New indices measure which brands own LLM category answers
A new empirical framework quantifies brand ownership in AI recommendations. For B2B brands, LLM visibility is now a pipeline metric, not an awareness one.
Key takeaways
- A mean Gini of 0.28 masks extreme category-level concentration where one brand captures most LLM mentions.
- The Competitive Vacuum Index flags categories with no dominant AI-recommended brand, signalling an unclaimed position.
- The Displacement Score reveals asymmetric substitution: some rivals are actively taking your citations, others are merely present.
- For financial services and industrial procurement, absence from LLM answers means removal from a buying conversation before it starts.
- Brands with strong editorial presence and third-party citations may be able to shift their Category Ownership Index through content strategy.
Three large language models, 3,750 responses, 50 brands, and a single uncomfortable finding: in most product categories, a small number of brands absorb the majority of AI-generated recommendations, and most competitors do not register at all. The study, published on arXiv by researchers mapping brand visibility across GPT-4, Google Gemini 3 Flash, and Perplexity sonar-pro, is among the first to apply formal concentration metrics to the question of who actually owns a category inside an LLM answer.
The researchers ran 250 brand-free category queries across five industries, repeating each five times under a stability protocol. From 3,750 responses, they derived three metrics: the Category Ownership Index (COI), which measures a brand's share of mentions within a category; the Competitive Vacuum Index (CVI), which flags categories where no single leader exists; and the Displacement Score (DS), which quantifies asymmetric substitution between brand pairs. The mean Gini coefficient across categories was 0.28, below the 0.60 threshold typical of winner-take-all markets, but the distribution was uneven. Some categories showed sharp concentration; others showed none. Both findings carry consequences.
Concentration is moderate on average, dangerous at the extremes
A mean Gini of 0.28 might sound reassuring. It should not be. Averages obscure the categories where one brand captures three or four of every five model mentions while rivals receive rounding-error shares. For a brand in one of those concentrated categories that is not the incumbent leader, the practical effect is near-total invisibility in AI-mediated discovery. The buyer never sees a shortlist; the model produces a name.
The Displacement Score makes the competitive asymmetry precise. In a pair of competing brands, substitution is rarely symmetric: brand A losing share to brand B is a different event from brand B losing share to A, with different magnitudes. For marketing strategy, this matters because it identifies which rivals are actively eating recommendation share and which are merely present. A financial services firm mapping its AI visibility can now, in principle, distinguish between a competitor that is taking its citations and one that is simply co-mentioned.
The Competitive Vacuum Index is the metric with the most immediate opportunity signal. Categories flagged by CVI have no dominant model-recommended brand. In those spaces, the competitive position has not yet been claimed inside the LLM's implicit ranking. That is not a stable equilibrium. Models will learn from new data, fine-tuning and retrieval will shift, and the brand that builds the strongest corpus of authoritative, frequently-cited content in that category before consolidation happens will likely be the one that owns it afterward.
What this means for B2B brands that do not sell toothpaste
The study covers consumer and B2B categories together, but the implications fall hardest on sectors where AI-mediated discovery is becoming the first point of contact: financial services, where a treasurer searching for trade-finance providers increasingly starts with an LLM prompt; multilateral and UN procurement contexts, where consultants and vendors are evaluated through AI-assisted briefings; and industrial groups, where procurement teams query models for supplier shortlists before any RFP is issued.