Demystifying AI in E-commerce
"AI" has become a marketing buzzword applied to everything from smart thermostats to basic if-then logic. In e-commerce, though, machine learning represents a genuine capability shift, one that's particularly powerful for product merchandising.
Let's be specific about what machine learning actually does in the context of Shopify collection sorting.
The Core Problem: Multi-Dimensional Optimization
To understand why ML helps, you have to understand why the problem is hard.
A collection with 200 products has roughly 200! (factorial) possible orderings. That's a number so large it defies comprehension. Manual sorting means a human picking one ordering based on gut feel. Simple algorithmic sorting means ranking by one metric (revenue, date, etc.).
Neither approach accounts for the interdependencies between variables:
- A product might have high revenue but low inventory (dangerous to promote heavily)
- A product might have low absolute sales but a very high conversion rate (hidden gem)
- A new product might have high potential but insufficient data to evaluate fairly
- Products on sale might temporarily deserve boosted placement
Machine learning can hold all of these variables simultaneously and find orderings that optimize for a composite goal.
How SortLab's Algorithm Works
SortLab uses a scoring model that evaluates each product in a collection across multiple dimensions, then generates a composite score used to determine position.
Revenue Signal (weight: high) The model calculates a rolling revenue contribution for each product, weighted more heavily toward recent performance. A product that sold well 6 months ago but has flatlined recently gets a lower score than one whose revenue is growing.
Conversion Signal (weight: high) Revenue can be high simply because a product appears early in the collection and gets more traffic. Conversion rate normalizes for this. A product with 8% conversion is valuable regardless of its current position.
Inventory Health (weight: medium) The model applies a penalty for low-inventory products. Promoting a product that's about to go out of stock creates a poor customer experience and wastes a prime position.
Recency Boost (weight: adjustable) New products don't have historical data. The model applies an adjustable boost factor to newly added products to give them a chance to prove themselves before being sorted purely on historical performance.
Margin Contribution (weight: configurable) For merchants who share margin data, the model can prioritize products that generate more profitable sales over high-volume, low-margin items.
The Scheduling Component
Static optimization isn't enough. Your best-selling products today may not be your best-sellers tomorrow. Inventory changes. Trends shift. New products launch.
SortLab re-runs the optimization algorithm on your chosen schedule, ranging from manual control to once an hour. Each run incorporates the latest sales data, inventory levels, and behavioral signals.
This continuous optimization is what separates AI-powered merchandising from manual sorting. The algorithm doesn't get tired. It doesn't have opinions about which products it personally likes. It follows the data.
What the Algorithm Can't Do
Intellectual honesty matters here. Machine learning is powerful but not omniscient:
It can't predict true demand. If a product would be popular if only customers knew about it, the model won't know that. It works from existing behavioral data.
It can't account for brand strategy. If you're intentionally promoting a product for strategic reasons (a new brand partnership, a seasonal push), you'll want manual controls to override the algorithm.
It needs data to learn. For very new stores with minimal transaction history, the algorithm has less signal to work with. Recommendations become more accurate over time as your sales data grows.
It optimizes for the goal you give it. If you optimize purely for revenue, you may inadvertently accelerate inventory imbalances. Setting the right objective function matters.
The Result
The combination of multi-signal scoring and continuous re-sorting consistently outperforms any single-metric approach. In tests across SortLab's merchant base, AI-optimized collections generate 24% more revenue per visitor on average compared to the same merchants' previous sorting approach.
That number will vary by store, but the direction is consistent: systematic, data-driven sorting beats intuition.