A wave of 2026 reports now ranks brands by a single AI-citation score. But a single number averages over engines that barely agree and answers that do not hold still.
2026 has been the year of the AI-citation leaderboard. Report after report now ranks brands, publishers and domains by a single visibility or citation score, built from hundreds of millions of tracked AI answers. The ambition is good: AI search is where attention is moving, and people want a number to manage. But the number is the problem. When we measured what actually sits underneath one of these scores, using our own data, we found it is an average of things that do not average well: engines that barely agree with each other, and answers that do not even agree with themselves.
Through 2026 a steady stream of AI-citation reports has arrived, each ranking brands and domains by an aggregate citation or visibility score drawn from hundreds of millions of AI answers. One widely cited example, 5WPR’s State of AI Citations 2026, reports that the top 20 sources captured roughly two thirds of one engine’s citations, a useful read on how concentrated AI’s sourcing is. The instinct to turn all this into a single score is understandable. The trouble is what the score has to flatten to exist.
When we asked the main AI engines the same questions and traced every citation, only about 8% of the domains they cited were shared across engines. The other 91% were unique to a single engine. That is the finding in AI engines cite different sources, and it breaks the idea of one score cleanly: a brand can be cited constantly by one engine and never by another, and a single cross-engine number reports neither. It reports the average, which is true of no engine in particular.
Worse for a fixed score, the engines do not even agree with themselves. When we asked the identical question again, about half of the sources an engine had cited were gone, replaced by others. That is our volatility study: being cited once has roughly a coin-flip chance of repeating on the very next identical query. A score computed from one snapshot is reporting a moment, not a position.
None of this means citations do not matter. It means the unit is wrong. The facts worth managing are which engine named you, for which question, and how stable that is across repeated checks. Keep the actual answer as evidence, read it per engine, and watch it over time. That is the whole of our argument for AI search: prove it, do not score it. You can see the shape of it on your own domain with the free AI visibility check, which shows you which engines cite you right now with the answer attached, then lets you watch it move.
A single number, usually 0 to 100, that aggregates how often AI engines cite a brand or domain across many questions and engines. The 2026 reports use it to rank who is "winning" AI search. Our argument is not that citations do not matter, it is that collapsing them into one cross-engine number hides the facts that matter.
Because it averages over two kinds of disagreement. Across engines, only about 8% of the domains cited overlap, so a brand can be cited heavily on one engine and invisible on another. Within a single engine, about half of the cited sources change when you ask the identical question again. One score cannot carry either fact.
Three things a single score throws away: which engine named you, for which question, and how stable that is over repeated checks. That means per-engine results, the actual answer kept as evidence, and repeated measurement over time rather than a one-off snapshot.
No. They are useful for seeing the shape of AI search, like how concentrated citations are in a few large sources. The caution is specific: do not manage your own visibility from a single ranked score, because the underlying signal is engine-specific and volatile.
The 8% cross-engine overlap that a single score averages away.
Read it →The within-engine instability a one-off score cannot see.
Read it →Per engine, with the answer attached, tracked over time, not one number.
Read it →Unsourced captures which AI assistants cite you, proves which crawlers really fetched your pages, and re-checks after you act — evidence, not a score.
© Unsourced, the evidence layer for AI search. Our data and findings are free to quote and cite — please attribute to Unsourced and link to unsourced.app.