Always-on Commerce Growth Agents

Why this problem mattered

I kept meeting commerce operators who spent their week stitching together GA4, Clarity, agency advice, and instinct, yet still felt they were running growth in the dark. I chose this problem because the product should do more of that monitoring and synthesis for them. It should turn scattered evidence into an always-on growth system and return time to the people running the business.

Overview

UserApproved.ai began as an effort to simulate how different shoppers experience real ecommerce journeys before and after launch. The public product retrospective describes a multi-layer system: a Persona Agent for beliefs, desires and emotional state; an Orchestrator for plans and interrupts; and an Execution Agent for browser interaction.

The user problem

Analytics can show where shoppers drop off, but teams still need evidence about why trust breaks, expectations fail or a journey becomes confusing. The product direction focuses on repeatable, persona-grounded observation rather than a single generic agent or a fluent synthetic interview.

What the team shipped

The team released a public beta for browser-based shopper journeys, developed persona-grounded evaluations, and evolved the architecture toward separate cognition, orchestration and execution layers. The current public product connects journey evidence with continuous conversion diagnosis; confidential implementation details are intentionally excluded here.

Product decisions

Publicly described decisions include narrowing from general persona simulation to shopping behavior, separating persona reasoning from browser execution, waking the Persona Agent only when meaningful interrupts occur, and grounding agents in real interview narratives rather than trait fields alone.

Publicly supportable impact

The founder’s public retrospective reports an internal evaluation improvement from 19/50 for a baseline agent to 34/50 for a persona-grounded agent, with 99% journey completion. These are disclosed internal results, not independent validation, and should be presented with that limitation.

Lessons

Behavioral realism is not the same as persuasive prose. One synthetic journey should never be treated as ground truth; value comes from repeated patterns, preserved traces, clear boundaries and comparison with real-user evidence.

Related public announcements and coverage

Public sources: UserApproved.ai product site, https://www.userapproved.ai/; product and architecture retrospective, https://www.linkedin.com/pulse/what-we-learned-building-ai-agents-think-like-real-shoppers-wu-ruzcc

Team acknowledgment

This was collaborative work. The public retrospective thanks Canoee Liu, Sky Yin, Chao Wang, Steve Higgins, Arnold Yang, Frankie Tam, Wesley Chen, Jonas Tirona and others; final publication should confirm individual attribution and current team roles.