What the Agentic Customer Changes About the Market
The central question is no longer whether software can help someone shop, but what happens when that software becomes a recurring intermediary between people, firms, and platforms. This final chapter connects the evidence about personalized curation and agent-ready websites with a market model of search, learning, pricing, and feedback. It gives the reader a conditional way to judge an AI customer: not only by whether it finds a suitable option now, but by what its information flow makes possible—or impossible—later.
The distinction matters because commerce has a memory. A recommendation changes one person's choice, but the search and transaction may also produce signals that shape what future customers see. Those signals can improve collective knowledge, narrow competition, preserve a useful business, or prematurely bury one. The same agentic capability can therefore look beneficial at the level of a single task and ambiguous at the level of a market.
Search creates a market memory
The market model in Agentic Markets makes this long-term effect concrete. It describes a stream of consumers choosing among businesses. Before buying, a consumer can pay to inspect whether a business fits their particular need; quality is learned through the transaction. The platform records inspection outcomes and post-transaction feedback, then exposes the accumulated history to later consumers. In this setting, search is not merely a private cost. It is also a process that determines which businesses receive enough attention to become understood.
A business can follow one of two long-run paths. It can be learned, meaning that repeated observations allow market beliefs about its fit and quality to approach the underlying parameters. Or it can be lost: beliefs fall far enough that no later consumer has a positive reason to inspect it, so its information stops updating. The lost business need not be universally bad. It may be a good match for a narrower group, but the platform may see only a series of failed inspections and fail to distinguish an idiosyncratic mismatch from a broadly relevant deficiency.
This mechanism gives a different meaning to exploration. An agent that stops after the first plausible result may save effort today while reducing the chance that other businesses are ever examined. An agent that searches more widely can create more opportunities for learning, but only if the resulting observations retain enough context to be interpreted. The market's future option set is partly produced by the sequence of decisions that earlier agents made under incomplete information.
The model distinguishes two ways to improve search. The first is to make it cheaper. Lower search cost encourages consumers to inspect more businesses, and the paper proves within its model that the learned set weakly expands and long-run consumer utility weakly rises. The direct effect is obvious: the current consumer pays less to investigate. The less obvious effect is that more businesses remain visible to future consumers instead of crossing the threshold into permanent neglect.
The second improvement is to make search more informative before purchase. This can help the current consumer by revealing more selective fit information. But if the platform observes only whether a business passed or failed, richer screening can make collective learning worse. A failure caused by a consumer-specific requirement—say, a niche dietary need—can look identical to a failure caused by a deficiency that matters to almost everyone. As screening becomes more selective, even broadly useful businesses may accumulate enough failed inspections to be abandoned.
This is the counterintuitive point: more information for the individual is not automatically more information for the market. The difference lies in how correlated the signal is across consumers and how much of its explanation survives aggregation. A pass/fail record answers what happened for one person.