SIX|Documentation

Functional Overview

The Netflix Paradox

Why has retail failed to match streaming's personalization despite 15 years of promises? The answer lies in three fundamental constraints that AI agents can finally overcome.

The Question Worth Asking

"Every e-commerce platform claims to be 'personalized.' Yet shopping still feels like wandering a store designed for someone else. Netflix knows my taste better than my local boutique. Why?"

Three Constraints That Broke Personalization

1. The Data Density Gap

Sparse signals in a world that demands dense understanding

Streaming services capture hundreds of behavioral signals per user per week: play, pause, skip, rewatch, hover, scroll-past, abandon-at-minute-23. This creates a dense preference map.

Retail captures 2-3 purchases per category per year. You bought a winter coat in 2021. Based on that single data point, the algorithm recommends coats for the next decade. The signal-to-noise ratio is catastrophic.

Netflix (streaming)

~500 behavioral signals/week/user

Knows: genre preference, pacing tolerance, mood by time of day

Retail (commerce)

~3 purchase signals/year/category

Knows: you bought a coat once

2. The Inventory Physics Problem

Atoms don't scale like bits

Netflix can recommend any movie to any user simultaneously. Digital inventory is infinite. The marginal cost of one more stream is effectively zero.

A retailer has 3 units of that jacket in medium. Recommend it to 1,000 users who all want medium? 997 hit dead ends. Worse: the algorithm just trained users to expect frustration.

The compounding problem:

  • • Static recommendations ignore inventory state
  • • Out-of-stock frustration erodes trust in personalization
  • • Users learn to distrust "recommended for you"

3. The Intent Mismatch

One interface for two fundamentally different modes

Users alternate between two distinct shopping modes—often within the same session:

Goal-Oriented Mode

"It's raining. I need a waterproof jacket. Under $200. Available now."

  • • Goal-directed, urgent
  • • Wants efficiency, filters, stock status
  • • Friction is the enemy

Browsing Mode

"I want to feel something. Show me what I didn't know I needed."

  • • Discovery-oriented, aspirational
  • • Wants storytelling, curation, surprise
  • • Speed is secondary to experience

Traditional e-commerce serves a single, compromised interface. Too efficient for inspiration. Too cluttered for transactions. Satisfying neither mode fully.

What Changes With Generative AI

The breakthrough isn't better recommendations. It's generative interfaces.

Instead of fetching a pre-built page template, we generate the layout in real-time based on who you are, what you need, and what's actually available.

Data Density → Intent Capture

Instead of inferring from sparse history, we let users explicitly select their mode. Goal-Oriented or Browsing. Clear signal, instant personalization.

Static → Inventory-Aware

The AI only considers in-stock products. Recommendations are grounded in reality, not wishful thinking.

One Layout → Infinite Layouts

Each user gets a unique interface. Goal-Oriented shoppers get dense grids. Browsing shoppers get editorial spreads. Same store, different experience.

The Retail Store Metaphor

Imagine walking into a physical store where the shelves rearranged themselves the moment you entered. Based on your visible intent (rushed vs. browsing), the weather outside, and what's actually in stock today.

Traditional E-commerce

"Here's our store. Same layout for everyone. Good luck finding what you need. Also, half of what we're showing you is out of stock."

Generative Commerce

"I see you're in a hurry and it's raining. Here are waterproof jackets in your size, in stock, sorted by fastest shipping. Grid layout. No distractions."

Why This Wasn't Possible Before

CapabilityBefore (2015-2022)Now (2024+)
Layout GenerationRule-based templatesAI generates complete UI schemas in <3s
Structured OutputUnreliable JSON parsingNative JSON mode with Zod validation
Context UnderstandingKeyword matchingSemantic understanding of intent + context
Cost per Request$0.50+ per generation~$0.01 per layout generation
Latency5-10s (unusable)1-3s (acceptable for page load)

How We Know This Works

Technical Validation

  • • Layouts generate in <3 seconds
  • • Zero hallucinated product IDs
  • • All recommended products are in-stock
  • • Persona-appropriate component selection

Experience Validation

  • • Goal-Oriented layout feels efficient, transactional
  • • Browsing layout feels editorial, aspirational
  • • Context (weather, urgency) influences product selection
  • • Same product catalog, different experiences

Continue Reading