How similarity is built
- Lookalikes inside an ad network use that network’s own signals and inventory. Effective for activation, opaque in method and not usable elsewhere.
- Behavioural lookalikes find audiences with similar affinity and interest structures across the open market, independent of one network.
- Demographic lookalikes match on age, gender, income and region. The weakest form, because demographic similarity predicts behaviour poorly.
The seed decides the outcome
- A seed of existing customers produces more people like those already won: the group with the least headroom.
- A seed from one channel reproduces that channel’s skew. A newsletter list is the part of the audience that reads newsletters.
- A seed that is too small fits similarity to noise.
- A seed from one market does not transfer. Using a domestic seed abroad assumes the audience has the same structure there, which is the assumption worth testing.
A more productive seed is often the audience of a club, artist or media title your customers overindex on. It points outward, to people not yet reached.
Lookalike and overlap are opposites
A good lookalike has high affinity similarity and low overlap with the source: it behaves like your audience and consists of different people. High overlap means the expansion mostly re-reaches the same group, which looks efficient in campaign reports and produces no growth.
Common mistakes
- Customer list as default seed. Convenient, and the seed with the least growth potential.
- No check on overlap. Without it, you cannot tell expansion from repetition.
- Copying seeds across countries. Build the seed in the market you are entering.
How Rascasse uses it
Rascasse is used before activation: it shows which brands, clubs, artists, media and interests an audience overindexes on, per country, so the seed can be chosen deliberately instead of defaulting to the customer list. The audiences are then built in the ad network where the campaign runs.
Related use cases: Paid Social & Programmatic, Media Planning, Audience Profiling.