Headline stats about AI shopping can hide who really wins or loses underneath the average.

This explores how big-picture numbers about AI commerce can hide who actually gains and who pays. The corpus points to a recurring pattern: the total looks like one story, and splitting it apart shows a different one. Sometimes the people paying aren't in the measurement at all. Sometimes the harm only shows up when two causes are added together. And sometimes the cause gets credited to the wrong thing.

Start with the people left out of the frame. Agarwal and Sen's field experiment found that Google's AI Overviews cut clicks to publisher sites by 39.8% without making users any more satisfied or improving click quality Do AI Overviews actually reduce clicks to publisher sites?. A study that only measured users would conclude that nothing changed. The real cost landed on publishers, who were never part of the user sample. BCG's survey has a similar shape at a larger scale: about $1.3 trillion in spending shifting toward agent-assisted shopping sounds like a single market shift, but the same analysis says consumers and platform networks take most of the gains while merchants face a structural disadvantage Who wins and loses as AI agents reshape shopping?. One growth figure hides a transfer from one group to another.

Averages can also point the wrong way about who loses. The popular story is that AI agents will cut out travel middlemen. Goswami's data shows the reverse for one vulnerable group: independent hotels became more dependent on online travel agencies (63.4% of bookings), because the agents connect through those agencies' systems rather than going around them Are AI agents actually removing online travel agencies from bookings?. Geography works the same way. Chainalysis found that agent-to-agent crypto payments look like one global market, but US wallets made 88.1% of AI model inference payments, while Chinese wallets clustered in security services Which countries dominate agent-to-agent AI payments on chain?. Even the value of an algorithm depends on the setting: a formal model shows that an optimal recommender's advantage over ordinary price competition shrinks to almost nothing as more sellers enter. Its edge exists mainly in thin markets Does algorithmic recommendation lose its edge as sellers compete more?. Pool thin and thick markets together and you misjudge both.

The less obvious problem is that measuring one cause at a time can undercount the harm. In a multi-agent simulation, a platform tilting its rankings and sellers manipulating product descriptions each harmed AI shoppers. Together, the harm was more than double what either caused alone, with no coordination, because both exploited the same agent biases Do platforms and sellers exploit AI shoppers together more effectively?. An audit of platforms alone, or of sellers alone, would miss most of the damage. Attribution can also go wrong: Magentic Marketplace traced lost welfare to agents making mistakes in complex negotiations, not to agents serving someone else's interests Do agent mistakes or misaligned incentives cause marketplace welfare loss?. Someone looking only at the total loss could blame the wrong cause and choose the wrong fix.

One more twist: the yardstick itself may move. If shopping agents help shape what consumers want through repeated back-and-forth Do AI shopping agents reshape what consumers actually want?, then 'consumer satisfaction' is partly measured against preferences the agent helped create. That echoes a broader warning that agent metrics track what is easy to score rather than real economic value Why do agent benchmarks not predict real economic value?. Taken together, the corpus suggests asking three questions of any AI commerce number: who was left out of the sample, which harms were measured separately when they should have been measured together, and whose preferences the outcome is being judged against.