Does a shopping algorithm's design decide who wins: you, the seller, or the platform running the show?

This explores whether a recommendation algorithm or AI shopping agent leaves people better off depending on whose gains it is built to look after: buyers, sellers, the platform, or everyone at once. The corpus says yes, though in a less obvious way than you might expect. Economic models suggest an algorithm's design matters most when there are few sellers. Agent experiments suggest that one sentence about who the agent works for can change what it recommends.

Start with the economics. One formal model shows that a recommender that never handles money can still match the outcome of an ideal market mechanism. It does this by committing to a fixed rule in advance and treating sellers' posted prices as signals of their real costs Can an algorithm recommend products without matching an ideal mechanism?. The catch is that this edge comes from the algorithm sometimes withholding trades, which limits how much sellers can profit from knowing more than buyers. As more sellers enter, ordinary price competition does that job anyway, and the algorithm's advantage shrinks toward nothing Does algorithmic recommendation lose its edge as sellers compete more?. So the question of whose surplus the algorithm protects matters most in thin markets, where a few sellers could otherwise capture the gains.

AI agents bring the same question down to ground level. Azhar argues that an agent paid through merchant referral fees cannot be fully loyal to both the merchant and the user. On this view, the conflict comes from who pays the agent, not from any limit in the technology Can an AI agent serve both merchant and user interests fairly?. An experiment shows how fragile that loyalty is and how cheap it can be to fix. Agents employed by a platform went easier on penalizing sponsored products. Adding one sentence to the prompt that named the consumer as the beneficiary removed that bias, even though the platform still employed the agent Can naming the consumer fix AI agent sponsorship bias?. In other words, the agent's sense of who it is working for partly decides whose surplus it protects, and that sense can be set in a few words.

Naming a beneficiary does not guarantee good outcomes, though. In a simulated marketplace, frontier LLM agents reached near-optimal welfare only when search conditions were ideal. Even then, they tended to accept the first offer they received and weighted speed over price by 10 to 30 times Do LLM agents achieve fair marketplace outcomes when given ideal search?. An agent loyal to you but naive about offers can still leave your money on the table. As agents take over shopping, sellers will compete for the agent's attention rather than yours, so the agent's built-in priorities become what sellers optimize against Will agents compete for attention just like users do?.

The last twist is that protecting a single user may be the wrong goal at scale. One line of alignment research argues that serving one user loyally is not enough once millions of agents interact. It calls for agents that balance many parties' interests, with transparency about how they do it and clear liability for developers and deployers Should AI agents serve one user or balance many interests?. Taken together, these notes say the welfare question has three layers: market thickness (does design even matter here?), the agent's stated principal (whose side is it on?), and the agent's competence (can it actually get the deal?). The corpus has strong formal results on the first layer. It has early but striking evidence on the other two.