I built a 9-week integrated campaign around a hypothesis that turned out to be wrong. Here's what the data showed, what I'd do differently, and how I confirmed my conclusions a year later.