What to do when everything is getting better and also somehow worse?

It’s your first Monday as the first head of billboard advertising at a prominent local dealership: three stores along the same stretch of highway, selling everything from $19,000 commuter sedans to $90,000 trucks. On the lot, a sale is a sale. In the back office, no two are alike. A truck comes with a service package that pays for years; a commuter sedan’s barely covers the oil changes. The cheaper the car, the likelier its loan goes into default. The pricier the car, the likelier the deal is fraud.

Your brief fits on a sticky note, so the whole marketing team can rally behind it: sell more cars. And you’re good at this. You hold out a control on every campaign, test every headline, and tune every step from the billboard to the signed paperwork. By spring, cost per approval has never been lower and approval rates have never been higher. The showrooms are packed on Saturdays, the finance office stays open late to keep up, and the business is running at its limits. Every report says you did it.

Then the first full year of the new book comes back, and it returns less than the legacy portfolio predicted. Optimized for volume at every stage of the funnel, it filled up with the cars that cost the most to carry: the loans that default and the service packages that barely pay. The buyers who used to drive off in $90,000 trucks are still out there; your billboards just stopped speaking to them. Every metric you were judged on assumed one thing: that more volume wouldn’t lower quality.

Out on the highway, your best-performing billboard still proudly reads “Free Lamborghini!” It has the lowest cost per sale the dealership has ever recorded, and the best cost per lead and per prospective buyer on the lot. The obvious fix is to chase the expensive cars instead, but that’s where the fraud is. Value was never the sticker price.

The averages the system optimized were carried by a minority of accounts. Swap the billboards for campaigns and the cars for merchants or subscribers, and this is the story of most recurring-revenue books. The answer isn’t to stop testing; it’s to test against the right target. Optimization has to be grounded in the signals that drive above-baseline behavior, identified as early in the selection process as possible.