pith. sign in

arxiv: 2606.23057 · v1 · pith:NBHV5ECQnew · submitted 2026-06-22 · 💻 cs.IR · cs.CL· cs.CY· cs.LG

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

Dmitrij \.Zatuchin This is my paper

Pith reviewed 2026-06-26 06:52 UTC · model grok-4.3

classification 💻 cs.IR cs.CLcs.CYcs.LG
keywords AI recommendationsbrand category ownershiplarge language modelscompetitive intelligenceGini coefficientrecommendation concentrationmulti-industry analysis
0
0 comments X

The pith

Large language models show moderate concentration in brand recommendations rather than strong dominance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper empirically maps which brands LLMs recommend when given brand-free category queries across five industries. It defines three metrics to quantify ownership share, absence of a leader, and substitution patterns between brands. In the sampled queries the mean Gini coefficient sits at 0.28, well below the 0.60 power-law threshold, competitive vacuums appear in only 8 percent of cases, and the three models agree on the top brand just 41.6 percent of the time. A sympathetic reader would care because these patterns indicate that recommendation ownership is distributed and model-specific rather than locked to a single winner.

Core claim

Across 3,750 responses spanning 50 brands, five industries, and 250 category queries on three models, recommendation concentration was moderate with a mean Gini coefficient of 0.28, competitive vacuums were rare at 8.0 percent, cross-model agreement on the top-recommended brand was 41.6 percent, and displacement ratios averaged 2.4 to 1 with industry variation.

What carries the argument

The Category Ownership Index (COI) for brand share within a category, the Competitive Vacuum Index (CVI) for categories lacking a clear leader, and the Displacement Score (DS) for asymmetric substitution between brand pairs.

If this is right

  • A top position on one model does not reliably transfer to another.
  • Displacement between brands is industry-dependent rather than uniform.
  • The models name at least one sampled brand in the great majority of queries.
  • The three metrics supply a reproducible procedure for competitive-intelligence tracking.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Brands may need separate visibility strategies for each major model instead of a single approach.
  • Applying the same metrics to real user search logs could test whether the observed patterns hold outside controlled prompts.
  • Low cross-model agreement suggests that ownership of AI recommendations is more fragmented than platform-level market share data would imply.

Load-bearing premise

The 50 brands and 250 brand-free category queries adequately represent the competitive landscape and user behavior in the five industries.

What would settle it

Repeating the protocol on a substantially larger set of categories and finding a mean Gini coefficient above 0.60 or competitive vacuums in more than 15 percent of queries would falsify the moderate-concentration result.

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents an empirical analysis of brand recommendations generated by three large language models across five industries. Using 250 brand-free category queries repeated five times each, yielding 3,750 responses, the authors define three metrics—Category Ownership Index (COI), Competitive Vacuum Index (CVI), and Displacement Score (DS)—and report moderate recommendation concentration (mean Gini coefficient of 0.28 with 95% CI [0.16, 0.41]), low incidence of competitive vacuums (8.0%), 41.6% cross-model agreement on top brands, and industry-dependent displacement ratios (unweighted mean 2.4:1).

Significance. If the brand and query sampling is representative of the underlying competitive landscapes, the work supplies a reproducible, statistics-backed procedure for competitive-intelligence analysis of AI recommendations, including confidence intervals from repeated queries and a stability protocol. The three metrics offer falsifiable quantities that future studies can apply or extend, providing a concrete empirical counterpoint to winner-takes-all claims within the studied scope.

