← Back to Research

When the Buyer Is a Machine: The Diagnosticity-Exploration Tradeoff in Agentic Commerce

Lee, S., D. Shin, & S. P. Han
Accepted at ICIS 2026 (Lisbon) and INFORMS CIST 2026 (San Francisco)
In preparation for Information Systems Research
Working Paper

Overview

Consumers are beginning to hand purchase decisions to AI agents that search, compare, and choose on their behalf. Marketing was built around human decision-making, so the question is whether it still works once the buyer is a machine. We find that it does, but not in the way it works on people.

The cues that move an agent are not the cues that move a person. Agents follow aggregated social proof and largely ignore the scarcity and urgency appeals that drive human buyers. We argue the reason is architectural rather than psychological: an effective cue shortens the agent's search rather than persuading it.

Experimental Design

Sandboxed shopping environment

  • A replica of a live e-commerce interface, built so that every page element can be controlled
  • Fictional brands with matched attributes, so brand familiarity cannot drive choice
  • Price, rating, and display position randomized; marketing cues manipulated alone
  • Roughly 23,000 agent purchases collected across several commercial models

Categories

  • Four product categories spanning durables and replenishables
  • Eight candidate products per choice set
  • Model version, temperature, and seed fixed for reproducibility

What We Find

🔍

Agents barely compare

With eight candidates on screen, agents typically open one or two before deciding.

📊

Cues are not equal

Aggregated social proof such as a best-seller badge moves choice sharply, while scarcity and urgency do not.

🔒

Forcing search barely helps

Requiring the agent to open every candidate leaves most of the badge effect intact.

🧭

A different reference point

People weigh alternatives against each other; agents judge each candidate against the request itself.

Interpretation

Why aggregated signals dominate

When each candidate is judged on its own against a stated request rather than against rival options, only cues that already carry ranking information inside them remain evaluable. A best-seller badge encodes what other buyers chose; a countdown timer does not. The gap is structural, which is why influence survives delegation but changes shape: it shifts from the buyer's emotions to the agent's search.

Implications

Sellers

Cue portfolios tuned for human shoppers lose much of their effect once agents mediate purchases.

Platforms

Ranking and badge systems carry more weight than intended when agents rely on them as the main evaluable signal.

Agent designers

Shallow search is not only a cost decision; it determines which signals can influence the outcome at all.

Policy

Consumer protection built around persuasion of people may not reach the mechanism that operates here.

Keywords

AI Agents Agentic Commerce Algorithmic Delegation Marketing Cues Consumer Search Discrete Choice