Rules vs ML repricing: when to switch your pricing engine

Every e-commerce repricer is one of two things: a rule engine that does exactly what you tell it, or a machine-learning model that does what the data tells it. Most teams pick the wrong one for the stage they're at — and pay for it in either lost margin or lost control.

How each one actually works

Rule-based repricers

You write the policy. "Match the cheapest Buy Box competitor minus €0.01, but never go below cost + 12%." The engine runs the policy on a schedule (from minutes to once a day, depending on the tool) against the prices it collects. Predictable, auditable, and exactly as smart as whoever wrote the rules.

ML repricers

You set goals (margin target, win rate, sell-through) and constraints (floors, ceilings, MAP). The model learns which price wins the Buy Box, which price maximises expected profit given elasticity, and updates continuously. It will sometimes price higher than competitors and still win, and sometimes price lower and lose — both correctly, both for reasons you'd struggle to explain in a meeting.

When rules are the right answer

• Small SKU count (< ~500). The training data per SKU isn't there for ML to beat a well-written rule.
• Heavy MAP or brand constraints. If 80% of your decisions are "don't break the policy", an ML model adds variance you'll have to clip anyway.
• Regulated categories. Pharma, alcohol, financial services — you need to be able to defend every price.
• New catalogue. No sales history means no training data. Start with rules; switch when you have signal.

When ML is the right answer

• 5,000+ SKUs. The rule maintenance cost grows faster than ML's training cost.
• High-velocity categories. Electronics, fashion, consumables — anywhere elasticity moves week to week.
• Marketplace presence. Buy Box dynamics on Amazon and similar are non-linear enough that hand-tuned rules consistently underperform.
• Margin-driven, not match-driven. If your goal is "extract the maximum profit per visit" rather than "always be cheapest", rules can't see the elasticity curve and ML can.

The hybrid that almost always wins

In practice the best repricing stacks are layered:

1. Constraints layer — hard rules: cost-plus floors, MAP, brand ceilings, channel rules. ML never overrides these.
2. Optimisation layer — ML chooses the price inside the allowed band based on elasticity, competitor signals, and inventory.
3. Override layer — humans force prices on hero SKUs, promos, and clearance. These bypass ML for the duration of the override.

This stack is what most "ML repricers" actually are under the hood — the constraints and overrides do much of the work, and ML handles the contested middle.

Amazon repricer alternatives, briefly

If you're shopping for a repricer beyond Amazon-native tools (Buy Box-focused, opaque about logic):

• Pure rules — RepricerExpress, Sellery. Great for tight control on small catalogues.
• ML/algorithmic — Aura, Bqool, Feedvisor. Beat rules on large catalogues but require trust.
• Off-marketplace, own store — Luska, Prisync, Competera. Optimise across your own site + comp set, not just Buy Box.

What "good" looks like

• You can explain any price. Either "rule X triggered" or "ML chose €Y because elasticity Z and competitor gap W". Black boxes don't survive contact with a CFO.
• Floor compliance is 100%. Not 99% — one MAP breach can cost a brand relationship.
• Margin / unit, not just sales / unit. A repricer that increases revenue while compressing margin is working against you.

See also competitive price monitoring guide and price elasticity of demand for the data inputs that drive both repricer types.