Background
Building positive customer experience and engagement, in a world overwhelmed with hundreds of thousands of mobile apps, is a big challenge. Customers are broadening the scope of their preferences and the opportunities to nurture customer relationships are becoming more diverse. Hence, more and more mobile apps are focusing on acquiring high lifetime value (LTV) users and forging a bond with their target audiences to stretch beyond just boosting purchases.
The card app vertical has a classic appeal to both young and old alike and presents a huge opportunity for mobile game developers to increase their app engagement in this ecosystem. In this success story, we discuss how you can ensure your app marketing success through leveraging first-party data. With the demise of cookies and without access to the identifier for advertisers (or IDFA), marketers need to look at and better understand their first-party data to provide audience insights that help build and feed machine learning models for their campaigns.
The Challenge
A popular gaming app company partnered with Aarki to run a user acquisition campaign for a well-known card app game, with the aim of meeting and exceeding their key performance indicators (KPIs). Their main goal was to increase brand awareness and attract a wider base of users who regularly visit the app, who actively and frequently engage with its features, and who make in-app purchases. They needed a partner with machine learning expertise, a master of producing engaging creative design across various gaming apps. The campaign targeted users from Europe who owned Android mobile devices.
The Solution
We started the campaign with the Explore phase. This phase is crucial for any programmatic campaign because it helps machine learning models to develop a better understanding of the app’s users’ behavior and preferences and to make more accurate predictions. We analyzed the first-party data shared by the client to understand their target audiences, uncover their needs and categorize the factors that influence their buying behavior. The Explore phase allowed us to collect enough installs and impressions to enable our algorithms to learn.
We then kicked off the Install Optimization Phase with the objective of delivering optimal CPI and establishing volume for event optimization analysis. Sophisticated machine learning models are the key to success for a user acquisition campaign. We developed models based on the client’s first-party data to effectively learn the user behavior and deliver scale at an optimal cost.
Creative Optimization
In addition to our machine learning algorithms, we leveraged our capabilities in Dynamic Creative Optimization which supported the campaign's success. For this, we utilized creative testing tools to determine the most relevant ad variation, such as the optimal CTA, the best performing background, and other elements to maximize the campaign’s performance.
At Aarki, we believe that understanding the needs and wants of the target audience increases the likelihood of the users engaging with the ad. Ad creative is the backbone of our campaigns, thus knowing the optimal ad format for the app and target audience helps us scale our campaigns. The ad formats that succeeded in this campaign were interstitial video and display.
Our deep category expertise and our creative research helped us design relevant ads to attract high-quality audiences and encourage them to use the app.
Results
Machine learning algorithms leverage users’ historical data to target similar audiences and predict the probability of future behavior. It determines whether users’ actions are likely to result in a conversion. We use custom models for each campaign and leveraged two models that proved to be a success for this campaign. As the campaign ran, increasing amounts of data points were collected. After the data collection process, both models improved their learning and accuracy.
Predicted Probability vs Install Rates
Testing of the machine learning models allowed us to continuously drive high install performance. After optimizations, the second model tested provided a more stable prediction for install probability, thus enabling a more consistent delivery against the required KPIs.
Performance Result