Once an algorithm decides who gets seen, sellers stop competing on price alone and start gaming the recommender itself.

This explores how sellers change their behavior once a recommendation algorithm, rather than a shopper browsing freely, decides which products get seen and bought. The corpus has two separate strands on this. One is formal economic modeling of pricing. The other is newer simulation work on sellers gaming AI shopping agents. Read together, they suggest the seller's main lever is moving from price to persuasion.

Start with price. In the economic models, sellers respond to a recommender by treating their posted price as a signal. If the algorithm commits to a fixed rule before sellers set prices, each seller's price ends up revealing its underlying costs. With that, a recommender that never handles any money can match the welfare of an ideal market design (Can an algorithm recommend products without matching an ideal mechanism?). The commitment is what makes it work: sellers can't game a rule that's already fixed. That edge has limits, though. As more sellers compete, ordinary price competition does nearly as well, and the algorithm's advantage shrinks toward nothing. It matters mainly in thin markets, where it can hold back trades to keep sellers from extracting extra profit (Does algorithmic recommendation lose its edge as sellers compete more?).

The less obvious finding is that when the buyer is an AI agent, sellers have a cheaper lever than price: the words in the listing. In one experiment, a seller-side AI rewrote only its product descriptions, using competitor data. It gained 3.66 to 14.89 percentage points of market share against five of six buyer models, with no change to price, rating, or position (Can AI seller agents shift buyer preferences through description edits?). This works because AI shoppers are far easier to sway than people. Given identical price, rating, and nudge cues, agents shift their choices much more than humans and fall into rigid, rule-like patterns, while humans shift modestly (Do shopping agents respond more strongly to price and rating cues?). A seller facing a predictable buyer has every reason to optimize for that predictability.

The platform and its sellers can end up exploiting the agent together without ever coordinating. When the platform tunes its ranking and sellers tune their descriptions, each aiming at the same agent biases, harm to consumers more than doubles compared with either acting alone (Can platform ranking and seller descriptions separately harm AI shoppers?, Do platforms and sellers exploit AI shoppers together more effectively?). That's tacit collusion with no agreement behind it. The platform's own funding model can push in the same direction: an agent paid for by advertising has a built-in reason to favor paid placements, however well it's engineered (Do advertising-funded shopping agents favor paid product placements?). The broader feed research makes a related point. How a feed weights content shapes what producers choose to make, so recommendation design ends up steering the supply side as well as what buyers see (How do recommendation feeds shape what people see and believe?).

The gap: the corpus has little empirical data on how real human sellers adapt over time, such as A/B-tested listings, keyword stuffing, or dropping out of niche categories. Most of the evidence comes from formal models and LLM simulations. If you want the next question to pull on, it's this: the pricing models assume the algorithm commits to a fixed rule that sellers can't game, while the agent studies show sellers easily gaming the buyer's language model itself. How to get that kind of commitment from an LLM shopper is still an open question.