Variation of AI-mediated commercial advice across languages, locations, and time

Determine how the observed patterns in AI-mediated commercial advice vary across languages, locations, and time, and how they relate to specific consumer-protection and Digital Services Act obligations.

Background

The paper audits ChatGPT, Gemini, and Google AI Overviews using real product-recommendation queries and finds substantial variation in recommended products, recommendation framing, displayed sources, and the correspondence between consumer interfaces and provider APIs. The audit is limited to predominantly English-language queries, responses collected from the Netherlands, and a single time period, so it does not establish whether the observed patterns generalize across linguistic, geographic, or temporal contexts.

The paper also frames these systems within European consumer-protection and Digital Services Act requirements. Further investigation is therefore needed to determine how the empirical patterns change across contexts and how they map onto specific regulatory obligations.

References

Future work should test how these patterns vary across languages, locations, and time, and how they relate to specific consumer-protection and DSA obligations.

— "If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations  (2609.18729 - Marin et al., 16 Sep 2026) in Conclusion

It is very unclear if users of LLM agents today are aware of such preferences and if they potentially constitute unwanted/harmful biases on the platforms.

— Characterizing Web Search by Conversational LLM Agents: From Search Decisions and Strategies to Results and Responses  (2609.19244 - Amani et al., 16 Sep 2026) in Section 3, Subsection "Domain Preferences in Search Results" (summary paragraph)