When an AI shopping assistant charges to 'inspect' a seller, who really ends up paying for that search?

This explores what happens when an AI platform sits between buyers and sellers and charges for each 'inspection' (one close look at a seller's offer) while sellers can pay part of that cost to attract attention. The short answer from the corpus is that a profit-seeking platform sets the token price so that buyers look at more sellers than they need to. Buyers come out no better and no worse, but money moves from sellers to the platform, and the market as a whole becomes less efficient Does platform pricing of inspection tokens encourage excess search?. The surprising part is who pays. Consumers don't notice any harm, so the cost of the extra searching is easy to miss.

The subsidy side explains why sellers keep paying anyway. In the same kind of model, seller subsidies sort sellers by quality. The best sellers all offer the maximum subsidy and become indistinguishable at that ceiling. Middle-quality sellers offer gradually larger subsidies to stand out, and low-quality sellers never get inspected How do seller subsidies sort quality in search markets?. That sorting is useful: it reveals a lot about quality without an auction. It also means sellers are competing to fund buyers' attention, and once the platform controls the token price, it can capture that competition as revenue. A mechanism that reveals information well can still pull surplus away from the sellers who fund it.

The formal result matches a familiar argument from outside economics. Doctorow's 'enshittification' lifecycle describes platforms that first serve users, then squeeze business customers, then extract value for shareholders Do platforms inevitably decline through value extraction cycles?. The inspection-token model looks like a precise version of that middle stage: buyers are left alone while sellers are drained. Doctorow's evidence is illustrative rather than systematic, but the two sources point in the same direction from very different starting points.

It is also worth asking whether AI agents would search the way the model assumes. Microsoft's Magentic Marketplace experiments found that LLM agents get close to ideal outcomes only when search conditions are ideal. They also show a strong bias toward the first offer they receive, favoring speed over quality by 10–30x Do LLM agents achieve fair marketplace outcomes when given ideal search?. If real agents settle early instead of over-inspecting, the platform's pricing lever may work differently in practice than in theory. That gap is open, and the corpus doesn't resolve it.

The material here is thin: two notes describe the core model directly, and the rest is adjacent. The two direct notes alone show that AI shopping intermediaries can profit simply by setting the price of looking. Buyers wouldn't see that cost; it shows up in sellers' margins.