How do Facebook lookalikes based on platform-specific Custom Audiences perform?

Question

LKLs created from platform-specific Custom Audiences look significantly different than LKLs created from a Custom Audience containing both platforms. For my tests I used the same sample, just segmented differently (i.e. 100% of the users in the two platform-specific Custom Audiences are also in the platform-agnostic Custom Audience). 

A reach comparison shows that I can add substantial reach compared to LKLs created from a Custom Audience with both platforms. (Details see below).

Does anybody have a strong observation which segmentation of Custom Audience seed produces the (in ROAS) highest-performing Lookalikes?

 

Answer 1

I've hit & missed by tweaking targeting of my custom audiences, but for criteria like OS (where it's mandatory to split in adgroups) or sometimes others such as gender (not mandatory by the platform, but might be more effective if creatives are very differenciated or the product catters specifically for one), I recommend having them separated in the seed audiences, for the reason explained above

Taking a specific example, imagine a country of 100m users with 50/50 Android/iOS:
- a custom audience of mixed OS would generate a lookalike 2% of 2 million people, which once filtered will be cut by roughly half (not exactly, illustration purpose only) = 1m targeted users from a LAL2%
- a custom audience of 1 specific audience would generate a lookalike 1% of 1 million people, once filtered will be almost identitical = 1m targeted users from a LAL1%
The LAL2% is likely to be "broader" than the LAL1% hence less effective.

2 nuances:
Seeing this tactic working, I started building multiple seeds, adding things like age, languages etc but this reached a limit and at some point it wasn't worth the effort (added complexity, audience management...).
Also: this applies well for very large geos (or tiers), less so when creating audiences for smaller geos.

Answer 2

I've definitely seen platform-specific custom audiences improve the performance of LaL audiences, but I don't know how well that generalizes (eg. is it always the case? when I segment out iOS users, do I meaningfully increase average monetization enough to see it across both groups?).

My approach with LaLs is to make the underlying custom audience as specific as possible so long as I am hitting 10,000 seed users. At less than 10,000, I prefer to aggregate across dimensions for reach rather than achieving specificity with the audience definition.