Whose side is the AI shopping assistant on? Change its role and the penalty for sponsored results disappears

Key numbers

• Controlled choice experiments varied only the assigned role in the system prompt (naming either a traveller or a booking platform as the agent's principal) and compared how the model handled sponsored listings and reasoned about them.
• When the platform was the principal, the penalty the model applied to sponsored listings dropped significantly, and the scepticism that disclosure normally triggers in the reasoning trace weakened with it — replicated across models and reasoning depths.
• A second study decomposing the disclosure label found the divergence between the two principals widened significantly when the paid placement was attributed to the platform.
• Stricter wording (“Sponsored” instead of “Promoted”) lowered selection of paid listings, but did not close the gap once the platform was named as principal.

Method and sample

arXiv paper “Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders” (2026-09-16): controlled choice experiments manipulating the agent's assigned principal in the system prompt, with the model's choices and reasoning traces compared across models and reasoning depths, plus a second study isolating the disclosure label.

Limits

The experiments run under controlled prompt conditions rather than measuring live shopping assistants; how advertising and ranking are coupled in production is not public, so real-world magnitudes remain unknown.

Related research