The most counterintuitive finding in marketing research has a specific shape: it reveals that what practitioners have been optimising for is not only insufficient but actively misleading. RankCaster AI's study of five point two two million AI citation records, published Thursday, has that shape. The central finding — that pages receiving the highest volumes of AI citations are the pages most likely to disappear from AI answers within weeks, whilst pages with dramatically fewer citations maintain visibility for four months and beyond — inverts the implicit assumption that has driven AI visibility strategy since GEO emerged as a discipline.
The study analysed monitoring data from forty-four organisations across fourteen thousand one hundred and fifty-two unique URLs, tracking citations across six AI providers: Claude, ChatGPT, Gemini, Perplexity, DeepSeek, and Google AI Overview. The analytical framework distinguishes between two citation pattern types. "Spike" pages receive on average seven hundred and twenty-three citations but exhaust their AI visibility within weeks. "Persistent" pages receive on average forty-one citations — roughly seventeen times fewer — but maintain that presence for a hundred and thirty-nine days or more. The volume gap and the persistence gap run in precisely opposite directions. A brand that measured its AI visibility strategy by citation count would have congratulated itself on exactly the pages that were about to vanish.
The mechanism behind the divergence emerges from the study's most significant finding: cross-provider presence is the strongest predictor of citation persistence. Not a single URL cited by only one AI provider achieved Persistent status in the dataset. At the other extreme, URLs cited by all six providers showed a thirty-six point eight per cent Persistent rate — the highest of any measured variable. The implication is structural: AI citation persistence is not a property of the content itself but a property of how that content is received across the distributed AI ecosystem. A page that a single AI system has found useful is fragile. A page that six AI systems independently cite, for different queries, in different contexts, has embedded itself into the informational infrastructure that those systems draw on.
The second significant finding concerns semantic breadth. Pages that span multiple distinct query clusters — that are relevant to different categories of question simultaneously — show substantially higher persistence than pages optimised narrowly for a single topic or query type. The practical implication challenges a fundamental instinct in content strategy: the inclination to produce highly specific, deeply authoritative content on a single subject. Deep specificity may produce strong performance in traditional search, where keyword matching rewards precision. It produces fragile AI visibility, because narrow pages create few entry points for the diverse, contextual queries that AI systems handle.
The content formats that demonstrate highest persistence are, in retrospect, explicable by this logic. Directories, event lists, rankings, and entity-rich pages — formats that by nature cover multiple entities, relationships, and contexts — consistently outperform single-subject articles on persistence metrics. These pages are, architecturally, designed to be relevant to many things at once. That property, it turns out, is exactly what AI systems reward over time.
To operationalise these findings, RankCaster AI introduces the Citation Lifetime Score — a composite zero-to-one-hundred metric that assesses URL-level persistence across three dimensions: provider diversity, weighted at fifty per cent; semantic breadth, at twenty-five per cent; and late citation share, the proportion of citations occurring after the initial distribution period, at twenty-five per cent. The weighting reflects the study's empirical findings: provider diversity is the dominant predictor of persistence, with semantic breadth and temporal distribution as important but secondary factors.
"Traditional SEO asks: how do I rank at the top for a keyword?" said Andy Terekhin, chief executive of RankCaster AI. "AI Visibility asks a different question: how do I create a page that multiple AI systems will continue to use as a source for different query classes? Our research shows that persistence, not volume, is the true measure of long-term AI presence."
The CLS framework gives marketing teams a diagnostic tool that citation volume counts cannot provide. A page with a high citation volume and a low CLS is a liability: it is generating AI presence that will not compound. A page with a lower citation count and a high CLS is an asset: it is building the kind of distributed, semantically broad presence that tends to self-reinforce as AI systems learn from each other's citation patterns over time.
For enterprise marketing teams, the study's practical recommendations are unusually specific. The creation of entity-rich pages — lists, directories, and rankings that cover multiple search intents — is supported by the persistence data and represents a category of content investment that most organisations have underweighted relative to long-form single-subject articles. Distribution across the ecosystems accessible to different AI providers is now a distinct content distribution function, not an incidental benefit of general publishing. And the measurement of citation lifetime, rather than citation volume, requires tooling that the majority of marketing analytics stacks do not yet provide.
The study was conducted on a dataset that RankCaster AI acknowledges is concentrated — forty-four client organisations — and the company has outlined eight directions for future research, including the role of entity density in persistence, the temporal relationship between semantic breadth and cross-provider citation patterns, and the construction of probability models for citation survival at thirty, sixty, ninety, and one hundred and eighty days. The research represents, by the scale of the dataset and the novelty of the analytical framework, the most empirically grounded contribution to the GEO field to date.