A Lookalike Is Only as Honest as Its Seed: Auditing the Audience You Clone
A lookalike audience is a cloning instruction, and most of the argument about whether lookalikes work is really an argument about what got cloned. Feed a platform your best customers and it will find people who resemble them on the signals it can see. Feed it a seed that quietly mixes lifetime buyers with one time gift purchasers, full price customers with promotion only responders, and a cohort of accounts created by a misattributed campaign, and it will faithfully find more of all of that, in proportion. The output always looks like growth in the platform dashboard. Whether it is growth in customers you wanted is a question about the seed.
Define best before you upload anything
Best has to be a business definition, not a platform default. Revenue alone seeds you with whales whose behaviour may be unrepeatable. Frequency alone seeds you with small basket regulars. A working definition for most stores combines recency, at least two orders in a period, margin after returns, and no dependence on a single promotion. Write the definition down with its exclusions: staff, wholesale, known resellers, accounts with a return rate that erases their margin. If the seed cannot be stated in one paragraph a finance colleague would accept, it is not ready to be cloned.
Seed size and freshness trade against purity
Platforms ask for volume, and broadening the seed to reach a recommended size dilutes exactly the traits you were cloning. A smaller, purer seed in a larger target market usually outperforms a swollen seed that includes everyone who ever bought. Refresh cadence matters as much: a seed built on customers from two years ago clones the tastes, price sensitivity, and category mix of a store that may no longer exist. Rebuild seeds on the same cycle as your segment reviews, and version them, so that a change in lookalike performance can be traced to a seed change rather than argued about from memory.
Judge the clone on cohort behaviour, not first order
Lookalike reporting loves the first conversion because it is fast and flattering. Judge the acquired cohort at ninety days against the seed's own behaviour and against your other prospecting cohorts: second order rate, return rate, discount depth needed for the second purchase, and support contact rate. Clones of a good seed should resemble it in shape, a little worse on every measure is normal, dramatically worse on returns or repeat rate means the platform cloned the visible surface, the demographics and device signals, and missed the behaviour that made the seed valuable.
Keep a holdout or you are grading your own homework
Platform attribution will claim credit for lookalike conversions that would have happened through other channels, especially where the clone overlaps your existing remarketing pools. Exclude current customers and site visitors from lookalike targeting, and hold out a geographic or audience slice where the lookalike does not run. The comparison does not need to be perfect to be useful. Without any holdout, the budget conversation reduces to the platform's report against nobody's, and the seed, the thing fully within your control, never gets the scrutiny it deserves.
Keep reading
Related guides on this site: RFM Segments: Setting Thresholds From Your Own Data, Not a Template Churn Risk Scores Are Useless Without an Action Window Practical answers about personalization.
Frequently asked questions
What is a seed audience for a lookalike?
The source list of existing customers a platform uses as the pattern for finding new people with similar characteristics. Its composition decides what the lookalike actually clones.
How big should a lookalike seed be?
Big enough for the platform to model, but purity matters more than volume. A smaller seed of genuinely repeat, profitable customers usually outperforms a large seed diluted with one time or promotion only buyers.
How should lookalike performance be judged?
On the acquired cohort's later behaviour, second orders, returns, margin, compared with the seed and with other prospecting cohorts, not on first order platform attribution alone.