Part 3: The Aarki Team Talks About Apple Privacy Changes: The DSP Perspective


DSP Perspective of Apple Privacy Changes

IDFA is at the heart of media buying for in-app advertising today, but this reality is going to change with Apple’s privacy changes, to be implemented with the release of iOS14.5. With this, advertisers need to receive the user’s permission through the App Tracking Transparency framework to track them or access their device’s advertising identifier. Though the release of iOS14.5 is just around the corner, there are still questions that require answers. 

We brought you the predictions of MMPs and Ad Exchanges on the impact of Apple’s privacy changes in our latest articles. Now, to complete the trilogy, we interviewed a few members of the Aarki team to give you comprehensive answers to the questions that have arisen on the DSP side. 

Aarki Team headshots, Nicol, Anton, GraceAarki Team headshots, Igor, Kim


Question 1: Will a DSP’s models still work with SKAdNetwork postbacks?

The short answer is yes; the complicated answer is that every buying partner, DSP, Network, SAN has to contend with weaker signals from SKAdNetwork, and incomplete event coverage, especially in the early days of a campaign.

Apple has a privacy threshold on two signals in the SKAN postback:

  1. The source app - where the ad that drove an install was displayed
  2. Conversion value - these values represent the conversion events tracked in the advertiser app post install

Looking at source app, the privacy threshold means that until some threshold of installs per campaign, per source app have been reached the source app of the install will remain unknown to any buyer optimizing with SKAN. The same is true for conversion value and we have seen that these values do require a non-trivial amount of installs per campaign ID before thresholds are met. 

In the case of conversion value, the conversion model and events chosen by an advertiser will also influence how likely conversion values signals are going to be received so it is important that advertisers and their buyers have open and honest discussions about SKAdNetwork conversions measured.

While the data is more delayed and anonymized to protect user privacy, the means of optimization for a buying model do not fundamentally change. A buying partner can decode conversion values into in-app events, and train and deploy models optimizing for these events.   

Another source of data for campaign optimization and models can still come from MMPs. While the scale and utility of the data may vary per advertisers app campaign (and bias needs to be accounted for) privacy compliant postbacks from users with the necessary consent can provide rich optimization data.

The best way to bootstrap SKAdNetwork optimization is to start running SKAdNetwork attributing campaigns at as high a scale as possible before ATT enforcement begins. This early start can provide advertisers and buying partners time to collect and analyze SKAdNetwork conversion data to build optimized models and campaign strategies prior ATT enforcement. 

- Nicol Cseko, VP of Product

Question 2: How do SKAN models differ from can-track models? 

The most challenging part about SKAN attribution for machine learning models is the inability to match each impression to a conversion (install, engagement, purchaser). This fundamentally changes the training set, and different modeling approaches must be considered.

Instead of a classical binary classification dataset, we must aggregate data per SKAN campaign.

IDFAOne way of extracting valuable information from such a dataset is to fit a distribution per each set of features (source_app, skan_campaign_id) and then sample from this distribution to make predictions (Thompson Sampling). The main challenge here is to create the optimal SKAN campaign id as it's going to be the most powerful feature.

Another approach is to build models to predict CTR using all contextual features; in this case, we have access to impression-level attribution. Then, we can model the remaining click to install conversion funnel on aggregated data using approaches outlined above. On one hand, this allows us to capture richer signals when it comes to user click propensity. On the other hand, this approach is susceptible to noise from click-bait, fraud and accidental clicks.

- Anton Protopopov, Data Scientist

Question 3: How will I be able to test new creatives on iOS 14.5 campaigns? 

Creative testing will have to get creative! Today creatives, and even specific versions of a dynamic creative can be measured from impression to install to post install with the post-install behavior clearly attributed. SKAdNetwork today blocks this granular measurement by obfuscating the install from the creative and the ad impression.   

There are still ways to test and optimize ad creative. First, creative optimization focused on top of the funnel metrics like clicks, time spent, and engagement are not impacted by SKAdNetwork and can continue to be used.

For optimizing to install the easy seeming option is to have SKAdNetwork campaign IDs by creative because you will always have an install signal from SKAdNetwork. This, however, may not work for advertisers looking to optimize creative down the funnel.  

