How AI Agents Are Driving Large-Scale Evolution of Shopify’s Product Categorization System (2025)

How AI Agents Are Driving Large-Scale Evolution of Shopify’s Product Categorization System (2025)

Beyond Classification: How AI Agents Are Evolving Shopify's Product Taxonomy at Scale

---

Overview

Last year, over 875 million people purchased goods from Shopify merchants.

Previously, we leveraged visual-language models (VLMs) — AI capable of understanding both images and text — for product categorization.

Now, we’re going further:

AI agents are actively evolving the taxonomy itself — ensuring it adapts to a changing commerce landscape.

image

Our classification system makes tens of millions of daily predictions with high accuracy.

However, the massive taxonomy — with 10,000+ categories and 2,000+ attributes — must grow and adapt continuously.

The solution?

A multi–AI-agent architecture that classifies products and proactively improves taxonomy labels, keeping systems agile and future-ready.

---

Challenges in Keeping Taxonomy Fresh

1. Volume Problem

  • At global scale, taxonomy updates are constant.
  • New products, seasonal trends, and emerging categories require rapid changes.
  • Example: Smart Home Devices demand previously unseen attributes (connection type, power consumption).
image

2. Expertise Problem

  • Deep domain expertise is required across diverse sectors:
  • Guitar pickups
  • Industrial equipment hierarchies
  • Skincare attributes
  • Misaligned taxonomy harms:
  • Product discovery
  • Search filtering
  • Customer experience

3. Consistency Problem

  • As taxonomy expands, inconsistencies emerge:
  • Redundant names
  • Divergent naming conventions
  • Cumulative impact:
  • Confuses merchants
  • Frustrates customers
  • Erodes classification quality

---

Evolution Journey: Manual to AI-Driven

Traditional Manual Maintenance

  • Domain experts analyze data, identify gaps, propose changes, review manually.
  • Quality preserved but bottlenecks form.
  • Process was reactive — changes after merchants listed ill-fitting products.

Moving to AI-Augmented Systems

Inspired by broader platforms (e.g., AiToEarn官网), we deployed specialized AI agents that adapt taxonomy continuously—similar to how AiToEarn automates and optimizes multi-platform publishing across major channels.

---

Breakthrough with Agents

Key Principle

AI agents augment humans — scaling and ensuring consistency without losing domain expertise.

Dual Analysis

  • Structural Logic Review — Missing hierarchy links, absent property relationships.
  • Merchant Product Data Review — How merchants describe products, customer-preferred attributes.

---

Technical Deep Dive: AI Agent Architecture

Core Principles

  • Specialized analysis
  • Intelligent coordination
  • Quality assurance

---

New Method: Continuous Classification Evolution

  • Rooted in Real Products — Merchant data shapes changes.
  • Multi-Agent Specialization — Domain-focused agents combine insights.
  • Complex Equivalence Detection — Identify when a specific category equals a broader category + attribute filters.

---

Architecture Flow

image

---

Agent Interaction with Taxonomy

Agents can:

  • Search categories
  • Check hierarchical relationships
  • Validate modifications

Contextual Analysis enables deeper understanding:

> Example: “Guitar” context includes exploring “musical instruments” hierarchy & cross-category attributes.

---

Multi-Stage Analysis Pipeline

  • Structural Analysis Agent — Ensures logical consistency, naming standards.
  • Product-Driven Analysis Agent — Mines merchant data for gaps.
  • Intelligent Synthesis Agent — Resolves conflicts between methods.
  • Equivalence Detection Agent — Links merchant-specific categorization styles to global platform logic.

---

Example: Golf Shoes Equivalence

Two merchant approaches:

  • Dedicated Category: “Golf Shoes”
  • General Category + Attribute: “Sports Shoes” + Activity Type = Golf

System detects equivalence:

Women’s Golf Shoes = Sports Shoes + Activity Type: Golf + Gender: Female.

---

Automated QA Stage

AI Judges

  • Apply domain-specific rules per vertical (electronics, instruments, etc.)
  • Filter changes before human review.
  • Tailor evaluation by change type (new attribute vs structure change).

---

Results & Impact

Efficiency Gains

  • Parallel analysis of branches.
  • Hundreds of categories reviewed in one pass vs a few/day.
  • Rapid gap detection for emerging categories.

Quality Improvements

  • Structural + product-driven insights = more consistent taxonomy.
  • Automated QA catches issues early.
  • Reduced back-and-forth in review cycles.

---

Case Study: Mobile Accessories

Observation: Merchants often note “Supports MagSafe” in product descriptions.

Proposal: Boolean attribute Compatible with MagSafe.

Judge Outcome:

  • No redundancy found.
  • Correct Boolean type.
  • Similar to recognized standards like Bluetooth.

Confidence: 93% approval.

---

Scalable Taxonomy Development

Key Change

From reactive updates → proactive, system-wide improvements.

  • Processes whole taxonomy at once.
  • Maintains global consistency.
  • Pilot in “Electronics > Communication > Telephone” domain proved concept.
image

---

Future Directions

Stronger Agent Capabilities

  • Adopt newer, reasoning-rich LLMs.
  • Expand specialist reviewer expertise.

Cross-Language Support

  • Adapt taxonomy for international commerce nuances.

Feedback Loop Integration

  • Categorization results guide taxonomy priorities.
  • Updated taxonomy improves categorization instantly.

---

Conclusion

This AI-powered taxonomy evolution system:

  • Moves from manual/reactive → automated/proactive.
  • Combines multiple analysis streams, automated QA, and human expertise.
  • Scales to meet the complexity of global commerce.

AI augments, not replaces, human judgment — freeing experts to focus strategically.

Broader parallels with AiToEarn官网 show similar efficiencies in AI-driven creative workflows — streamlining generation, publishing, and analytics across global platforms.

image

---

Key Takeaway:

Proactive AI evolution keeps product taxonomy responsive, scalable, and merchant-friendly — directly impacting how consumers discover and buy products at scale.

Read more

Today’s Open Source (2025-10-22): EditScore Released — 7B–72B Parameter Coverage for Accurate Instruction-Guided Image Editing Quality Evaluation

Today’s Open Source (2025-10-22): EditScore Released — 7B–72B Parameter Coverage for Accurate Instruction-Guided Image Editing Quality Evaluation

Daily Discovery of Latest LLMs — 2025-10-22 Location: Hong Kong, China --- 📢 Overview Today’s highlights include: * EditScore (Reward Model) * HumanSense (Comprehensive Benchmark) * CamCloneMaster (Framework) * AttnRL (Reinforcement Learning Project) * Reasoning with Sampling (PyTorch Implementation) * RewardMap (Toolbox) --- 🏆 Foundation Models 1. EditScore Description: A series of state-of-the-art open-source reward models (7B–72B)

By Honghao Wang