Measuring AI Search. But Not What Counts.

Visibility in AI Search has become the obsession of the moment. Every brand wants to appear in ChatGPT responses, Google AI Overviews, or Perplexity answers. The question we get most from clients: "How do we make sure AI mentions us?"

The problem is not that they are invisible. The problem is they do not know whether they are visible, because they are measuring the wrong things. Dashboards full of numbers that look impressive but say nothing about what matters: whether the brand is recommended, in what context, and with what effect.

A recent article on Search Engine Journal puts its finger on the same wound. Tom Capper from STAT says it plainly: "Citations have URLs. That's still valuable tactical data, but I don't think it should be your KPI." He is right.

And this is not just a tooling problem. It is a mindset problem. Many teams apply classic SEO logic (rankings, clicks, impressions) to a completely different medium. AI Search does not work on ordered results. It works on synthesis, context, and recommendation. The old metrics do not fit, but nobody has replaced them yet.

Mentioned does not mean recommended

The first confusion we see with clients: they treat any appearance in AI responses as a win. But there are three very different levels of visibility.

A mention means your brand appears in the text. The AI knows you exist. It is a signal, not a victory. A citation means the AI links to your site. More concrete, but still does not tell you whether you are recommended or just listed. A recommendation is what actually matters: the AI says "use X" or "X is a strong option for Y." That drives business.

We have seen brands appearing in 80% of relevant responses, but always in third position, inside a list without context. Visibility? Technically, yes. Impact? Almost zero. We have written about this before: AI visibility depends enormously on how you ask and whom you ask.

Three metrics that actually matter

If you drop "how many citations did we get this month" as your primary metric, what replaces it? Three things.

Presence: does your brand appear in the AI response for a given prompt? Not how many links you have, but whether you are there. Simple, binary, but much harder to track than it sounds. Responses vary by location, device, and context. You need samples, not isolated screenshots.

Recommendation context: when you appear, how do you appear? Are you recommended as a solution, mentioned in passing, or associated with a problem? The difference is enormous. A brand can be "visible" but mentioned exclusively in negative contexts. Without this level of detail, your dashboard is lying.

Visual prominence: where do you appear on the page? In AI Overviews, above the fold matters. Position one in a carousel that requires scrolling is not the same as position one that is immediately visible. Visual prominence complements classic ranking and works as a leading indicator.

A concrete example from our practice: an e-commerce client appeared in AI Overviews for "best laptop for students" but always below the fold, in the third carousel. Technically visible. Practically invisible. When we re-optimized the cited source content and added clearer structured data, the visual position improved. The classic ranking did not change at all.

The prompts you track are probably wrong

Many teams take their SEO keyword list and paste it into an AI tracking tool. The approach is fundamentally flawed.

An SEO keyword and an AI prompt are different things. Nobody types "digital marketing agency Bucharest" into ChatGPT. They write "I want a marketing agency to help with Google Ads, budget around EUR 2,000 per month, in Romania." Context, intent, budget, location, all of it enters the prompt.

The correct approach: start from real questions. People Also Ask on Google is a solid starting point. Then think about the people behind those questions: what roles they hold, what budgets they have, what industries they are in. Build a prompt sample that reflects the real diversity of your audience.

Treat response variability as a sampling problem, exactly how rank tracking handles local and device-based variations. In a previous article we showed that AI works as a matchmaker, not as a search engine. Your tracking prompts need to reflect that.

In practice, we use sets of 30 to 50 prompts per client, grouped by intent (informational, comparative, transactional) and by persona. We run them periodically across multiple AI engines and compare results. We do not look for absolute numbers but for trends. If in a comparative prompt your brand appeared last month and now does not, that is a signal. If it appears but in a negative context, that is an even more important signal.

Citations are evidence, not objectives

The most counterintuitive thing in this whole discussion: citations are not useless, but their role is investigation, not reporting.

When a competitor is recommended and you are not, look at the pages cited. Are they review sites? Industry publications? Directories? That is where the action is: improve your brand presence on those sources and monitor whether that changes AI recommendations too.

The process looks like this: identify cited sources, analyze what content they have about you versus the competition, act on sources where you can influence, then measure the impact. Not "we had 47 citations this month" but "we identified 3 sources where the competitor appeared and we did not, we fixed that, and recommendations shifted in 60% of our sample prompts."

Search Console already sees AI traffic, but it does not see mentions without a click. That is why the combination of tracking data, post-conversion surveys, and source analysis becomes essential.

Four steps toward realistic measurement

For teams that want to start correctly, not complicated:

1. Audit what is missing. Open your current dashboard and look for what is not there: mentions without a link, recommendation context, visual prominence. If you only see citations and traffic, you have gaps.

2. Define consistent comparisons. Before tracking trends over time, document the entities, competitors, prompt types, and collection environment. Without a stable baseline, any trend is noise.

3. Connect source research to actions. Every source you investigate or influence needs to be logged. Otherwise you cannot link a change in recommendations to a specific action.

4. Tie visibility to business outcomes. Compare visibility trends with revenue, conversions, and survey responses. AI visibility is a leading indicator, not a final objective.

Most of our clients started exactly here: an audit of dashboard gaps, a prompt set built around the real audience, and a methodology for investigating sources. Not spectacular, but it works.

AI Search is not optional and no longer new. But measuring it is still in its early stages. Those who start correctly now gain time, not just data.