Who Owns the AI Recommendation? A Multi-Industry Empirical Map of Brand Category Ownership Across Large Language Models

Large language models now mediate how buyers discover products and services, making the competitive structure of AI-generated recommendations a strategic concern for brands. A basic question has lacked large-scale empirical answers: in a given category, which brand does a model recommend, and how concentrated is that ownership? Across 3,750 responses spanning 50 brands, five industries, and 250 brand-free category queries on three models (GPT-5.2, Google Gemini 3 Flash, and Perplexity sonar-pro), each query repeated five times under a dice-roll stability protocol, we propose three exploratory metrics: the Category Ownership Index (COI), a brand's share of mentions within a category; the Competitive Vacuum Index (CVI), flagging categories with no single leader; and the Displacement Score (DS), quantifying asymmetric substitution between brand pairs. In this sample, recommendation concentration was moderate: the mean Gini coefficient was 0.28 (95% CI [0.16, 0.41]), below the 0.60 power-law threshold we set. Competitive vacuums were rare, appearing in 8.0% of queries, so the models named at least one sampled brand in most cases. Cross-model agreement on the top-recommended brand was 41.6%: a top position on one model did not reliably hold on another. Displacement was industry-dependent, from co-recommendation in consulting (0.4:1) to one-directional substitution up to 4.3:1, with an unweighted mean of 2.4:1 across the five industries. A BERTopic check placed only 4.2% of discovered topic clusters outside the original categories. Within the scope studied, these results sit in tension with a strong winner-takes-all narrative around AI recommendation, and the three metrics offer a candidate, reproducible procedure for competitive-intelligence analysis that future work can validate.

Summary

Main Finding

Across 3,750 LLM responses spanning 50 brands, five industries, 250 brand-free category queries, and three models (GPT‑5.2, Google Gemini 3 Flash, Perplexity sonar‑pro), AI recommendation was moderately concentrated but did not show a strong winner‑takes‑all pattern. Mean industry Gini = 0.28 (95% CI [0.16, 0.41]); competitive vacuums were uncommon (8.0% of queries); cross‑model agreement on the top brand was low (41.6%); and asymmetric displacement between brands was industry‑dependent (pooled mean displacement ratio ≈ 2.4:1, with consulting exhibiting co‑recommendation).

Key Points

Dataset and scope
- 3,750 responses: 250 distinct category queries × 3 LLMs × 5 iterations (dice‑roll protocol).
- 5 industries (SaaS, Consulting, FinTech, E‑commerce, HealthTech), 10 brands per industry (50 brands), 77 alias entries for brand matching.
- Models: GPT‑5.2 (parametric), Gemini 3 Flash (multimodal/web grounded), Perplexity sonar‑pro (RAG with citations).
Metrics proposed
- Category Ownership Index (COI): brand mentions intensity per query (uses mention counts, so can exceed 1.0).
- Recommendation Share (RS): brand’s average COI across an industry’s queries.
- Competitive Vacuum Index (CVI): 1 − max_b COI(b,q); CVI > 0.50 flags a vacuum; CVI < 0.20 flags monopoly.
- Displacement Score (DS): P(A | ¬B) − P(A | B), measures asymmetric substitution vs co‑occurrence.
Principal quantitative results
- Concentration: mean Gini = 0.28 (95% CI [0.16, 0.41]); industries vary (e‑commerce highest, Gini = 0.53; SaaS lowest, Gini = 0.18). H1 (power‑law concentration) not supported.
- Displacement: pooled displacement:co‑recommendation ratio ≈ 1.3:1; excluding consulting the unweighted mean across the other four industries ≈ 2.9:1; industry ratios ranged 2.0:1 to 4.3:1. H2 partially supported (asymmetry exists but not uniformly >3:1).
- Vacuums: 20/250 queries (8.0%) had CVI > 0.50. H3 (≥20% vacuums) not supported.
- Emergent categories: BERTopic found only 4.2% of clusters that did not map closely to the original queries (threshold cosine < 0.60). H4 (≥15% emergent) not supported.
- Cross‑model agreement: top brand agreed across the three models in 41.6% of queries. H5 (agreement <50%) supported.
Notable industry patterns
- Consulting: strong co‑recommendation / enumeration of major firms (MBB, Big‑4), leading to co‑occurrence rather than displacement.
- E‑commerce & SaaS: stronger displacement patterns (some challenger or niche brands show asymmetric substitution behavior).
- HealthTech: more competitive vacuums, especially in niche clinical subsegments.
Robustness / operational choices
- Dice‑roll protocol (5 repeats, temperature 0.3) used to accommodate LLM output variability.
- Regex alias matching (transparent/reproducible) used for extraction; short aliases (e.g., “EY”) noted as possible false positives.
- COI computed from mention counts (intensity) rather than binary presence.