The privacy thresholds on a SKAdNetwork campaign ID mean having many campaigns live without the right scale will prevent quick enough learnings on post install events via conversion values and the cost and time to learning could be prohibitive to some. For these advertisers there can be a few approaches:

  1. One way to get around the problem is test your creatives in existing and mature SKAdNetwork campaigns. You could roll in a new creative and track performance changes. 
  2. A more data driven approach could be to split traffic within the campaign itself, such as by geo, and then look at performance by your geo splits to understand creative level performance. For now this works as some geographic data can be derived from the SKAdNetwork postback. Geos need to be split carefully to minimize bias, but can allow for another way to measure post install conversion on a creative level.
  3. Learnings can also be applied from Android campaigns, or users who have consented to in app tracking but it is important to be aware that they aren’t guaranteed to be transferable.  

Getting clear creative performance is something that everyone in the mobile marketing ecosystem has been asking for and Apple seems to be listening. The latest SKAdNetwork version, 2.2, references 3 new ways to describe ads: adType, adDescription, adPurchaserName. These properties aren’t currently used and are not part of a 2.2 postback but we are cautiously optimistic that Apple will continue to update SKAdNetwork to provide meaningful creative level attribution. 

- Grace Oabina, Sr. Manager - Analytics

Question 4: How will you use the SKAdNetwork campaign IDs?

SKAdNetwork provides 100 campaigns for each advertiser app, per DSP or ad network. There are multiple considerations on the optimal number of campaigns to use, given Apple’s limitations on postback utility for granular campaigns. Putting these concerns aside for the moment, there are several ways the budget of campaigns can be utilized.

SKAN campaigns ultimately serve two purposes:

  1. To define cohorts which are useful for reporting and evaluating performance; e.g., to split traffic by ad format or geographic region, which would allow the DSP and the advertiser to make performance-driven campaign management decisions.
  2. To split traffic into cohorts which maximize downstream information gain for training predictive models.

As a result, we allocate campaigns roughly as follows.

  1. Allocate M1 campaigns to the advertiser.
  2. Retain M2 campaigns internally for manual feature exploration.
  3. Optimize the remaining M3 = 100 – (M1 + M2) campaigns for downstream model training.

Of the categories above, 3. presents the most interesting challenge. In a sense, we are attempting to solve a constrained dimensionality reduction problem. We must compress a ~500K-dimensional feature space into a single categorical feature with M3 levels, in a way that enables us to encode the largest amount of information about the underlying samples. Several approaches come to mind.

  1. Decision tree-based. Decision trees split a dataset to maximize information gain, so they are a natural choice for this problem.
  2. Clustering-based. We can learn an embedding of context vectors, and build a clustering of these vectors; this implicitly assumes that similar users within the embedded space exhibit similar conversion patterns.
  3. Neural network (autoencoder)-based. We can jointly learn optimal cohorts and conversion probabilities based on these cohorts.

We are actively exploring and experimenting with all of these strategies.

- Igor Raush, Data Scientist

Question 5: Everyone is talking about SKAdNetwork readiness. Has the delay in Apple’s App Tracking Transparency had an impact on readiness for SKAdNetwork on the supply side? 

To attribute with SKAdNetwork, not only does the app being advertised needs to be ready, the entire supply chain does as well, from the impression opportunity on the publisher side to the exchanges that broadcast bid requests to buyers. 

As a DSP we’ve seen the delay of iOS 14.5 helping in two areas: 

  1. Giving more time for publishers monetizing to add or update the info.plist file that contains a list of SKAdNetwork certified buying partners. Without this list added or up to date, publishers will miss out on future iOS app performance marketing demand, and bidders who don’t get their SKAdNetwork ID out in the market will miss out on opportunities to bid.
  2. Time for more exchanges and SSPs to support SKAdNetwork. For bidders buying via exchanges, the SSP SDK is required to pass the SKAdNetwork campaign parameters and signature from the DSP to Apple at the appropriate time, this means that all SDKs that deliver ads have to support this attribution framework. Most mobile first in-app supply supports SKAN at scale but some other players still lag behind. While they may not all be pure play mobile supply sources, significant mobile performance spend flows through these channels. 

The good news is that on the supply Aarki now sees that 62% of all iOS 14+ bid requests we receive are SKAdNetwork eligible.

Furthermore, looking deeper into the supply landscape, and focusing on where the overwhelming majority of Aarki buys are transacted today for iOS 14+ users, 91% of those publishers support SKAdNetwork already.

- Kim Aquino, Director of Business Development

Aarki is relentless in its work to continue providing the same exceptional campaign performance expected by top mobile marketers in the more privacy conscious advertising ecosystem of the future. To learn more about the strategies that we at Aarki have developed, drop us a message here

 

Topics: Marketplace Insights