Predicting the unpredictable – Complete iOS measurement using predictive analysis

By Michel Hayet
Predicting everything in iOS - square image

Our predictive analytics solution has come a long way since we initially started working on it in late 2019. 

The product’s initial purpose was to save UA managers precious time when making their optimization decisions, by providing them with fast predictive insights — eliminating the uncertainty that comes with campaign optimization. 

In June 2020, Apple’s announcement of their intentions to launch SKAdNetwork, introduced the mobile ecosystem into a new privacy-centric era — that brought about an entirely new set of challenges.

Predict was identified as the key to overcoming many issues the market faced following these changes. Fast, anonymous insights, providing measurement clarity based on a limited measurement window — was now the main task at hand, and our sights were set on making this happen.

Our latest product version introduced UA managers with a solution that delivers aggregate predictive Day 30 insights in the form of ROAS, ARPU, retention rate, and percentage of paying users. All while seamlessly communicating with the SKAN protocol.

So what was next for us?

We set out to provide accurate performance predictions across all traffic types.

With SKAN predictive measurements already running, the most challenging aspect of iOS traffic was achieved (and is being optimized with every day that goes by). 

From that point, the next logical milestone was providing predictive insights for all iOS traffic types.

What does offering predictive analytics to all iOS traffic actually mean?

What does offering predictive analytics to all iOS traffic actually mean?

In the iOS 14+ era, iOS traffic can be broken into several channels:

All of the above required different measurement methods to be in compliance with the different privacy and measurement constraints relevant to them.

These measurement types can be addressed under two main attribution methods:

SKAN only

Mobile app campaign attribution that strictly relies on SKAN postbacks.

MMP only

Where attribution is performed either through ID matching for consenting users (who have allowed their user-level data to be shared with other companies), or through probabilistic modeling across all users.

Generating predictive insights for each attribution method required adapting the product to each solution’s specific requirements and constraints, which was a challenge to say the least.

But for us, providing advertisers and media partners a full picture that takes both methods into account was a must, simply because neither could really manage and optimize their activities without complete and comprehensive data at hand.

How does it all play out?

Predict mode in the Conversion Studio

Once the SKAN integration structure was in place and field tested successfully, it was time to introduce Predict into AppsFlyer’s Conversion Studio solution.

The Conversion Studio provides advertisers with maximum flexibility to make each of their six bits count. You determine precisely what measurement data to encode within the conversion value, which includes the activity you’d like to measure and the post-install measurement window length.

Predict is now one of the measurement modes available in the Conversion Studio, alongside the existing Revenue, Conversion, Engagement, and Custom modes.

Predict in SKAN conversion studio
Predict in SKAN Conversion Studio

Within the Conversion Studio, Predict utilizes 5 of the 6 available bits to produce predictive insights for Day 30 KPIs. 

While other Conversion Studio measurement modes offer actual indication of events measured during the 24-72 hours window, Predict provides indication for events and measurements that would likely occur beyond that time frame, based on advanced machine learning analysis.

This provides great value for advertisers that base their optimization logic on these later events (e.g. subscription, purchase, advanced levels, etc.), but lack this down-funnel indication under the existing SKAN structure. 

An upside of using Predict through the Conversion Studio is that the sixth remaining bit can be utilized for gaining additional indication for events taking place during the activity window, alongside predictive insights. 

Integrating Predict into the Conversion Studio meant that our SKAN alignment was ready to go. 

Predict beta advertisers can now enjoy full transparency into their conversion value schema mapping, and allocate the additional sixth bit according to their needs — all from the same dashboard.

AppsFlyer attribution

As mentioned earlier, AppsFlyer attribution can be achieved through either ID matching for consenting users, or probabilistic modeling across all users.

Applying predictive modeling to AppsFlyer attributed installs, makes it even more accurate than before.

Predictive insights are still delivered within 24 hours, as the strict SKAN work model and privacy constraints still apply. However, the predicted KPIs of pARPU, pROAS, pRetention rate — are all available with the predicted value itself, replacing the value ranges used in SKAN.

With the sixth bit’s limitation removed, there’s no need to fit these values into predefined ranges, enabling better accuracy for each predictive insight. 

And, in addition to eliminating the need to translate predictive insights into SKAN conversion values and communicate them to the SKAN protocol, the process itself also becomes easier. 

Aggregate, anonymous predictive KPIs are simply associated with the relevant media partner, and clearly presented on the dashboard. 

Single source of predictive insights

While operating with both attribution types enables covering the entirety of iOS traffic, this also creates the risk of duplicating attribution records.  

These duplications occur when installs are attributed through both AppsFlyer and SKAN.

AppsFlyer’s Single Source of Truth (SSoT) solution addresses this risk head-on, in an effort to present advertisers with an accurate view of their iOS attribution data.

All attributed installs go through a process of deduplication, meaning that in cases where both SKAN and AppsFlyer attribution records are found, AppsFlyer attribution will be the only one recorded.

When presenting predictive insights, keep in mind that this logic should also be applied, as a prediction is generated per each and every install, and we don’t want to create duplicated predictions that might change predictive KPIs for a specific cohort of users.

Another important aspect to take into account is the fact that organic installs are out of SKAN’s scope. So, in order to accurately measure the predicted behavior of an organic cohort, SSoT must be included. 

Organic users now represent the users who weren’t attributed by neither AppsFlyer nor SKAN, making them an important piece of the iOS performance puzzle.

With only five of the six available bits utilized for producing predictive insights, an alternative (and recommended) usage of the sixth bit would be for distinguishing between users attributed by AppsFlyer and ones attributed by SKAN — in order to present a single source of attribution reality.

Predicting everything under the iOS umbrella

The ability to integrate between different AppsFlyer solutions that were constructed around the new iOS landscape — enables Predict to produce predictive insights across all attribution methods.

A simple dashboard toggle can take you from AppsFlyer’s attribution predictive insights, to SKAN insights, and into the duplicate-free SSoT view of your data.

Predicting everything under the iOS umbrella

Predict’s business value remains the same — providing you with accurate predictive insights. The only difference is in the attribution method for which these insights are delivered.

Marketers can now leverage early foresight into key LTV elements like ARPU, ROAS, retention rate, and percentage of paying users across all of iOS — to help make their optimization decisions faster and more accurate than ever before.

Shifting between Predict iOS modes
Shifting between Predict iOS modes

Poor performing campaigns no longer have to run for a 2-4 weeks learning period, wasting precious marketing budgets while marketers are forced to wait for clear LTV results, only to be paused too late.

Good performing campaigns can now be capitalized early on and generate more value per each dollar invested — starting from day one.

And now that iOS’s predictive insights have been made available, our sights are set on the next frontier — predictive insights for Android.

Stay tuned for more to come.

Michel Hayet

A former digital entrepreneur, Michel has experienced all aspects of the mobile marketing ecosystem. Studying the intricacies of the digital advertising space, Michel explores technology innovations surrounding mobile ad fraud, predictive analytics, and various tech-stack solutions that make mobile marketing safer and more efficient.

Follow Michel Hayet

Ready to start making good choices?