Lookalike Modeling

Under the hood the platform builds a feature vector for every user it knows about, fits a model that separates your seed members from everyone else, and then ranks the remaining population by that score. The percentage you select is a cutoff on that ranked list rather than a similarity threshold: a one percent audience is simply the top slice of the country's users by model score. Only the portion of your seed the platform can match to real accounts contributes, so upload size and match rate are two different numbers.

Seed size, homogeneity and recency pull against each other. A seed too small leaves the model fitting noise; a seed padded with every buyer you ever had mixes several distinct patterns and produces an audience that resembles no one in particular. Widening the percentage cutoff buys reach at the cost of precision, and each step out dilutes toward the general market. Value-weighted seeds, where each member carries a revenue figure, usually beat flat lists because they tell the model which patterns are worth more.

Practically, teams build a small ladder of seeds: closed-won accounts, high-lifetime-value customers, and recent high-intent leads, then test narrow and broad cutoffs against each. Existing customers are excluded from delivery so budget goes to acquisition, and extra demographic layers are kept light because heavy manual targeting fights the model it sits on. A scorecard quiz is a convenient seed source, since only leads above a scoring threshold are exported, which keeps the seed intentionally narrow rather than accepting every form fill.

A lookalike can only find people who resemble whoever is already in your seed, so it reinforces the shape of your current base and will not discover a segment you have never sold to. Any accidental skew in that base, a channel that over-delivered one region or one job family, is reproduced and amplified at scale. The model is also bounded by what the platform observes about its own users, which is opaque and shifts without notice, and in narrow B2B markets the qualifying population may simply be too small to rank meaningfully.

Judge a lookalike on downstream quality, not on click metrics. Compare cost per qualified lead and win rate against a broad-targeting control running the same creative, because a lookalike that only improves click-through is selecting for curiosity. Track the share of leads from the audience that match your fit criteria; if that share is no higher than from untargeted traffic, the seed is not carrying the signal you assumed.

Watch decay over the life of the audience. Model scores are recomputed as the platform learns, but your seed ages, so refresh it on a regular cadence and compare performance before and after. Frequency and reach also tell you when a narrow cutoff has exhausted itself: rising frequency with flat conversion means the audience is saturated and widening the percentage is cheaper than raising bids.

The most frequent error is seeding on the cheapest available list, usually all newsletter subscribers or every form fill. The model then learns what a content downloader looks like and the resulting audience converts to content rather than to revenue. Seed on the outcome you actually want more of, closed-won accounts or customers past a retention milestone, even when that shrinks the list. A smaller seed built on the right outcome beats a large one built on availability.

The second error is stacking heavy manual targeting on top of a lookalike, adding job titles, interests and narrow age bands. Each layer removes people the model ranked highly for reasons you cannot see, and the audience shrinks until delivery becomes expensive. Use exclusions freely, since removing current customers and recent converters is unambiguous, but add positive filters one at a time and only when a test shows the unfiltered audience underperforms.