A PR team reports that brand mentions climbed last quarter and share of voice now leads the category. The AEO dashboard, which runs the same prompts every week, shows the brand cited in about the same handful of AI answers as before. Both numbers are accurate. The question is which one tells you where AI visibility is heading. Vendors increasingly pitch share of voice as a leading indicator for AEO. Some of that pitch holds up under the evidence, and some of it does not.
Key Takeaways
- Brand mentions shape AI visibility through two real mechanisms: models learn entities from repeated documents, and AI search engines retrieve mostly from third-party sources.
- Most evidence linking mentions to AI visibility compares brands at a single point in time, and large brands score high on both, so it cannot show that adding mentions will lift your citations.
- Treat share of voice as predictive only after your own week-by-week data shows mentions in the sources that engines cite moving ahead of your Share of Citation.
Why Mentions Became the Default Leading Indicator
Share of Citation is the metric AEO teams actually care about: how often AI models pull a brand into answers and name it as a source. It is also slow to move, it varies by engine, and a small prompt set makes it noisy. Mentions are cheap to count, PR teams already report them, and they move every week. When a metric is easy to collect and tells a plausible causal story, it ends up in the deck.
The causal story is real, at least in outline. AEO works on two layers. Training-data presence is earned slowly through wide coverage. Retrieval presence is earned technically through structured, current content. Mentions feed the first layer directly and the second layer partly. A “leading indicator” makes a stronger claim than that, though. It claims a change in mentions today predicts a change in citations later. That claim needs evidence about timing and causation, and most of what gets quoted does not provide it.
What the Research Actually Supports
The strongest evidence for the training layer comes from a 2022 paper titled Large Language Models Struggle to Learn Long-Tail Knowledge. The researchers counted how many pre-training documents contained the entities in a given question, then measured whether models answered it correctly. They found strong correlational and causal relationships between accuracy and the number of relevant documents, across several datasets and model sizes. For a brand, the implication is direct. A model that has seen your name alongside your category in many documents is more likely to know what you do.
The same paper found that retrieval augmentation reduces this dependence on pre-training data. So the training layer is only half the picture. Live retrieval is the other half, and here the question is which sources engines reach for.
A 2025 study, Generative Engine Optimization: How to Dominate AI Search, compared AI search engines with Google across verticals, languages and query paraphrases. Its authors report that AI search shows a “systematic and overwhelming bias towards Earned media” over brand-owned and social content, in contrast with Google’s more balanced mix. They also found that engines differ from one another in domain diversity, freshness and sensitivity to phrasing. Third-party coverage, which is what mention counts measure, is the pool AI engines draw from.
Mentions do not settle the outcome on their own. The original GEO paper from 2023, which introduced the term, showed that changes to content alone could raise visibility in generative engine responses by up to 40%, and that what worked varied by domain. Whether you get retrieved depends on how often you are mentioned, and it also depends on the page that holds the mention.
Where Share of Voice Breaks as a Predictor
The first problem is confounding. The 2025 study names a “big brand bias” that niche players have to overcome. Large brands have more mentions and more citations at the same time, so any chart that plots mentions against AI visibility across many brands mostly measures company size. It cannot tell you whether a mid-market brand that adds mentions will see its citations rise.
The second problem is source mismatch. Share of voice counts every mention the same way. Engines do not. ChatGPT pulls from across the open web, Claude leans on high-authority publications and documentation, and Grok reads the real-time social web. A spike in social mentions might move Grok and do nothing for Claude. One blended share of voice figure hides this, which is how a brand ends up with high share of voice and near-zero Share of Citation.
The third problem is timing. The training-layer effect shows up only when a model is retrained. The retrieval-layer effect can show up within days of a page being indexed. A single share of voice number mixes two very different time lags. The fourth problem is agreement. Trust is a graph of agreement across independent sources, so a thousand mentions that describe your product in different ways do less than a few hundred that describe it the same way.
A mention predicts a citation only when it sits in a source the model actually reads and says what the other sources say.
This is why PR goals work better when they are set inside AI answers rather than in a clippings count. For NewDay USA, PR topic visibility was measured directly on AI engines.
PR topic visibility for NewDay USA against a +10% target, tracked weekly on five engines over six months.See the case study.
How Do You Test Whether Your Mentions Predict Anything?
Run the test on one brand over time, because comparisons across brands are muddied by company size. Fix a prompt set that reflects real buying questions, then record Share of Citation separately for each engine every week. Log which domains each engine cites for those prompts. After a month you will have a list of domains that each engine actually draws from in your category.
Next, split your mention data in two: mentions on domains that appear in that citation list, and everything else. If share of voice is a genuine leading indicator for you, growth in the first group should come before growth in Share of Citation on the matching engine. Growth in the second group should not. If both groups move together with citations, or neither does, the metric is not yet telling you anything you can act on.
Earned placements give you a faster test. When a story runs on a domain an engine already cites, check whether that URL appears in answers over the following weeks. A retrieval effect shows up quickly or not at all. A training-layer effect will not appear until the next model update, so do not judge it on a quarterly chart. A few months of weekly data will give you a direction to work from. It will not prove the effect.
What to Do Before the Next Reporting Cycle
Start by changing the report. Show share of voice next to Share of Citation, per engine, and stop presenting one blended mention figure as a proxy for AI visibility.
- Build a citation-domain list for each engine from four weeks of prompt tracking, and update it monthly.
- Split mention reporting into mentions on citation domains and all other mentions.
- Point PR targets at the citation domains, and set the goal as visibility in AI answers on those topics.
- Audit how your top ten cited sources describe your category, pricing and differentiators, then correct the ones that disagree.
- Review the lag between citation-domain mentions and Share of Citation every quarter, and drop the leading-indicator label if no lag appears.
Citations compound. Once a model cites a brand, the next citation becomes more likely. That is the case for finding the mentions that really predict citations, and for cutting the ones that only make the chart look busy.
FAQs
Is Share of Voice a Leading Indicator for AI Visibility?
Only partly. Research shows that models learn more about entities that appear in more documents, and that AI search engines draw mostly on third-party sources. Overall share of voice also counts mentions that engines never read, so it becomes predictive only when you limit it to the sources each engine actually cites.
What Is the Difference Between Share of Voice and Share of Citation?
Share of voice measures how often a brand is mentioned across the web overall. Share of Citation measures how often AI models pull a brand into answers and name it as a source. A brand can have high share of voice and near-zero Share of Citation.
How Long Does It Take for New Brand Mentions to Show Up in AI Answers?
It depends on the layer. A mention on a domain that an engine retrieves from can appear in answers within weeks of being indexed. A mention that only affects training data has no effect until the model is retrained, which can take months.
Which Brand Mentions Matter Most for AEO?
The mentions that matter most are on domains that AI engines already cite for your category’s prompts, and that describe your brand the same way other sources do. Research comparing AI search with Google found a strong bias toward earned, third-party media over brand-owned content.