A lookalike audience is not a list of people who resemble your customers in any way a human would describe. It is the output of a model trained on a seed group.
The seed defines everything downstream
An advertiser supplies a set of known people, usually past purchasers or engaged visitors. The platform examines what those people have in common across the signals it holds.
It then scores its wider user base for similarity on those dimensions and returns the closest matches as an addressable audience.
The model has no notion of what makes a good customer. It reproduces whatever pattern exists in the seed, including patterns the advertiser would not have chosen.
A contaminated seed produces a confident mistake
If the seed contains large numbers of one-time discount buyers, the model finds more discount buyers and does so efficiently.
The audience will look successful on immediate response and disappointing on retention, because it faithfully replicated the wrong group.
Seeds built from high-value repeat customers usually perform better despite being smaller, since the pattern being copied is the one the business actually wants more of.
Size and precision trade against each other
Lookalike audiences are usually offered at several breadth settings. A tight setting returns the most similar people and a small addressable pool.
A broad setting returns many more people with a weaker resemblance, which suits campaigns that need scale more than precision.
Neither is correct in general. The right setting depends on whether the constraint is finding enough people or finding the right ones.
Similarity is measured on platform signals only
The model can only compare people on data the platform holds, which is behaviour inside its own environment plus whatever it has inferred.
Attributes that matter commercially but leave no trace there, such as procurement authority or household circumstances, cannot enter the calculation.
This is why lookalikes often work well for consumer products with visible interest signals and poorly for business purchases where the deciding factors are invisible.
Refreshing the seed matters more than rebuilding the model
A seed list ages. The customers acquired two years ago may have come through channels and offers that no longer represent the business.
Rebuilding the audience from a stale seed simply reproduces the older pattern with more confidence, which is worse than not rebuilding at all.
Maintaining the seed as a live definition of a good customer, updated as the business changes, does more for performance than adjusting any setting on the audience itself.