Data & Methods

Experimental design
- Fully crossed observational design: 5 industries × 50 queries/industry × 3 models × 5 iterations = 3,750 responses.
- Query types: direct recommendation, comparison, category exploration, problem‑solution, trust/authority.
- System prompt and generation settings: temperature 0.3, max tokens 1,024, simple “helpful assistant” system prompt; rate limiting and checkpointing described.
Mention extraction and aggregation
- Brand mentions found by case‑insensitive alias pattern matching (77 aliases → 50 canonical brands).
- COI computed using raw mention counts across all model×iteration outputs for each query; RS is mean COI over an industry’s queries.
Displacement and network analysis
- DS computed for brand pairs per query; thresholds |DS| > 0.10 used to classify displacement (DS > 0.10) or co‑recommendation (DS < −0.10).
- Aggregated counts of pairwise displacement/co‑recommendation and directed network analysis (PageRank) used to identify brands with displacement power/vulnerability.
Topic discovery
- BERTopic pipeline: BGE‑M3 embeddings (1,024 dim), UMAP (15 neighbors, 10 components), HDBSCAN (min cluster size 20, min samples 5).
- Clusters matched to query embeddings; clusters with best‑match cosine < 0.60 labeled emergent.
Hypothesis tests reported against pre‑specified thresholds (e.g., Gini 0.60 cutoff, vacuum rate 20%, emergent >15%, cross‑model agreement <50%).

Implications for AI Economics

Recommendation gatekeeping is platform‑dependent, not universally winner‑takes‑all
- Moderate concentration and low cross‑model agreement imply that LLMs differ materially in which brands they surface. Brands face platform heterogeneity instead of a single monopolized AI discovery channel.
Strategic consequence for firms and GEO practice
- Brands aiming for visibility must monitor multiple LLMs and optimise for generation‑centric signals (what the paper calls GEO), not only traditional SEO. Divergent top‑brand lists across models create both risk and opportunity.
Competitive dynamics vary by industry and query type
- Consulting shows bundled/co‑recommendation behavior (models enumerate incumbents together), reducing direct displacement. In contrast, several industries permit asymmetric displacement, meaning dominant mention of one brand tends to exclude others—this can amplify discovery advantages.
Measurement and competitive intelligence tools
- COI, CVI, and DS offer transparent, reproducible metrics for tracking AI recommendation ownership; they can serve as a starting standard for GEO/CI benchmarking, subject to further validation (binary vs intensity choices, alias matching precision).
Policy and market‑structure concerns
- Even if winner‑take‑all effects are not universal, LLMs act as algorithmic gatekeepers with nontrivial influence on consideration sets. Regulators and platform designers should consider transparency, auditability, and potential bias (e.g., demographic sensitivity found in prior work).
Research and commercial next steps
- Link LLM recommendation patterns to downstream economic outcomes (clicks, signups, revenue) to estimate causal effects on market shares.
- Broader monitoring: more models, languages, geographies, dynamic tracking over time as model updates and fine‑tuning alter recommendations.
- Improve extraction and evaluation: move beyond regex aliasing to entity linking and citation‑level grounding to reduce false positives and better attribute web‑grounded recommendations.

Limitations noted by the authors (relevant for interpreting economic implications) - Single‑time snapshot (Feb 23–24, 2026) — model behaviors evolve rapidly. - Model sample limited to three platforms; other assistants and vertical agents may differ. - Mention extraction via alias regex can produce false positives; COI based on mention counts captures intensity but may overweight verbose responses. - Construct validity of COI/CVI/DS requires further external validation (e.g., user behavior linkage).

Takeaway: LLM‑mediated brand discovery is consequential but heterogeneous. The proposed COI/CVI/DS framework provides reproducible, industry‑aware metrics for mapping AI recommendation ownership; early empirical evidence suggests moderate concentration, frequent platform dependence, and industry‑specific displacement dynamics rather than universal winner‑takes‑all outcomes.

