Miinto set out to steer by profit on each order

A marketplace needs to know what each sale leaves

Some products earn Miinto a lot and some very little, so on some it pays to spend a lot on ads and on others almost nothing.

Miinto is a marketplace with no cash tied up in stock. It earns little on each order, so a small error in what it expects to earn can matter.

Finance knew the total profit each month. What nobody could see was what each order, product and campaign earned. That was the gap my team and I had to close.

The same number came in several versions, and nothing was written down.

We made wrong calls because nobody agreed which number was right. With the leadership team, we agreed what to steer by and wrote each meaning in the metrics book. For returns, that means choosing whether a sale counts on the order day or comes off when the item comes back. There isn't one right answer, so the rule has to be written down.

I started with the orders and what was in them, then added products and customers, payments, the books and ad costs. Once we got the return estimates right, the figures matched the books very closely. Checks in the data catch problems, but not all of them. Systems change, and the setup can break when they do.

A new product needs a return estimate before its first returns arrive.

We estimate how often a product will come back from its type, brand and colour. As real returns arrive, the guess gets sharper. We use those estimates to work out the expected gross profit on each item and order.

Miinto sends expected gross profit to the ad platforms through ProfitMetrics. Since September 2025, Miinto has used it to set the ad budget.

From February to August 2026, expected gross profit after ads on orders without a voucher was 171% higher than in the same months of 2025, before Miinto began steering by expected gross profit in September 2025. That work covered the ad budget, dynamic pricing and which products got the best spots on the site.

The figure is an average for July to September 2026, with my own requests left out. It counts requests, not correct answers. In early 2026, I moved my focus to tools on the shared numbers.

I set up a way for people at Miinto to build their own tools and automate their own processes safely on the shared data. For Product Center, I built the base and security, and the product team built the rest with Claude. I built Marketing Intelligence for marketing to see the ad budget, ad spend and campaign plan together. Finance got tools that put month-end work in one place.

When someone asks Claude, the connection checks who is asking and what that person may see before Claude gets any numbers. Claude reads the meanings in the metrics book. Claude still picks the wrong number now and then, so we check the answers that matter.