Start with behavioral data, not assumptions — session recordings, heatmaps, or funnel analytics that show where visitors actually drop off. Pick the single biggest leak, form one hypothesis, and run one controlled test before expanding the program.