Self-Promotion in LLM Recommendations

Document Type

Conference Proceeding

Role

Contributor

Published In

WebSci '26: Proceedings of the 18th ACM Web Science Conference 2026

Publisher

Association for Computing Machinery

First Page

310

Last Page

319

Publication Date

5-25-2026

Abstract

Large language models (LLMs) are used across diverse applications, creating complex choices for users selecting among competing models for varying use cases. When users turn to AI systems themselves for guidance, do AI models exhibit self-promotional bias when recommending competitors? We conducted an audit of three major providers for self-promotion in LLM recommendations: OpenAI, Anthropic, and Google. Using prompts spanning coding, mathematics, scientific reasoning, general knowledge, and operational domains, we observed how each company ranks itself and its competitors. We compare the collected rankings to performance benchmarks to separate legitimate benchmark-based suggestions from promotional bias. Our analysis reveals self-promotional bias across all providers. Models consistently rank their own companies’ products 0.2 positions higher than their benchmark performance would justify. Vendor-specific analysis shows substantial variation, with OpenAI exhibiting the strongest promotional tendencies. These findings raise important questions about transparency and fair competition as AI recommendation systems increasingly influence technology adoption decisions.

Link to Full Text

Share

COinS