If an AI agent does your shopping, companies may end up competing to impress the bot, not you.

This explores whether the fight for human attention (clicks, ad placements, search rankings) gives way to a fight to be picked by an AI agent acting for the user. The collection's most direct answer is yes, at least as a trajectory: as people hand goals to agents, services have to win the agent's selection rather than a person's glance, and the research expects that to produce agent-facing versions of the familiar machinery: discovery systems, rankings, and recommendation infrastructure that mirror today's ad ecosystem Will agents compete for attention just like users do?. So the claim is not that attention competition disappears. It moves one layer down and changes who it's aimed at.

What does the agent actually see when it chooses? Increasingly, not a web page. Agents that call APIs directly instead of clicking through interfaces finish tasks 65–70% faster at nearly the same accuracy, and some systems can automatically build APIs out of existing apps Can API-first agents outperform UI-based agent interaction?. The consequence is easy to miss: if agents favor clean, callable interfaces, then the thing a business competes on stops being visual design and persuasive copy and becomes how well its service can be read and invoked by a machine. In this market, a polished landing page counts for less than a well-documented endpoint.

There's a stranger twist. Agents may not consult outside services at all when they can answer from what they already know. Researchers have trained agents against search engines that are simulated entirely from a model's internal knowledge, and a 14B-parameter simulator matched or beat real search engines for that purpose Can LLMs replace search engines during agent training?. That points to a competitor no click-based economy ever had: the agent's own memory. A service doesn't only have to beat rival services. It has to be worth calling at all compared with the agent answering for itself.

The human doesn't vanish from the loop either. Work drawing on conversation analysis argues that agents drift away from what the user meant when they silently chain tool calls, and that they should stop and ask the user to clarify or narrow the request When should AI agents ask users instead of just searching?. Those check-in moments are where human preference re-enters, so a realistic picture is a hybrid: agents filter and shortlist, and people weigh in at a few decision points. Cost pressure also shapes which agents do the choosing. If cheap small models handle most routine agent steps Can small language models handle most agent tasks?, the selection decisions businesses compete over may often be made by lightweight models with their own biases, not by frontier systems.

A direct caveat: the collection has one note that addresses this question head-on. The rest are adjacent pieces about how agents call tools, where their information comes from, and when they defer to users. The corpus does not yet have empirical work on how agents rank competing services, or on whether agent-targeted manipulation (the agent equivalent of SEO spam) is already happening.