Make SKAN work in the real world: 15 tips for specific verticals and monetization models

How do you make SKAN work in the real world? Well … you can start by watching this video. Hit the play button and keep scrolling …

https://youtu.be/gddm7w6h3pI

Apart from SKAN on iOS, the measurement options are challenging. You might try to use modeled ad network data for campaign measurement. You might use media mix modeling to attribute advertising results. You might still be trying to use fingerprinting. You might be using first-party data and logic. In fact, you probably should be using many methods. 

But on iOS, SKAdNetwork is the only option that is deterministic, ensuring that (under certain conditions) you will get postbacks for app install attribution, and ad networks will get the data they need for campaign optimization. So it makes sense to use it also, and to learn to use it well.

The problem is that to make SKAN work in the real world, there are plenty of challenges:

• Choosing the right model
• Selecting the right key measurement events and thresholds
• Encoding enough information into the limited space in your postbacks
• Designing the right app experience to get quick feedback via SKAN on results
• Managing ad partner and campaigns to avoid losing too much data to privacy thresholds (SKAN 3) or crowd anonymity (SKAN 4)

So I asked 2 of Singular’s smartest SKAN whisperers to spend some time with me on making SKAN work in the real world. Both of them have worked with literally hundreds of clients to help them tweak their settings, models, and apps and extract the absolute maximum amount of information from SKAdNetwork.

• Victor Savath, VP Solutions Consulting
• Nabiha Jiwani, Customer Success Team Lead

Here’s what they told me …

1. Realize it’s not 1 and done

Under IDFA, you could collect everything and figure out what you need after. Under SKAN, you need to be much smarter in selecting the right measurements.

But you’re not smart enough. None of us is. Realize you’ll be iterating to get it right, and realize you’ll be iterating more to get it righter. And furthermore, realize that as the world changes — Apple switches something, a partner changes something — you’ll be iterating yet again.

“There needs to be a mentality and understanding that it’s not a one-and-done exercise,” says Savath. “We see iteration as part of the core philosophy of approaching SKAdNetwork. Not just because your KPIs change or your product changes … the ecosystem changes.”

2. Understand the goal

Do you mostly want measurement for internal teams? Do you specifically want optimization for ad partners? Is ROAS the key metric you want, or is something like the customer journey more important? What does the product team need, specifically, versus the growth team?

“With my clients specifically, I’ve seen a deeper focus on how partners can understand and read the SKAN data and make sure that those events connect,” says Jiwani. “The second layer would be, okay, let’s focus on revenue.”

SKAN is technology, but you make SKAN work first and foremost by having a clear strategy.

3. A big benefit of ad monetization measurement models under SKAN

Ad monetization has a massive advantage over monetization models like subscription when using SKAN, because feedback is far quicker.

“When I think about conversion models more holistically, I’m always thinking about what gives you good early signals and then what can also give you strong predictors of quality user, strong LTV over time,” Savath says. “So it’s often a blend of discrete variables, continuous variables. But in the world of admon, you can actually have both because you’re going to have early signals (ad impressions, generally speaking) so you have these signals with a high level or decent amount of variance.”

That means you can segment users easily just based off volume of ad impressions alone, and you can use those segmentation to make and test predictions about which cohorts might be more likely to buy IAPs or even subscribe later on, if you have those options.

Note: Singular has had ad monetization SKAN measurement models for some time, and recently added those to our free product tier.

4. Real-world SKAN: compare to Android

There’s a lot of modeling under SKAN 3 thanks to the very quick measurement period. That modeling can be very good, but it’s still modeling. 

So take some of the early signals you’re measuring with SKAN and compare that with your Android data.

“Taking those smaller signals like a session start or some sort of indication there and coupling it with what you know of how your Android users have been performing and comparing those 2 subsets have helped a lot of my clients understand where their iOS data has been going,” says Jiwani.

Plus of course, with SKAN 4 you’re going to get longer measurement periods, which will help considerably. That may not arrive at scale, however, until early next year. SKAN 4 postbacks are currently only around 15% of all postbacks that Singular is seeing, though the trend is rising.

5. Make SKAN work: understand your users better than they do themselves

Clearly you want to use early indicators to predict future behavior. The first step is understanding what is actually happening in your app versus what you just happen to be currently measuring.

Once you do that, you can build your SKAN models off real behavior.

