See here for a step-by-step guide to setting up your first Lookalike Model.

How are the models created?

The goal of lookalike (LAL) segments is to expand a segment (the "seed segment") by finding similar users. In order to create a lookalike segment, Permutive uses machine learning to generate a lookalike model. The model learns the common features (in this case, what segments a user is in) and their relative importance.
For example, if we want to expand a segment of "Tennis Lovers", our lookalike model may learn that many users in the segment are also in a "Sports Lover" segment. This means that if we see a user that is in "Sports Lover", within some given probability, they are "almost" a "Tennis Lover" as well. Conversely, the model may learn that users who are "Tennis Lovers" are not "Basketball Lovers", and a negative importance can be given to the "Basketball Lovers" segment.

What do the red dots mean on my lookalike graph?

A red dot represents that a lookalike segment has been created with the given similarity and reach.

What happens when I choose 1P or 1P + 3P?

1P - Only first-party data is used to create the lookalike model. Depending on the amount of data available in your project, the model can be less accurate than one that uses both 1P and 3P.
1P + 3P - This means both first and third party data is used and thus much higher accuracy is possible.

Why can't I use just 3P data?

Using both 1P and 3P data is more beneficial as the LAL model produced will have more similarity and reach than if only 3P data is used.

Why does it take up to 24 hours to create my model?

Model creation takes up to 24 hours, but in many cases will finish within less than an hour, depending on the amount of available data. As we schedule model trainings at a specific time of day, hence in a worst case scenario,  the model will be available 24 hours after you select 'Build lookalike model...'

How many segments will I need in my project to create a model?

There is no minimum, but we recommend at least 10. If you use 1P-only, the more segments you have (and the more users your segments have in them), the more accurate the lookalike model will be.
If you use 1P and 3P, then this is less important as 3P data already has thousands of segments available.

What is the Minimum segment size for the seed set? 

We suggest a segment have at least 5,000 users or 0.02% of your overall audience to create an accurate lookalike model.

Will I be able to use the segments for targeting in my ad server?

Yes, lookalike segments behave like normal segments for all intents and purposes. Think of them as segments with automatically generated conditions.

What happens if I delete the seed segment the model was created from?

All lookalike segments created from a seed segment will not be automatically deleted and will still work. However,  since models are trained daily, using that day's data, when a seed segment has been deleted, users will no longer be entering the seed segment and the model will become less accurate.

Can I rename the segments I create from my model?

Yes, LAL segments will behave as any other segment.

Can I edit the lookalike segment properties once it has been created? 

No. However, the seed segment properties can be edited to define the users that will be in the LAL model segment. The model will then take 24 hours to retrain.

Can I refer to the LAL model segment in the building of another segment ie. like a building block?

Yes. LAL segments can be used like normal segments.

Example of a good/bad seed segment:

  • Email subscribers - highly engaged customers who have performed a concrete action. 
  • Over 5k users or 0.2% of your overall audience
  • Should not be based off a users presence in another segment (eg. using Segment Entry/Exit events)
  • Avoid using demographic segments as seed segments (eg. Male/Female)

Note: A CPM cost for delivering campaigns against lookalike segments will apply if third party data is being used in the calculation.

Did this answer your question?