major comments (2)
  1. [Abstract] Abstract: the central claims of moderate concentration (mean Gini 0.28 below the 0.60 threshold) and rare competitive vacuums (8.0%) are computed exclusively over a fixed set of 50 pre-selected brands and 250 queries; the absence of any selection protocol, market-share coverage verification, or query-validation step makes these quantities describe an artificial subset rather than the actual recommendation landscape, directly undermining generalization to the five industries.
  2. [Abstract] Abstract: the 0.60 power-law threshold used to interpret the Gini coefficient is introduced without reference to prior literature on concentration indices or sensitivity checks, rendering the statement that observed concentration is 'below the threshold' dependent on an arbitrary cutoff whose effect on the tension with winner-takes-all narratives cannot be assessed.
minor comments (2)
  1. The abstract references a 'dice-roll stability protocol' and BERTopic validation (4.2% of clusters outside categories) but supplies no implementation details or robustness checks for either.
  2. Consider adding a table or appendix listing the five industries, the 50 brands, and sample query phrasing to allow readers to evaluate coverage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and metric interpretation. We address each major comment below and will revise the manuscript to improve transparency and rigor.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of moderate concentration (mean Gini 0.28 below the 0.60 threshold) and rare competitive vacuums (8.0%) are computed exclusively over a fixed set of 50 pre-selected brands and 250 queries; the absence of any selection protocol, market-share coverage verification, or query-validation step makes these quantities describe an artificial subset rather than the actual recommendation landscape, directly undermining generalization to the five industries.

    Authors: We agree the abstract and methods require greater explicitness on sampling to support the claims. The 50 brands were the top 10 by market share per industry drawn from public reports (e.g., Statista, Euromonitor), and the 250 queries were generated from standard industry category taxonomies; a BERTopic validation already showed only 4.2% off-category clusters. We will add a dedicated 'Brand and Query Sampling' subsection in Methods detailing selection criteria, data sources, and coverage checks, and revise the abstract to state that findings apply to this representative sample of major brands rather than claiming exhaustive industry coverage. This clarifies scope without altering the reported statistics. revision: yes

  2. Referee: [Abstract] Abstract: the 0.60 power-law threshold used to interpret the Gini coefficient is introduced without reference to prior literature on concentration indices or sensitivity checks, rendering the statement that observed concentration is 'below the threshold' dependent on an arbitrary cutoff whose effect on the tension with winner-takes-all narratives cannot be assessed.

    Authors: The referee is correct that the 0.60 threshold lacks citations and sensitivity analysis. We will insert references to prior work on Gini and concentration indices in digital markets (e.g., studies of platform power-law distributions) and add a sensitivity table in Results showing classification stability at 0.50 and 0.70 cutoffs. The abstract will be updated to note the threshold as a heuristic benchmark with these robustness checks, allowing readers to evaluate its role in the moderate-concentration finding independently. revision: yes

Circularity Check

0 steps flagged

No circularity: all reported statistics are direct aggregations of raw response counts

full rationale

The paper's central results (mean Gini 0.28, 8% competitive vacuums, 41.6% cross-model agreement, displacement ratios) are computed from explicit mention counts in 3,750 LLM responses using standard definitions of share, Gini coefficient, and agreement. COI, CVI, and DS are defined as direct functions of those counts within the fixed sample of 50 brands; no parameter is fitted on a subset and then re-predicted, no self-citation chain justifies a uniqueness claim, and no ansatz or renaming reduces the output to the input selection by construction. The derivation chain is therefore self-contained as descriptive statistics on the observed data.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claims rest on the assumption that the chosen brands and queries form a representative sample of industry competition and that repeated queries under the stability protocol reflect stable model behavior; no new physical entities or mathematical axioms beyond standard statistics are introduced.

free parameters (2)
  • Power-law concentration threshold = 0.60
    Arbitrarily chosen cutoff of 0.60 used to interpret Gini values as indicating strong concentration.
  • Number of query repetitions = 5
    Fixed at five to implement the dice-roll stability protocol.
axioms (1)
  • domain assumption The 50 brands and 250 category queries are representative of the five industries.
    Generalization of the moderate-concentration and low-agreement findings depends on this sampling being unbiased and comprehensive.

pith-pipeline@v0.9.1-grok · 5894 in / 1350 out tokens · 29094 ms · 2026-06-26T06:52:11.416114+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. How Large Language Models Source Brand Reputation Across Languages and Markets

    cs.IR 2026-06 unverdicted novelty 5.0

    LLMs cite third-party domains for 85.7% of brand attributions, with Wikipedia dominant in most languages, a long-tailed domain distribution, and market-specific shifts such as YouTube and HR sites in Poland.

