Persona Conditioning of Brand Recommendations in Re... | AI Research

Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat: A Prominence-Stratified Cross-Provider Audit
When you ask an AI for a recommendation, the answer you receive is shaped by more than just your prompt; it is also influenced by the AI’s understanding of who you are. This research investigates how much a user’s persona—such as being a solo founder versus an enterprise executive—changes the brands an AI recommends. By auditing 2,000 chat interactions, the authors demonstrate that AI recommendations are not persona-agnostic; instead, they shift significantly based on the context provided, suggesting that AI-driven brand perception is highly sensitive to the buyer's identity.

How the Audit Was Conducted

The researchers tested three different AI model configurations using a set of 10 distinct personas and 8 common commercial prompts. To isolate the effect of the persona, they kept the prompt, temperature, and system instructions constant, only changing the persona description prefixed to the user's message. They measured the "Jaccard similarity"—a way of calculating how much the list of recommended brands overlaps—to see how often the AI swapped its suggestions when the persona changed. They also categorized brands by their market prominence (from category leaders to regional players) to see if some brands are more "persona-resistant" than others.

Key Findings: The Mid-Market Shift

The audit revealed that changing a persona causes a noticeable shift in recommendations, with similarity scores dropping by 12% to 20%. This effect is not uniform across all brands. Category leaders (the most famous brands) are largely "persona-resistant," meaning they are recommended regardless of who is asking. In contrast, mid-market brands are highly sensitive to persona, with up to 75% of the recommendation set changing depending on the user's context. The researchers also observed that the Anthropic model showed a larger shift in recommendations compared to the OpenAI models, which may be linked to how much the models rely on their internal training data versus external search results.

Why This Matters for AI Recommendations

The findings challenge the idea that AI assistants provide a neutral, objective list of "best" products. Instead, the study suggests that AI models are already performing a form of de facto market segmentation. Because the models lean on their internal "priors"—the patterns they learned during training—they naturally tailor their suggestions to fit the persona they perceive. This means that if a brand wants to understand how it is perceived by AI, it cannot rely on a single, aggregate measurement. Instead, brand perception must be analyzed through the lens of specific buyer personas, as the "best" recommendation is highly dependent on who the AI thinks is asking.

Important Considerations

While the results show a clear impact of persona on AI output, the researchers note a few limitations. The study focused on a specific set of models and a hand-curated list of personas, meaning the results might not apply to every AI assistant available. Additionally, while the Anthropic model showed a larger point-estimate effect, the statistical confidence intervals for different models overlap in some areas, meaning the exact ranking of which model is "most" responsive to personas is not definitive. Finally, the researchers emphasize that this study measures the distribution of recommendations rather than the accuracy of the advice, as there is no single "correct" answer in commercial recommendations.