether your work is landing. Track it against named competitors to see who owns the answer space in your category. A rising citation share, especially relative to competitors, is the clearest single indicator that GEO is working.
Signal 3: AI referral traffic
Not every win is clickless. When users do follow a citation, that traffic shows up in your analytics with telltale sources: referrals from ChatGPT, Perplexity, and other AI surfaces. Segment it out and watch it as its own channel.
- Build a segment in your analytics for known AI referrers so you can see the channel cleanly.
- Watch behavior, not just volume. AI-referred visitors often arrive with high intent because the answer pre-qualified them.
- Expect it to undercount. Because so much value is clickless, referral traffic is a floor on your impact, not the whole of it.
Read this signal alongside citation share. Citation share tells you about presence in answers. Referral traffic tells you about the slice of that presence that converts into a visit.
Signal 4: Brand mention tracking
Citations with links are the headline, but engines also mention brands without linking, and those mentions shape perception. Track how the engines describe you when asked. Do they name you in your category? Do they get your positioning right? Do they recommend you for the use cases you want to own?
This is qualitative, and it is worth the effort. The story the engines tell about your brand is increasingly the first story a buyer hears. Catching a wrong or outdated description early lets you fix the underlying sources feeding it, which ties back to the entity and authority work in our guide on getting cited.
Setting a baseline before you start
Measurement only means something against a starting point. Before you change any content, run your prompt set once and record the state of things. That first pass is your baseline: your citation share on each engine, which competitors own the answers, and how the engines currently describe your brand. Everything after is measured against it.
A baseline also protects you from a common trap, which is attributing every movement to your own work. AI answers shift for reasons outside your control, including the engines changing their own retrieval and ranking behavior. When you have a stable prompt set and a baseline, you can at least separate a broad platform shift, which moves everyone at once, from a gain you actually earned, which moves you relative to competitors. Without a baseline, every reading is just a number with no reference.
On cadence, monthly suits most businesses. It is frequent enough to catch trends and course-correct, and spaced enough that you are measuring real change rather than day-to-day noise in how the engines phrase things. Faster-moving categories may warrant a tighter loop, but resist the urge to check daily and react to every wobble.
What a monthly report looks like
Pulling it together, a useful monthly AI visibility report has five parts. This is close to what we hand clients.
- Citation share by engine, this month versus last, so the trend is visible at a glance.
- Competitive share, your citation share against two or three named competitors on the same prompt set.
- Prompt-level detail, the questions where you gained or lost presence, with the actual answers quoted.
- AI referral traffic, volume and behavior from AI sources, segmented from the rest.
- Narrative and next actions, what changed, why, and the specific content or authority moves planned in response.
The last part is what makes it a report rather than a dashboard. Numbers without a decision are decoration.
An honest caveat on the tooling
This measurement space is young. The engines do not expose citation data through clean APIs the way search consoles do, sampling has real limits, and third-party tools vary in quality. We would rather be honest about that than pretend there is a precise, complete number for your AI visibility. There is not, yet. What there is, done consistently, is a reliable directional read: are you present in the answers that matter, is that presence growing, and are you winning it against competitors.
It is also worth being clear-eyed about who does the sampling. Manual sampling, where a person runs the prompts and records the answers, is slow but gives you the real, current behavior and the actual wording the engines use. Automated tools promise scale, but they vary widely in quality and can drift from what a real user sees. Our stance is to treat automated tools as a way to widen the prompt set, not as a substitute for periodically checking the answers with your own eyes. The engines are the ground truth, and the ground truth is what your buyers actually see.
That directional read is enough to run a real program on. For how the strategy behind these numbers fits together, return to the GEO guide, or see GEO vs SEO for why your old dashboards cannot answer these questions on their own.