The lookalike that failed because the seed was too good

The lookalike that failed because the seed was too good
Myth: "Feed your lookalike only your highest-value rebilling users — purest seed, best audience."

Logical, and a team building a 1% lookalike for a mainstream offer did exactly that: seeded with 600 top rebillers. The lookalike underperformed broad targeting by 30% on CPA for six weeks.

They rebuilt it with 9,000 all-converters (including front-end-only), no value filter. CPA dropped 24% below broad.

The pure seed was too small and too narrow — 600 people gave the algo a cramped, statistically noisy pattern, and "high-value rebiller" correlated with traits (older, higher income) that were hard to find at scale cheaply. The messy 9k seed gave a richer signal of "people who sign up for this" — which is what the front end actually needed to optimize.

Reality: a tiny pristine seed starves the algorithm — for dating, a large noisy seed of all converters usually beats a small clean one of your best.