The 95% of your catalog nobody is pricing

Walk through how prices actually get set at a company with six thousand SKUs. The top sellers, maybe three hundred items, get real attention. Someone watches competitors, runs promotions, argues about those prices in meetings. The other five thousand seven hundred items were priced once, usually at cost plus a house margin, on the day they entered the catalog. Some of those days were eight years ago.

Nobody decided those prices in any meaningful sense. There is no time. At five minutes per SKU per quarter, repricing the rest of that catalog would consume 475 hours, so it does not happen, and a default from 2018 quietly governs revenue in 2026.

Where the money sits

The long tail looks individually trivial and is collectively large. An item that sells thirty units a year at $42 is not worth a meeting. Three thousand such items are several million dollars of revenue priced by a rule of thumb that nobody has revisited since the rule was written.

Two things drift while nobody watches. Costs move, so if landed cost rose 18 percent since an item was priced and the price moved 10, margin quietly compressed and keeps compressing. And willingness to pay moves, so if the market would bear 8 percent more on a slow-moving niche part with no close substitute, that money was simply never collected. Across a deep tail, recovering two to four points of margin is a normal outcome, not a heroic one. On $3M of tail revenue, three points is $90,000 a year, every year, on products you were already selling to customers you already had.

What the math actually does

The system has two halves. The first is demand estimation. From your own sales history, the model estimates how each product's volume responds to its price. Some items are price-sensitive, some barely notice a change. No single slow mover has enough history to estimate alone, so the model pools information across similar products, which is the same reasoning a good category manager uses when she says that things like this tend to behave like this.

The second half is constrained optimization. You set the rules before anything moves. Floors, so nothing sells below landed cost plus a minimum. Ceilings, so nothing drifts into territory that insults a loyal customer. Brand rules, so the good, better, and best tiers stay in order. Competitive pegs on the handful of items where you genuinely fight for the order. Inside those boundaries, the optimizer finds the set of prices that does best on the objective you chose, usually margin dollars.

None of this is new mathematics. The methods are decades old and well understood. What changed is that the data plumbing and the computing needed to run them across six thousand SKUs no longer require an enterprise budget or an enterprise timeline.

You must be able to ask why

A pricing system you cannot interrogate is a pricing system you cannot trust, and a system the team does not trust gets switched off within a quarter, usually right after the first recommendation a veteran salesperson dislikes.

So every price change has to carry its reasons in plain terms. Why did this item move up 6 percent? Because its cost rose, because demand at the old price barely responded to the last increase, because the floor that used to bind no longer does. When the pricing manager can audit that reasoning and override it with one click, the system gets used and the overrides become training data. When she cannot, the system becomes an expensive report.

This is why we keep the whole pipeline deterministic and inspectable. The estimates come from your transactions, the rules come from you, and the optimizer's work can be replayed step by step. There is no point in recovering margin with a method your own team is afraid of.

When you should not do this

Honesty requires the other half of the argument. Demand estimation only works with enough history. A catalog that turned over completely last year. A business with eighteen months of clean sales data. A category where one customer is 60 percent of volume. In those situations the estimates would be noise wearing a suit, and an optimization built on noise does not recover margin, it redistributes it at random.

We have told prospective clients exactly that, and told them to come back after four more quarters of clean data. As a rough floor, you want two to three years of transaction history, a few thousand SKUs so the tail is real money, and cost data you actually believe.

If that describes your business, the five percent of the catalog you actively price is probably fine. The money is sitting in the ninety-five percent nobody is deciding.