Reference graph

Works this paper leans on

27 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S.: On the dangers of stochastic parrots: Can language models be too big? In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21), pp. 610–623. ACM, ??? (2021). https://doi.org/10.1145/3442188.3445922

    work page doi:10.1145/3442188.3445922 2021

  2. [2]

    Journal of the Academy of Marketing Science 48(1), 24–42 (2020) https://doi.org/10.1007/s11747-019-00696-0

    Davenport, T., Guha, A., Grewal, D., Bressgott, T.: How artificial intelligence will change the future of marketing. Journal of the Academy of Marketing Science 48(1), 24–42 (2020) https://doi.org/10.1007/s11747-019-00696-0

    work page doi:10.1007/s11747-019-00696-0 2020

  3. [3]

    In: Advances in Neural Information Processing Systems, vol

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A.,et al.: Language models are few-shot learners. In: Advances in Neural Information Processing Systems, vol. 33, pp. 1877–1901 (2020)

    1901

  4. [4]

    PREPRINT (Version 1) available at Research Square (2026)

    ˙Zatuchin, D.: Gender Bias in Large Language Model Brand Recommendations: A Three-Study Analysis of Prompt-Induced Disparities Across Seasonal and Recip- ient Contexts. PREPRINT (Version 1) available at Research Square (2026). https://doi.org/10.21203/rs.3.rs-8883056/v1 18

    work page doi:10.21203/rs.3.rs-8883056/v1 2026

  5. [5]

    optimize your brand for LLMs

    Dubois, D., Dawson, J., Jaiswal, A.: Forget what you know about search. optimize your brand for LLMs. Harvard Business Review (2025). https://hbr.org/2025/06/ forget-what-you-know-about-seo-heres-how-to-optimize-your-brand-for-llms

    2025

  6. [6]

    Public Opinion Quarterly36(2), 176–187 (1972) https://doi.org/10.1086/267990

    McCombs, M.E., Shaw, D.L.: The agenda-setting function of mass media. Public Opinion Quarterly36(2), 176–187 (1972) https://doi.org/10.1086/267990

    work page doi:10.1086/267990 1972

  7. [7]

    Academy of Management Journal33(2), 233–258 (1990) https://doi

    Fombrun, C., Shanley, M.: What’s in a name? reputation building and corporate strategy. Academy of Management Journal33(2), 233–258 (1990) https://doi. org/10.2307/256324

    work page doi:10.2307/256324 1990

  8. [8]

    International Journal of Engineering Business Management17(2025) https://doi.org/10.1177/ 18479790251351889

    Lopez-Lopez, D., Bara Iniesta, M.: The impact of conversational AI on con- sumer decision-making: A systematic review and cluster analysis. International Journal of Engineering Business Management17(2025) https://doi.org/10.1177/ 18479790251351889

    2025

  9. [9]

    2463– 2473 (2019)

    Petroni, F., Rockt¨ aschel, T., Riedel, S., Lewis, P., Bakhtin, A., Wu, Y., Miller, A.: Language models as knowledge bases? In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 2463– 2473 (2019). https://doi.org/10.18653/v1/D19-1250

    work page doi:10.18653/v1/d19-1250 2019

  10. [10]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pp

    Mallen, A., Asai, A., Zhong, V., Das, R., Hajishirzi, H., Khashabi, D.: When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pp. 9802–9822 (2023). https: //doi.org/10.18653/v1/2023.acl-long.546

    work page doi:10.18653/v1/2023.acl-long.546 2023

  11. [11]

    In: Advances in Neural Information Processing Systems, vol

    Bolukbasi, T., Chang, K.-W., Zou, J.Y., Saligrama, V., Kalai, A.T.: Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In: Advances in Neural Information Processing Systems, vol. 29, pp. 4349–4357 (2016)

    2016

  12. [12]

    In: Proceedings of the 13th ACM Conference on Recommender Systems (RecSys ’19), pp

    Ekstrand, M.D., Burke, R., Diaz, F.: Fairness and discrimination in recommenda- tion and retrieval. In: Proceedings of the 13th ACM Conference on Recommender Systems (RecSys ’19), pp. 576–577. ACM, ??? (2019). https://doi.org/10.1145/ 3298689.3346964

    arXiv 2019

  13. [13]

    Federal Reserve Bulletin79(3), 188–189 (1993)

    Rhoades, S.A.: The Herfindahl-Hirschman index. Federal Reserve Bulletin79(3), 188–189 (1993)

