// SOFTWARE
The black box problem in marketplace repricing
7 min readnijitech
If a system that changes your prices automatically cannot say why it changed them, mistakes cannot be found. Nobody notices until the margin has melted. How to build auditable repricing.
Full post
On a marketplace, price is a number that can change dozens of times a day. Past a certain product count, tracking that by hand becomes impossible and automation becomes inevitable. The problem is not the automation itself; it is that the system changing the price cannot say why it changed it.
If the reasoning behind a pricing decision is not visible, that decision cannot be audited. And a decision that cannot be audited does not get noticed when it is wrong — the margin erodes slowly and nobody can point to a single moment where it broke.
What the black box hides
When a price drops, the question is simple: did the drop answer a real competitor move, or the leftover listing of a seller who is out of stock? On screen the two look identical. The difference only appears when you can inspect the reasoning behind the decision.
Mistakes that go unnoticed in a system with no visible reasoning
- Treating an out-of-stock competitor listing as a valid price signal
- Comparing prices with and without shipping on the same plane
- Mistaking a different variant (colour, count, bundle) for the same product
- Taking a short campaign price as the permanent market price
Individually these are all small errors. In a system producing thousands of decisions a day there is no such thing as a small error; every one of them gets multiplied by volume.
The price spiral
The best-known failure mode of automated pricing is a mutual downward spiral. If two sellers are both set up to undercut their rival, both prices fall below cost with nobody intervening. From its own point of view the system worked correctly: it applied the rule it was given.
Guardrails matter more than the rule
The real design of a repricer is not in how it sets the price but in where it stops. A margin floor, a daily change ceiling and a manual lock for specific products — these are not a brake on the system; they are what keeps it alive.
Minimum guardrails
- Never go below cost plus target margin — an absolute floor
- A maximum percentage change within a single day
- The ability to switch automation off entirely for specific products
- Human approval before applying changes above a threshold
A system with no decision log cannot be fixed
If every price change carries a record of the signal it rested on, then when the margin drops you can go back and find which signal was misleading. Without that record you are left with the outcome alone, and you cannot work backwards from an outcome to a cause. The cost of keeping the log is a few fields; the cost of not keeping it is switching the system off.
This matters even more for multi-variable targets like the Buy Box. If you cannot tell whether you lost the box because of price, delivery time or seller performance, every change you make afterwards is a guess.
In short
When evaluating a repricer
- Can you see the reasoning behind every price change?
- Where is the floor, and who decided it?
- Are competitor signals verified — stock, shipping, variant?
- Are changes above a threshold put to a human?
Building a fast repricer is not hard. What is hard is explaining, six months later, why you sold at that price.
Products mentioned in this post
From the glossary: Explainable AI (XAI) · Guardrail · Buy Box