Market Shares under Delegation
Model heterogeneity drives different product shares; demand can concentrate.
Research Highlights
How ACES - Agentic e‑CommercE Simulator- Works
It starts with a simple prompt. We provide an AI agent with a shopping task, like "find the best fitness watch".
AI Model Preferences Analysis
AI Model Preferences: Fitness Watches
See how different AI shopping agents choose completely different products
🟠 Claude Sonnet 4 strongly prefers Fitbit Inspire (45.2%)
Fitbit Inspire 3
$79.95
👑
Fitbit Versa 4
$149.95
Garmin Forerunner 55
$166.40
Garmin vívoactive 5
$220.89
Smart Watch (Generic 1)
$24.99
Smart Watch (Generic 2)
$29.99
WHOOP 5.0
$359.00
Generic Fitness Tracker
$28.99
Performance Heatmap Analysis
Webpage Heatmap of AI Shopping Agents
Selection rate of the same product across different positions for three AI agents.
Position 1
11.5%
Position 2
27.0%
Position 3
23.4%
Position 4
15.4%
Position 5
3.4%
Position 6
7.9%
Position 7
6.9%
Position 8
4.5%
Feature Impact Analysis
How Product Features Influence AI Agent Selection
Explore how different product attributes affect AI agent behavior across models and providers.
OpenAI
Google
Anthropic
Being named "Overall Pick" has the largest impact — Gemini 2.5 Pro jumps to 64.5% selection probability (+54.5 p.p. over baseline).
Key Research Takeaways
Key Takeaways
Agentic e-Commerce: Market Shares, Platform Levers, Seller Strategy
Platform Levers and Position Effects
Strong top-row bias; rankings, promotions, and sponsorships affect models differently.
Seller-Side Listing Agents Move Share
Minor, model-targeted listing edits can materially shift market share.
Team
Citation
@article{allouah2025aces,
title={What is your AI Agent buying? Evaluation, Implications and Emerging Questions for Agentic e-Commerce},
author={Allouah, Amine and Besbes, Omar and Figueroa, Josué D and Kanoria, Yash and Kumar, Akshit},
journal={arXiv preprint arXiv:2508.02630},
year={2025}
}