    1993

  14. [14]

    R., & Newman, M

    Clauset, A., Shalizi, C.R., Newman, M.E.J.: Power-law distributions in empirical data. SIAM Review51(4), 661–703 (2009) https://doi.org/10.1137/070710111

    work page doi:10.1137/070710111 2009

  15. [15]

    Contemporary Physics46(5), 323–351 (2005) https://doi.org/10.1080/00107510500052444 19

    Newman, M.E.J.: Power laws, Pareto distributions and Zipf’s law. Contemporary Physics46(5), 323–351 (2005) https://doi.org/10.1080/00107510500052444 19

    work page doi:10.1080/00107510500052444 2005

  16. [16]

    Quarterly Journal of Economics 135(2), 645–709 (2020) https://doi.org/10.1093/qje/qjz025

    Autor, D., Dorn, D., Katz, L.F., Patterson, C., Van Reenen, J.: The fall of the labor share and the rise of superstar firms. Quarterly Journal of Economics 135(2), 645–709 (2020) https://doi.org/10.1093/qje/qjz025

    work page doi:10.1093/qje/qjz025 2020

  17. [17]

    In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp

    Dodge, J., Sap, M., Marasovi´ c, A., Agnew, W., Ilharco, G., Groeneveld, D., Mitchell, M., Gardner, M.: Documenting large webtext corpora: A case study on the Colossal Clean Crawled Corpus. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1286–1305 (2021). https://doi.org/10.18653/v1/2021.emnlp-main.98

    work page doi:10.18653/v1/2021.emnlp-main.98 2021

  18. [18]

    In: Proceedings of the 40th International Conference on Machine Learning (ICML), pp

    Kandpal, N., Deng, H., Roberts, A., Wallace, E., Raffel, C.: Large language mod- els struggle to learn long-tail knowledge. In: Proceedings of the 40th International Conference on Machine Learning (ICML), pp. 15696–15707. PMLR, ??? (2023)

    2023

  19. [19]

    Springer, Berlin, Heidelberg (2010)

    Celma, `O.: Music Recommendation and Discovery: The Long Tail, Long Fail, and Long Play in the Digital Music Space. Springer, Berlin, Heidelberg (2010). https://doi.org/10.1007/978-3-642-13287-2

    work page doi:10.1007/978-3-642-13287-2 2010

  20. [20]

    Pearson Prentice Hall, Upper Saddle River, NJ (2007)

    Fleisher, C.S., Bensoussan, B.E.: Strategic and Competitive Analysis: Meth- ods and Techniques for Analyzing Business Competition. Pearson Prentice Hall, Upper Saddle River, NJ (2007)

    2007

  21. [21]

    Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A.: GEO: Generative Engine Optimization (2023)

    2023

  22. [22]

    Technical report, BrightEdge (2025)

    BrightEdge: AI search visits surging in 2025: But organic search remains the cornerstone of digital growth. Technical report, BrightEdge (2025). https://www. brightedge.com/resources/research-reports/ai-search-visits-in-surging-2025

    2025

  23. [23]

    In: Advances in Neural Information Processing Systems, vol

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A.,et al.: Training language models to follow instructions with human feedback. In: Advances in Neural Information Processing Systems, vol. 35, pp. 27730–27744 (2022)

    2022

  24. [24]

    Grootendorst, M.: BERTopic: Neural Topic Modeling with a Class-Based TF-IDF Procedure (2022)

    2022

  25. [25]

    Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., Liu, Z.: BGE M3- Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embed- dings Through Self-Knowledge Distillation (2024)

    2024

  26. [26]

    McInnes, L., Healy, J., Melville, J.: UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction (2018)

    2018

  27. [27]

    In: Advances in Knowledge Discovery and Data Mining (PAKDD 2013)

    Campello, R.J.G.B., Moulavi, D., Sander, J.: Density-based clustering based on hierarchical density estimates. In: Advances in Knowledge Discovery and Data Mining (PAKDD 2013). LNCS, vol. 7819, pp. 160–172. Springer, ??? (2013). 20 https://doi.org/10.1007/978-3-642-37456-2 14 21

    work page doi:10.1007/978-3-642-37456-2 2013