A buying assistant doesn't need to rig their ranking: just choose which question they ask

A soft commercial deposit does not change anything. A target deposit based on the sponsor's advantage increases the proportion of recommended sponsored products by 0.30.

The hypothesis tested is a blind spot in commercial assistant audits: a financial incentive does not need to enter the final ranking algorithm to weigh on the recommendation, it can simply influence the question of preference that the assistant chooses to pose. The authors make this distinction observable in a deliberately reduced synthetic device.

Each task has two products, three numerical attributes verified, a price limit and a private preference vector. An honest simulated user answers a single comparative question; a separate recommendation module receives the products and response, but not the sponsor information. Three regimes are compared: neutral question, soft commercial order, explicitly adversarial order asking to question the advantage of the sponsor by ignoring that of the competitor.

Of 40 test cases, the soft set does not change any selection. The target set increases the share of selected sponsored products by 0.30 and reduces the average synthetic utility by 0.0547 compared to neutral questioning for a given language model. A consistency judge evaluates the 20 targeted responses sampled as consistent — five with a regret of over 0.05; Only a separate dimension, dealing with the coverage of issues, indicates the unilateral nature of the solicitation. The authors explicitly limit their conclusion: the device does not establish typical advertising behaviour or an effect on real consumers.