Assessment

Paper Typedescriptive Evidence Strengthmedium — The paper reports a large (3,750-response) systematic snapshot across three major models, uses repeated queries to assess stability, and introduces quantitative metrics; however, it makes no causal claims, is exploratory, uses a limited set of brands, industries, and model snapshots, and does not link recommendations to real-world user behavior or economic outcomes. Methods Rigormedium — The authors employ a reproducible protocol (250 brand-free category queries, five model repeats per query, a dice-roll stability procedure), propose clear metrics (COI, CVI, DS), and validate topic coverage with BERTopic, but the study is constrained by sample selection (brands, categories, languages), single-time snapshots of closed models, potential sensitivity to prompt phrasing, and lack of robustness checks tying outputs to downstream measures (clicks, conversions, revenue). Sample3,750 model responses generated from 250 brand-free category queries covering 50 brands across five industries, collected from three LLM-based recommendation systems (GPT-5.2, Google Gemini 3 Flash, Perplexity sonar-pro); each query was repeated five times under a dice-roll stability protocol; topic validation via BERTopic. Themesadoption innovation GeneralizabilityOnly three proprietary models and specific model versions — results may not hold for other or future models, Limited to 50 selected brands and five industries — not representative of all product/service categories, Uses controlled brand-free category prompts rather than real-world user queries, which may change recommendations, Single cross-sectional snapshot — model behavior may shift over time with updates or fine-tuning, Likely English- and market-specific (geography/language not reported), limiting global generalizability, Does not measure user behavior (clicks, purchases) so economic implications for firms are inferred rather than observed

Claims (9)

Claim	Direction	Outcome	Confidence & Evidence	Details
The study collected 3,750 responses spanning 50 brands, five industries, and 250 brand-free category queries on three models (GPT-5.2, Google Gemini 3 Flash, and Perplexity sonar-pro), with each query repeated five times under a dice-roll stability protocol. Other	null_result	dataset/sample composition (responses, brands, industries, queries, models, repeats)	Reading fidelity high Study strength high	n=3750 0.3
The paper proposes three exploratory metrics for analyzing AI recommendations: the Category Ownership Index (COI), the Competitive Vacuum Index (CVI), and the Displacement Score (DS). Other	null_result	introduction of COI, CVI, and DS as metrics	Reading fidelity high Study strength medium	not reported 0.18
Recommendation concentration was moderate in the sample: the mean Gini coefficient was 0.28 (95% CI [0.16, 0.41]), below the 0.60 power-law threshold the authors set. Market Structure	negative	Gini coefficient of brand mentions within categories (recommendation concentration)	Reading fidelity high Study strength medium	n=250 0.28 (95% CI [0.16, 0.41]) 0.18
Competitive vacuums were rare, appearing in 8.0% of queries; models named at least one sampled brand in most cases. Market Structure	negative	proportion of queries with no single leader (competitive vacuum)	Reading fidelity high Study strength medium	n=250 8.0% 0.18
Cross-model agreement on the top-recommended brand was 41.6%; a top position on one model did not reliably hold on another. Market Structure	mixed	cross-model agreement rate for top-recommended brand	Reading fidelity high Study strength medium	n=250 41.6% 0.18
Displacement (asymmetric substitution between brand pairs) was industry-dependent, ranging from co-recommendation in consulting (0.4:1) to one-directional substitution up to 4.3:1, with an unweighted mean of 2.4:1 across the five industries. Market Structure	mixed	Displacement Score (ratio of asymmetric substitution between brand pairs)	Reading fidelity high Study strength medium	n=5 0.4:1 to 4.3:1, mean 2.4:1 0.18
A BERTopic check placed only 4.2% of discovered topic clusters outside the original categories. Other	null_result	proportion of topic clusters that fall outside the original category taxonomy	Reading fidelity medium Study strength low	4.2% 0.05
Within the scope studied, these results sit in tension with a strong winner-takes-all narrative around AI recommendation. Market Structure	negative	degree of winner-takes-all concentration in AI recommendations	Reading fidelity high Study strength speculative	n=250 0.03
The three metrics (COI, CVI, DS) offer a candidate, reproducible procedure for competitive-intelligence analysis that future work can validate. Other	null_result	reproducibility/usefulness of proposed metrics for competitive intelligence	Reading fidelity high Study strength speculative	not reported 0.03

LLM recommendations do not produce ubiquitous winner-takes-all outcomes: across 3,750 brand-free queries to three models, recommendation concentration was moderate and top-brand agreement across models under 42%, though displacement patterns varied sharply by industry.