Open three digital PR reports from three different agencies this month and you will find three different AI citation metrics, all called roughly the same thing. Nobody is being dishonest. The industry added a KPI before it agreed what the KPI was, and the bill for that is coming due in renewal conversations.
I have watched this play out twice now in the same quarter: a client sees the AI number fall, asks what happened, and the agency cannot answer because the metric they chose moves for reasons nobody in the room controls. That is not a reporting problem. It is a definition problem.
The metric everyone added and nobody defined
The proposals in circulation are all reasonable in isolation. Citation share by LLM — how often your brand appears in responses from each engine. Citation share by outlet — which of the publications you earned coverage in are actually being used as sources. Share of voice against competitors in AI answers. Accuracy and sentiment of how the model describes you. Earned-coverage-to-citation lift.
Five candidate metrics, and the industry picked whichever one its existing tooling already produced. Agencies with rank-tracking heritage report share of voice, because it looks like share of search. Agencies with media-monitoring heritage report by outlet, because that is what their clip reports already track. The metric got chosen by procurement history, not by what it explains.
Three units that disagree with each other
| UNIT | WHAT IT MEASURES | MOVES WHEN… | ATTRIBUTABLE TO PR? |
|---|---|---|---|
| Citation share by engine | How often you appear in a given engine's answers | The engine changes retrieval, which happens constantly | Partly — confounded by model updates |
| Citation share by outlet | Which earned placements get used as sources | You earn coverage in a source engines trust | Yes — this is the PR causal chain |
| Share of voice | Your citations as a proportion of the category's | A competitor does anything, or the category grows | Barely — half the inputs are not yours |
Look at the third column. Share of voice can fall in a month where your PR team did outstanding work, simply because two competitors published research that got picked up. Citation share by engine can collapse overnight because an engine adjusted how often it links out at all — which is precisely what happened to referral patterns across the board this spring. Neither of those is a performance signal, and both will be read as one.
“If your KPI moves when your competitor ships a press release, it is measuring the category. It is not measuring you.”
Why share of voice is the worst of the three
It is also the one clients ask for by name, because it is legible. Everybody understands a pie chart. That legibility is exactly the trap: a metric that is easy to present and impossible to attribute is the worst combination available in a retainer.
There is a structural problem underneath it too. AI citations concentrate heavily on a small set of domains, so category-level share is dominated by whether a handful of high-authority publishers happened to cover your space that month. Your share can halve because one publisher ran a roundup that named four competitors. We have written before about how AI citations concentrate on a few domains, and that concentration is what makes share of voice so volatile at the account level.
What the citation data actually rewards
The useful thing about the research that has landed this year is that it points at levers, not just outcomes. A Digital Applied study across 1,000 AI Overviews found pages carrying named-source citations were cited 2.1x more often than pages without, and pages over 2,500 words earned 1.6x more citations than shorter ones. Semrush's analysis of roughly 89,000 LinkedIn URLs found about 75% of cited authors had posted at least five times in the previous four weeks. SeRanking found 49% of AI Overviews on explicit review-intent searches cited at least one review platform.
Two findings that are already percentages — Semrush (89k LinkedIn URLs) and SeRanking
The 75% figure is the one PR teams should sit with. It says the engines are not just citing publications, they are citing people who are currently visible — recency of the author's own activity correlates with whether their work gets pulled into an answer. That makes executive visibility programmes a citation lever rather than a brand-building nicety, and it is measurable in a way most thought-leadership work never has been.
The review-platform finding cuts differently. Just under half of review-intent answers cite a review platform, which means for a large class of commercial queries your G2 or Capterra presence is competing with your earned coverage for the same citation slot. PR does not own that surface, and a report that counts all citations equally will show you gaining ground that your product marketing team actually won. Splitting citations by source type before you attribute them is the same discipline behind our AI citation authority framework.
| FINDING | EFFECT | SOURCE |
|---|---|---|
| Pages carrying named-source citations | 2.1x more likely to be cited | Digital Applied, 1,000 AI Overviews |
| Pages over 2,500 words | 1.6x more citations than shorter pages | Digital Applied, 1,000 AI Overviews |
| Cited authors active in the last four weeks | ~75% of cited LinkedIn authors | Semrush, ~89,000 URLs |
| Review-intent answers citing a review platform | 49% of AI Overviews | SeRanking |
Read those findings together and they describe a job that looks a lot like classic PR with the measurement moved. Get named in sources the engines already trust. Make sure the coverage attributes to a real person. Keep the people who speak for the brand visibly active. Show up on the platforms that carry category-defining opinion. None of that is new work — it is the work, with a different scoreboard.
Which is the argument for not tearing up the PR programme to chase a new metric. The tactics that earned citations before earn them now; what changed is that you can finally see which placements did the earning. Agencies treating this as a reason to rebuild their service offering are solving a measurement problem with a strategy change, and the two engines-level findings above suggest the strategy was mostly right already. The gap we see in practice is narrower than the discourse implies — it is between coverage that gets counted and coverage that gets cited, which is what the 2026 AI citation study set out to separate.
An AI citation KPI you can defend
Report citation share by outlet as the headline, and carry the other numbers as context rather than as targets.
The third play is the one that saves the relationship. Most of the awkward renewal conversations we hear about come from a number that dropped with no explanation attached. Having a control metric — total citations returned for the prompt set, regardless of who they name — lets you say plainly whether the category moved or you did, which is the same separation that makes AI citation tracking worth instrumenting properly rather than pulling by hand.
What to put in next month's report
One headline number: how many of your earned outlets are being used as sources for your category's prompts, and which ones. One control number: total citations returned across the prompt set. Everything else — per-engine share, sentiment, share of voice — sits below the fold as context, labelled as things you observe rather than things you are steering.
That report is less impressive and far more defensible than the pie chart, and it has a property the pie chart lacks: when it moves, you can point at the pitch that moved it. If you are building this into an existing link building programme, the outlet list you already maintain is most of the work — you are changing what you count, not what you do.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.