“Instead of just choosing revenue buckets based on, let’s say my average product price tags in an in-app purchase, let’s just look at my IDFV data set and say: what is the average amount generated on day one amongst users that complete the tutorial?” says Savath. “Perhaps you use that as your revenue bucket thresholds when defining a conversion model, because now it gives you a good segmentation group or cohort to observe.”

Watching that group over time — much more time than you can currently measure with SKAdNetwork — gives you good insight into monetization potential for that segment. You can then experiment with early predictive signals that indicate a new SKAN install should be assigned to that segment.

6. Iterate SKAN conversion models monthly when testing

Iterating in the real world with SKAN is a challenge. You can’t do it daily, because you need some amount of time to let SKAN campaigns flush through your ad network partners’ ecosystems. Iterating too quickly will be messy, and provide insufficient data to make smart decisions on.

But how often should you iterate when in testing?

“I’ve had clients change it month over month initially,” says Jiwani. “You might optimize towards tutorial complete, registration, some sort of account creation, right? Changing those initial metrics that indicate whether someone will or will not purchase, or will or will not deposit and then taking that data and then modeling it … changing that first initial event indicator has been really strong with some of our customers.”

Of course, you’re probably not going to continue that cadence forever, but you’re also not likely to keep the same model for a year. 

A quick note: initially when switching SKAN campaigns you had to pause everything, wait 48 to 72 hours, then restart.

Singular now offers technology that makes iteration much quicker and simpler: just switch and go. Singular also offers a product feature where you can simulate SKAN changes without actually making them, and see what your update measurement data would look like.

(Talk to us if you’re not a Singular customer and want to know more.)

7. Check your log-level data to optimize revenue buckets

Don’t just set up conversion value revenue buckets and forget about them. If your revenue buckets don’t match what users are actually doing, you’re essentially wasting bits and failing to maximize information return.

Make SKAN work by checking log data:

“Anyone who has a mixed model or IAP revenue model, you’ll notice in any sort of log-level data … there’s dips or there’s segments in your conversion model that are just unused,” says Jiwani. “Making sure you tweak those buckets … to actually capture that subset of users, or making sure that you either expand or minimize those buckets to ensure that there are no dips, that each actual conversion value is used in that model … will ensure that you’re getting the most amount of data you can.”

Jiwani says she’s done that exercise time and time again with customers, and it almost always results in better data capture.

8. SKAN 4 helps you break cohorts down into sub-groups for higher fidelity

In SKAN 3 you might be able to assign a user to a cohort like did_tutorial, observe later behavior, and set a revenue estimate based on early indicators for that kind of cohort.

In SKAN 4, you’re going to be able to go a lot deeper.

“You have a cohort of users that did ‘tutorial complete’ on day 1 and performed an in-app purchase …that’s operating as one of those 64 buckets,” says Savath. “That’s a user cohort. The beauty of P2 and P3 is you observe that user as they enter the P2 timeframe or the P3 timeframe and you say, oh … that single cohort now breaks into 3 additional groups … users that did ‘tutorial complete’ day 1 and purchased … and then in P2, they did a repeat session. They came back. Or they made another purchase, or they made a huge purchase, right? Now you have 3 groups: that one cohort split to three.”

All of it adds to modeling rigor and your ability to build better ROAS and LTV models based on early indicators.

A quick note about SKAN 4: yes, we still don’t have a majority of SKAN 4 postbacks yet. In fact, far from it.

But with Singular, you can set up a SKAN 4 conversion model, benefit from any SKAN 4 partners you’re working with, and not lose any data from SKAN 3 partners … because it’s backwards compatible.

9. Streaming verticals and subscriptions: yes, you do have early data

Subscriptions are hard under SKAN, especially if you used to have a 7-day trial period. 

But there are always early events you can use for proxies, and in streaming media verticals like music, entertainment, and video, there might be more than most. Make SKAN work by using them.

“If you think about any of your streaming services … you do have continuous variables that you could read signals into,” Savath says. “They’re good early indicators and they’re plentiful.”

Examples:

• How many times did they listen?
• How many movies did they watch?
• Did they stream to a bigger screen, or to an external speaker?

Also, there are segmentation indicators, like what kind of subscription they tested: a family plan, an individual plan, student plan, and so on. All of that gives you more information to build segmentation and look for predictive indicators.

Such as: in the family plan, did anyone else sign up and get added?

