What a Good AI Visibility Metric Actually Looks Like

In the last issue I argued that you can measure AI visibility, but you can't measure much else yet. That still stands. The more useful question now is what to do with the number you can measure, because on its own it tells you surprisingly little.

A good example landed the following week.

A headline number and its denominator

On 29 October, Microsoft Clarity published an analysis of more than 5.2 million website sessions involving AI agents, across 230,511 Clarity projects. The headline finding was that checkout was the most common thing agents did, appearing in 25.21% of agentic sessions against 14.65% for humans.

PPC Land then did the arithmetic. The percentages appear to be calculated on a base of around 1.22 million sessions, not the 5.2 million in the headline. Against the full figure, checkout drops to roughly 5.9%. Microsoft also says it can't tell which clicks the agent made and which the person made.

I'm not raising this to knock Clarity. It's a free tool doing genuinely useful work, and its new Agent Activity Dashboard is worth switching on if you have agentic traffic. The point is that a respected platform published a striking number, and the real story was in the denominator. If that can happen with Microsoft's own data, it can happen with yours.

Being visible is not the same as being well represented

Gianluca Fiorelli put it bluntly on LinkedIn recently: raw visibility is "a simple and extremely stupid metric". Strong words, but the logic holds. A citation count tells you that you appeared. It doesn't tell you whether appearing did you any good.

You can be cited with an out-of-date price. You can be cited for the wrong page, sending a buyer-ready query to a blog post from three years ago. You can be mentioned alongside three competitors in a sentence that makes you sound like the budget option. All of those count as visibility. None of them are wins.

Seven questions to ask of any AI visibility figure

The first six are adapted from Gianluca's list. The seventh comes from the Clarity example.

1. Are the facts correct? Pricing, services, locations, who you work with. Outdated facts cited confidently are worse than no citation.
2. Is it the right URL for the intent? A commercial query should land on a service page, not an old article.
3. What's the narrative? Positive, neutral, negative, or framed in a way you'd never choose yourself.
4. Who are you cited alongside? Co-citation shapes how you're positioned, whether you like the company or not.
5. Where does the link sit? Inline in the answer, or tucked into a footnote that few people open.
6. Are you there across the whole conversation? AI search is a conversation. Appearing in the first answer and vanishing on the follow-up is a different result from staying present throughout.
7. What's the base? Any percentage needs a denominator you can see. If a tool won't tell you what it's counting, treat the number as directional, not definitive.

What the tools can tell you, and what they can't

The tools are improving. On 13 October, Bing added a Topics Portfolio beta to the AI Performance report in Bing Webmaster Tools. It groups the queries your site is cited for into topics and, usefully, lists opportunity queries where other sites are cited and yours isn't. It's free, and it's the closest thing we have to a competitive gap view for AI answers.

It's also citation data only, with no clicks. Bing describes the figures as approximate, and the tab only appears if there's sufficient data for your site. Clarity's dashboard shows what agents do once they arrive, but not who decided what. GA4's AI channel shows referral traffic, which misses everything that happened inside the answer itself.

Each of these answers one question well. None of them answers all seven. That's not a flaw to wait out. It's the shape of the job for now, and it's why the reporting stack matters more than any single tool.

Meaning is still a human job

There's a related point that applies to the data you feed AI systems, not just the data you get back. Veronika Heimsbakk wrote recently about AI-generated ontologies. Her argument is that a generated ontology can look complete while skipping the part that matters: people agreeing on what things actually mean. She gives the example of a product catalogue where a reasoner treated every t-shirt variant as the same item. Her recommendation is a human-authored core, with AI used for drafts.

That applies directly to schema, product data and entity markup, which is a large part of what AI systems draw on when they describe you. As she puts it, "You can only refine what you understand." It's the same argument I made in Learn by Playing, Lead by Knowing: the tool speeds up the work, but the judgement comes from having done it.

What to actually do this quarter

You don't need a new platform to do this properly. You need a short list and a routine.

• Pick 10 to 20 prompts that matter commercially. The questions a buyer would actually ask before contacting you, not vanity terms.
• Run them monthly and check the answers against the seven questions. Note what changed since last time.
• Put that alongside what you can count. Bing citations and opportunity queries, GA4 AI referrals, Clarity's agent activity. Benchmark the trend, not the single reading.
• Report context, not just counts. "Cited 40 times" is less useful than "cited 40 times, correct on pricing, right URL in most of them, and losing the comparison queries to two competitors."

That last example is the difference between a visibility number and a visibility metric. One tells you that you showed up. The other tells you what to fix.

If you'd like help setting up a routine like this, or working out what your current AI visibility numbers are actually telling you, that's the kind of work we do.