Auditing Preferences for Brands and Cultures in LLMs

Jasmine Rienecker, Katarina Mpofu, Naman Goel, Siddhartha Datta, Jun Zhao, Oscar Danielsson, Fredrik ThorsenarXiv 2026 · DOI

TL;DR

Large Language Models (LLMs) are increasingly acting as intermediaries in our purchasing and cultural consumption decisions. This paper introduces ChoiceEval, a novel framework for auditing LLMs for "entity-perception bias" – their systematic favoritism towards certain brands or cultures. The study reveals that major LLMs exhibit stable, US-centric preferences, raising concerns about market fairness and cultural homogenization.

What Problem Does This Paper Address?

As LLMs like ChatGPT and Gemini become primary interfaces for information discovery, they are fundamentally changing how we interact with brands and cultural offerings. This shift creates an urgent need to understand the systemic risks associated with LLM-driven market intermediation. Specifically, this paper addresses the gap in auditing LLMs for "entity-perception bias," a form of bias where models systematically favor certain entities (brands, cultures) over comparable alternatives, potentially skewing market fairness, competition, and information diversity. Existing research has largely focused on social biases (gender, race, etc.), leaving this crucial aspect of economic and cultural influence under-examined.

Methods

The core contribution is ChoiceEval, a reproducible framework designed to audit entity-perception bias in LLMs. It tackles two key challenges:

  1. Realistic, Persona-Diverse Evaluation Queries: ChoiceEval segments users into psychographic profiles (e.g., budget-conscious, wellness-focused) based on established consumer segmentation research (like VALS). These profiles are then adapted to specific topics (e.g., running shoes, hotel chains) to generate realistic, consideration-phase prompts that mimic how users actually seek advice.
  2. Converting Free-Form Outputs to Quantitative Metrics: LLM responses are processed to extract normalized top-k choice sets. This involves a multi-expert extraction process using an LLM with stochastic decoding to simulate independent annotators, followed by majority voting to ensure robustness. These extracted entities are then analyzed for preference concentration, rigidity, and geographic bias using metrics like Preferred Entity Inclusion Rate (PEIR) and Log Odds Ratios (LOR).

The framework is designed to be model-agnostic and topic-agnostic, allowing for broad applicability.

Datasets

The paper introduces a new benchmark dataset generated using the ChoiceEval framework. This dataset comprises:

  • 10 topics spanning commercial (e.g., hotel chains, electric vehicles, laptops) and cultural (e.g., travel destinations, wine regions) domains.
  • Over 2,000 questions in total, with 23 questions per topic and user cluster pair.
  • User clusters derived from psychographic segmentation (VALS).

This dataset is made publicly available.

Key Insights and Ideas

  • LLMs are not neutral: The study strongly suggests that LLMs exhibit stable, inherent preferences, acting more like biased intermediaries than objective information providers.
  • The "US-centric" bias is pervasive: American-developed LLMs (Gemini, GPT) show a marked favoritism towards US entities across most commercial and cultural topics, often exceeding what real-world market data would suggest.
  • Systemic, not incidental bias: These preferences persist across different user personas and queries, indicating systematic algorithmic predispositions rather than random noise.
  • Implications for market dynamics: This bias can create artificial competitive advantages, reinforce market concentration, and potentially lead to cultural homogenization by amplifying the visibility of certain entities.
  • DeepSeek shows a less pronounced bias: While still exhibiting some US tilt, the Chinese-developed DeepSeek model demonstrates that alternative development methodologies and data curation can mitigate this geographic bias.

Key Results

  • Stable Preferences: All three tested LLMs (Gemini, GPT, DeepSeek) demonstrated stable preferences for specific entities across topics, with high correlation coefficients (Spearman's Rank Correlation > 0.952 for Laptops).
  • Geographic Bias:
    • Gemini and GPT showed statistically significant US preferences in 7 out of 8 entity-anchored topics when compared to Europe, and in all topics when compared to Asia.
  • For example, in the Weekend Getaway Cities topic, GPT's top ten recommendations featured nine US cities, while Gemini's consisted exclusively of US cities. This contrasts sharply with real-world tourism data.
  • DeepSeek exhibited a less pronounced US bias, with a significant tilt towards European cities in the Weekend Getaway Cities topic.
  • Entity Inclusion: Averaging across categories, Gemini, GPT, and DeepSeek included their top-ranked entity in 71.2%, 75.9%, and 70.3% of responses, respectively.

Limitations and Future Work

  • Limitations:
    • The study focused on three specific LLMs and a limited set of topics.
  • The geographic classification was simplified to US, Europe, and Asia due to data sparsity for other regions.
  • The study was conducted in English and without personalization, which are common in real-world LLM interactions.
  • Future Work:
    • Explore regional, linguistic, and personalization factors on LLM recommendations.
  • Conduct longitudinal studies to track how LLM biases evolve over time with updates and user feedback.
  • Investigate how different behavioral patterns of LLMs (e.g., direct recommendations vs. nuanced responses) affect user decisions.
  • Further research into the underlying factors driving these preferences, such as training data composition and semantic embedding structures.

Summary updated

BibTeX
@misc{rienecker2026auditing,
  author = {Jasmine Rienecker and Katarina Mpofu and Naman Goel and Siddhartha Datta and Jun Zhao and Oscar Danielsson and Fredrik Thorsen},
  title = {{Auditing Preferences for Brands and Cultures in LLMs}},
  year = {2026},
  eprint = {2603.18300},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.18300},
}
Figure