10. Multiple monetization methods helps drive better data

If you only monetize via subscriptions, you have one thing to measure, and it’s hard, and often takes longer than SKAN’s available measurement period. 

If you add ad monetization, then you’ve got additional signals that will come quickly and give you more information. And if you add in-app purchases, you’ve got even more information that will help you build smarter predictions and more accurate assessments of cohort value.

Even better, you’re automatically allowing your users/customers/players to segment themselves, and you’re allowing people who don’t want to make an immediate long-term commitment the ability to try before they buy.

11. Retail: first purchase is easy, subsequent value is hard

Retail apps often work fairly well under SKAN for initial purchases at least, because people often download a retailer’s app for a specific purpose, and they pull the trigger immediately.

So the first purchase often happens quite quickly.

The problem is getting adequate measurement for subsequent purchases.

That’s where you have to look at engagement variables and usage variables: sessions, views, searches, add to carts, and more to get a sense of how likely a specific newly acquired customer is to buy more.

A personal note of caution here: I have installed retail apps, purchased nothing, and then made hundreds of dollars of purchases literally months later. You simply cannot assume that if nothing happens right away, nothing will continue to happen forever.

SKAN 4 will help, but not be a panacea for that:

“The unlock with P2 and P3 postbacks is just having those additional signals that indicate someone’s coming back or someone’s making a different purchase, or they’re viewing another item in a specific catalog,” says Jiwani. “That will be helpful with SKAN 4.”

12. Retail: make sure you differentiate between the reporting layer and the partner optimization layer

It’s important to be able to measure the quality of profitability of new user cohorts. It’s also important to communicate the value of new users to ad partners. In retail apps, that often means a mixed SKAN conversion model with measurement for engagement events as well as revenue.

A key tool here: IDFV.

“You might have an engagement and events funnel model,” Savath says, referring to sign-ups, cart adds, etc. “If you have that type of model, Singular’s in a place where you still have revenue reporting because you could say: I’m using that funnel as a way of segmenting my IDFV data set, and I’m going to observe these cohorts and see the actual revenue that they generate over time, and report on those revenue inferences within the Singular reporting interface such that the networks can be optimizing off events based off how the model is configured, but from an analytics or LTV reporting perspective, I can also see the revenue.”

13. Fintech: mix engagement and revenue metrics

Fintech can be tough to measure under SKAN. There are great events to look for, such as account creation, connecting a bank, depositing money, but these are big steps for people to take, and they don’t always happen quickly.

Mixed conversion models, therefore, are the way to go:

“I’ve seen most customers do a mixture of both engagement metrics and revenue metrics to capture users that who initially have engaged with the app, have inputted a certain level of information into the app, have connected various accounts within that app as an indicator of how active that user is,” Jiwani says. “And then at the same time on the revenue front, we’re capturing potentially the amount of money deposited or the amount that’s been used in a transaction.”

Funnels models are more rare in fintech, but there is potential here, she adds.

14. On-demand: you lucky SOBs!

On-demand apps are super-lucky under SKAN: most people who download an on-demand app do it as part of a purchase or engagement process.

Example: you want a ride, you download Uber or Lyft, enter your payment information, and take a ride.

But there’s more to look at to make SKAN work when you want to estimate LTV, Savath says.

“Then you go into the world of continuous variables … revenue amounts are definitely not as common given the high level of variance between trip length duration. So it’s really around the engagement and frequency of utilization.”

More, generally, is good. (Of course.) But it can be misleading too: vacation or business travel users might be very sporadic.

15. Games: hyper casual vs mid-core

Hypercasual games might have been just made for SKAN. Admon and speed are both common factors here:

“It’s almost like it was designed for this use case because you’re talking about users that are engaging what they do within the first day, and then oftentimes the life cycle of hyper was much more truncated,” Savath says.

For mid-core, you need to go deeper:

• What’s the long-term monetization strategy?
• What events predict high-value users?
• Do high-value users watch more ads quickly?

Make SKAN work: so much more in the whole podcast

Subscribe to Growth Masterminds on YouTube and get the audio podcast as well. You’ll thank me later!

Also, if we can help in any way, experts like Victor Savath and Nabiha Jiwani work with customers and prospects every single day to ensure they maximize their marketing campaign values with Singular tools. Sometimes you even get Singular CTO Eran Friedman, perhaps the most knowledgeable privacy-and-measurement person on the planet.

Get in touch, book some time, and find out how we can help.