How Facebook could do look-alike audiences in SKAdNetwork and iOS 14
Look-alike audiences is an incredibly helpful tool in the mobile marketer’s arsenal, and it’s one that most user acquisition experts have written off in terms of the future of mobile marketing on iOS. Look-alikes (LAL) are largely powered by the IDFA, and the consensus is generally that IDFAs will be scarce. (Plus, don’t forget, you can’t just use first-party data like an email address as an identifier if you haven’t received App Tracking Transparency permission.)
But could SKAdNetwork power look-alike audiences?
In a privacy-safe way?
Without identifying specific users or devices? And without violating any of Apple’s policies?
That question recently popped up in the Mobile Attribution Privacy group on Slack, and a few of us took a bit of time to think it through. I want to be clear: I don’t have any special or private insight into what Facebook is or might be doing here. This is just speculation on my part as to how privacy-safe look-alike audiences might be able to work under SKAdNetwork on Facebook.
The question: look-alikes via SKAdNetwork?
A marketer asked if anyone knew whether we will be able to create look-alikes based on SKAdNetwork specific events or not.
One reply suggested it should be possible to build look-alikes out of SKAN conversion values … as long as Facebook is able to decode them according to how you encoded the six-bit conversion values. A concern, however, is how Facebook would be able to link the conversions back to individual people without violating privacy and tracking people who may not have consented.
Here’s one way I think it could work, and how Facebook could do look-alike audiences in iOS 14. Interestingly, it’s sort of similar to how Apple Ads Attribution works.
First: how look-alikes work now
Let’s start here: in terms of how look-alike audiences work today, you send Facebook your app events plus users’ IDFAs for the audience you want to do look-alike targeting on. Typically, of course, you’ll pick your best users, your paying customers, or your most-engaged community members.
Once Facebook gets that data, they match it row by row to Facebook users.
Then they look at their universe of Facebook users — and it’s a large universe. WhatsApp has two billion users, Instagram has well over a billion, and the core Facebook experience itself is closing in on three billion.
Post IDFA, all this changes significantly.
Sure, for all your App Tracking Transparency users you get direct data points, and if Facebook gets enough data points from its own users agreeing to be tracked AND enough data points from other third-party apps whose users also agree to be tracked, the old methods still apply. There is likely to be some of this, but whether it’s 5% or 50% of the ecosystem … time will tell.
If it turns out to be more than 20%, that’s still a significant number of people, and a significant number of your app’s users. You’ll be able to do some testing to see how accurate modeling based off these users is for your non-consenting app installers, and that modeling could be very useful.
(I do wonder if more consent will come from your most valuable users, though. It might be the case that those who trust you enough to buy significant amounts of in-app purchases in your app/game should trust you enough to let you track them. Again … time will tell.)
But if you don’t get consent, then Facebook doesn’t know who the user is when they get the SKAN postback, so the standard LAL model simply wouldn’t work.
Now, here’s how LAL could work in iOS 14 via SKAdNetwork
There may still be hope for look-alike audiences, but I reiterate: this is just my guess. This isn’t insider information.
OK, here we go:
- Facebook knows which of their users engage with which ads in significant detail. They’ll have impression, video view, video complete, and click data, plus probably much more. That is definitely already a signal about “people who might be interested in XYZ.” (In addition, Facebook has a lot of first-party data on each user based on things people have explicitly told Facebook.)
- Facebook also knows which users they place in a specific SKAdNetwork campaign. That means that when a SKAN postback arrives, it could be one of the users that was in that campaign during the relevant time window.
- For Facebook, each ad campaign will have a huge number of users. But over time, if you randomize how you assign users to campaigns (so that it’s not always the same people in the same campaign ID), you could start getting a signal of the probability that a user in that campaign is a whale. Also, if advertisers encode “install day” into the six-bit SKAdNetwork conversion value, it would help Facebook narrow down the set of users who could potentially belong to that anonymous SKAN postback.
- These are just one of the many reasons why Facebook is reserving 91 out of the 100 possible SKAdNetwork campaign IDs for their own algorithms, only giving advertisers nine campaign options.
- So in theory, Facebook just needs to get a lot better at building a probability that a user would be a whale in a certain game as opposed to an IDFA-universe knowing that someone is a whale like today. This, of course, is where their data science army comes in.
All of this is pure speculation, but it is one conceivable method how Facebook could maintain efficacy — probably lower than today, but still decent — in look-alike campaigns. I’m also sure they have dozens of other methods and ideas that I can’t even imagine yet since I’m just looking from the outside.
It’s more than Facebook, of course
I’ve used Facebook in this example because the big social network is the most prominent example of look-alike audiences for most mobile marketers. But many if not all of the big platforms are preparing to run similar technology, based on what I’m seeing. Whether you call it differential privacy or federated learning of cohorts or forest-in-the-trees privacy, the basic concept is generally equivalent: group individual devices and/or people and make them individually identifiable.
Then target by cohort, not person.
Ultimately, however, the main question is: will Facebook and other platforms get enough quality data to keep look-alike audiences as an effective targeting mechanism or not? How effective? And how much more data will be required to make cohorts targetable versus individual.
Plus, of course, if they can … will they enable that feature either explicitly or implicitly? You could imagine a scenario where platforms are able to do this, but keep the “secret sauce” of how it works behind the curtain.
Look-alike audiences are uncertain
But iOS 14.5 and increased privacy are not.
If you’d like some help preparing for iOS 14, SKAdNetwork, and all the massive amount of change the mobile marketing ecosystem is experiencing, we’d love to help.