Citation counts are a bad explanation of AI visibility. If your AI visibility dashboard is a leaderboard of who got cited, you’re measuring who got footnoted, not who won the answer. Those are different things, and this week three separate pieces of evidence landed showing just how different.

An AI answer can make you the winner and hand the citation to someone else

At the Search ‘n Stuff conference, Tom Capper made a case that Mark Williams-Cook’s Core Updates newsletter passed along, and it deserves to be repeated everywhere: he showed an AI answer to a car-buying query that named Audi in the first sentence, then recommended the Audi S6, the Audi A6, and the Audi R8. The citation under that answer went to a car-parts retailer.

Run that through a citation-counting dashboard and Audi scores zero. Four brand mentions, three specific models, and an outright recommendation: the most Audi-favorable answer a language model could plausibly generate, and the metric says Audi is invisible. Meanwhile the parts retailer “wins,” having sold exactly zero sedans.

A metric that scores the winner at zero in their best-case scenario isn’t a rough proxy that needs refinement. It’s the wrong number.

People absorb answers they never click

The second piece of evidence is an eye-tracking study by MJ Cachón and Diego Criado at Laika. They put 22 people in front of a Tobii eye tracker, ran them through 9 searches each, and measured the gap between what people look at and what they click.

The gaps are enormous. The link cards inside AI Overviews were viewed by 81.8% of participants and clicked by 13.6%. That’s six people reading those links for every one who clicks. The “Explore Further” AI element was viewed by 83.3% of participants and clicked by nobody. Image and product blocks hit 100% viewership with 0% clicks in some intents.

Yes, it’s 22 people in a lab in Spain, so hold the specific percentages loosely. But the direction is not subtle. Users are reading. The study clocked people fixating on the main AI Overview text for an average of 919 milliseconds, longer than they spent on organic results. Then they act somewhere else, later, or not at all. The influence happens; the click record just never hears about it.

If a brand can be recommended without being cited, and read without being clicked, then a dashboard built on citations and clicks is auditing the receipts for a transaction that happened off the books.

Google just told you whose numbers to trust, and it isn’t your vendor’s

The third piece landed on July 10, when Google updated its official guide to AI features in Search with a new measurement section. Two things in it matter.

First, there’s now first-party data: the Generative AI performance report in Search Console shows how often links to your site appeared in AI Overviews and AI Mode, broken down by page, country, date, and device. It’s impressions only, it’s rolling out gradually, and it’s not going to win any awards for depth. But it comes from the only company that can actually see the system.

Second, Google added this: “No third-party tool has access to our internal ranking or AI systems.” That sentence is aimed directly at the vendors selling AI visibility scores with implied insider precision. Per Google’s AI optimization guide, use third-party tools if they help your workflow, and evaluate their advice against official guidance. A confidence score on a citation count is still a count of the wrong thing, now with decimals.

”But citations are the only thing we can measure”

This is the reasonable objection, and it’s half right. Citations and referral clicks are the observable exhaust of AI search, so tools count them. It’s the old joke about the drunk searching for his keys under the streetlight because that’s where the light is. The light is real. The keys are still in the grass.

Being measurable doesn’t make a number meaningful. Batting average was measurable for a hundred years; the Oakland A’s won twenty straight games the year they decided it was the wrong number. Your citation dashboard is batting average.

Measure mentions, impressions, and outcomes instead

Here’s what actually tells you whether AI search is working for you:

1. Mentions, not citations. Is your brand named in the answer text itself? That’s the Audi position, and it’s the one that sells sedans. Sample the prompts that mirror your real buying journeys: specific products, real customer types, genuine constraints. Then check whether you’re in the answer, not the footnotes.
2. Impressions from the source. The Search Console Generative AI performance report is thin, but it’s real. Watch it by page and country and treat it as your ground truth for “did we appear,” because it’s the only number that comes from inside the building.
3. Outcomes. Growth in branded search volume. Growth in direct traffic. Conversions from people who arrive already knowing what they want because an answer engine briefed them. These lag, and they’re entangled with everything else you do. Welcome to marketing. But they’re the only metrics that pay invoices.

Why the industry keeps selling citation counts anyway

The psychology here is straightforward. SEOs spent twenty years staring at rank trackers, so when AI search arrived, the industry rebuilt the most familiar piece of furniture: a leaderboard with your domain on it. Vendors need a single defensible number to put on the pricing page, and “citations” is countable, chartable, and sortable. Certainty sells, even when it’s certainty about the wrong thing.

But a number you can chart precisely is not the same as a number that explains anything. Citation counts fail the moment you test them against an answer where the mention and the citation split, which, as Capper showed, is not an edge case. It’s Tuesday.

So before you renew that AI visibility dashboard, run one test: find an answer where you’re named and someone else is cited, and see what the tool says happened. If it says you lost, you’re not measuring your visibility. You’re measuring the shadow and calling it the object.