As AI makes searching free, the hard part of shopping may shift from finding options to knowing what you actually want.

This explores whether cheap, agent-driven search moves the hard part of shopping from finding options to saying what you want. The corpus's most direct answer is yes. A formal model shows that manual search breaks down as products get more complex, because people can't browse enough options to compare all the features. Agentic search avoids that breakdown by shifting effort from browsing to refining what the buyer is asking for. In that model, the limiting factor is no longer how big the catalog is. It is how costly it is for the buyer to articulate their preferences Why does manual search fail as product choices become complex?. Once the agent can scan everything, a vague request becomes the weak link.

If articulation is now the scarce resource, the next question is what fills the gap when people can't supply it. Behavioral research on annotation gives a useful warning. Many stated preferences aren't real, stable preferences. Some are 'non-attitudes', meaning answers given without any underlying opinion. Others are 'constructed' on the spot and shift with how the question is asked Do all annotation responses measure the same underlying thing?. So scarce articulation isn't only a problem of effort. Often there is no settled preference to report. Bilgihan and colleagues push this further for shopping agents that actually complete purchases. They argue preferences aren't formed beforehand and then handed to the agent. Instead, preferences take shape through back-and-forth loops between what the agent learns and how much the consumer trusts it Do AI shopping agents reshape what consumers actually want?. In that view, cheap search doesn't just reveal how little people articulate. It turns articulation into a joint product of person and agent.

There are tools for lowering the cost of stating preferences. LLMs can turn offhand complaints like 'doesn't look good for a date' into positive preferences that a search system can use, such as 'prefer more romantic' Can language models bridge the gap between critique and preference?. Personalization research finds that short summaries of what someone likes work better than retrieving their specific past interactions Does abstract preference knowledge outperform specific interaction recall?. Together these suggest that the practical fix is to let the agent infer and compress preferences rather than ask people to spell them out. That fix has a cost: the agent's own interpretation becomes part of the preference.

The main risk is that defaults take over when articulation is thin. In 15,000 simulations, six LLMs recommended the same side of every strategic choice they were tested on. Option order changed the outcome more than the business context did Do LLMs consistently favor the same strategic choices regardless of context?. In simulated marketplaces, agents showed a strong bias toward the first offer they received, favoring speed of response over offer quality Do LLM agents achieve fair marketplace outcomes when given ideal search?. When the buyer under-specifies, these built-in tendencies decide instead.

This connects to market-level findings that cheap search doesn't automatically help shoppers. One 2026 market design paper argues that imperfect agent search, easy-to-generate look-alike product descriptions, and entry by low-quality sellers can combine to leave markets worse off than before Can cheap agent search actually harm shopping markets?. Another model shows that a platform can price inspection in ways that push buyers to over-search, moving surplus from sellers to the platform Does platform pricing of inspection tokens encourage excess search?. The takeaway is that falling search costs make articulation the bottleneck. Whoever helps buyers articulate, whether the agent, the platform or the model's defaults, gains real power over what buyers end up wanting.