An AI shopping agent’s employer can change how it treats sponsored listings

Changing the employer, not the sales instructions

Davood Wadi and Yu Ma tested a specific question: does an AI shopping agent become more accommodating toward sponsored listings when its instructions name a booking platform, rather than a traveler, as its employer?

Their experiment did not tell the agent to maximize advertising revenue or favor sponsors. Instead, the central manipulation changed the employing party in a sentence beginning, “You have been employed by…” The authors call the resulting shift “sponsorship bias”: a more favorable evaluation of sponsorship under platform employment than under consumer employment.

That definition matters. Bias here does not necessarily mean choosing an advertisement over an otherwise preferable product. It means changing how much sponsorship counts against a listing when the assigned employer changes.

A controlled hotel comparison

The first study covered six generative large language models, which produced hotel selections, and six non-generative decision models, which returned probabilities over the available options.

Each task contained four hotels. The two competitive options had equal prices and positive reviews; two less attractive alternatives had higher prices and negative reviews. The researchers varied whether one competitive listing carried a sponsorship flag and whether the agent worked for the traveler or the platform. Hotel identities and review texts varied, with statistical controls accounting for their influence.

For each generative model, the researchers collected 500 selections in each of four conditions. For each decision model, they evaluated 360 position-balanced choice sets under all four conditions.

Across all 12 models, sponsorship reduced the target listing’s evaluation under consumer employment. Platform employment significantly weakened that penalty, although its size differed substantially between models.

For Gemini 3.1 Pro, the target hotel’s selection share fell from 53.0% without sponsorship to 2.8% with sponsorship when the agent worked for the traveler. Under platform employment, it fell from 54.2% to 25.0%. The platform-employed agent still penalized sponsorship, but considerably less.

Kev 9B showed a smaller effect with a different endpoint. Its average probability assigned to the target fell from 43.43% to 40.60% under consumer employment, but rose from 42.62% to 43.88% under platform employment. These are model-assigned probabilities, not observed bookings or repeated selection frequencies.

The central pattern also survived alternative role wording and a change from asking which hotel to “book” to which to “recommend.” In Gemini 3.1 Pro, it appeared at both tested reasoning-effort settings.

Clearer labels helped, but did not erase the role effect

A second study, using Gemini 3.1 Pro and Kev 9B, separated two features of disclosure: whether a listing was called “Promoted” or “Sponsored,” and whether the label added “by the platform.”

“Sponsored” lowered the listing’s evaluation relative to “Promoted.” But explicitly connecting the placement to the platform increased the gap between consumer- and platform-employed agents under both labels. Clearer advertising terminology therefore did not eliminate the role-dependent difference in these tests.

The researchers also scored skepticism expressed in several generative models’ pre-choice reasoning traces. Sponsored listings elicited less expressed skepticism under platform employment, and greater skepticism was associated with lower selection likelihood. This is evidence about generated explanations, not direct access to neural processing or proof that skepticism caused the choice.

Two ways to counter the shift

Adding “Act in the traveler’s best interest” to the platform instruction removed its relative leniency toward sponsorship in both models tested. However, Kev 9B then penalized sponsorship more strongly than the consumer-employed baseline. The instruction also changed estimated preferences associated with hotel names and reviews.

The authors next fine-tuned Kev 9B to reproduce its original consumer-employed probability distributions regardless of the assigned employer. Training used synthetic laptops, headphones, coffee makers and running shoes; a separate air-purifier category guided checkpoint selection. Hotels were excluded from training.

On the hotel benchmark, the fine-tuned model showed no statistically detectable remaining role-by-sponsorship interaction and preserved measured non-sponsorship preferences more closely than the prompt remedy. This result concerns one fine-tuned decision model, not a demonstrated universal fix.

What the benchmark cannot settle

These controlled hotel tasks did not measure real purchases or compare agents with human shoppers using identical interfaces. Equal prices and similarly positive reviews helped isolate sponsorship’s effect, but leave unanswered how agents trade sponsorship against a genuine discount or better quality.

Crucially, the mitigation target was the model’s own consumer-employed behavior—not independently established consumer welfare. The preprint shows that employer wording can change recommendations without an explicit advertising objective. It does not establish that the consumer-employed baseline always makes the best choice.