Over the past couple of weeks, we’ve shared a proposed solution for probabilistic attribution on iOS14. It may surprise some observers of the mobile marketing ecosystem to learn that we’re building an attribution model. In this post we wanted to discuss why we’re building this solution, and the benefits it will bring to our clients.

Data science problems require data science solutions

Deterministic attribution, or the task of matching an install to a marketing campaign, is a data engineering problem which has historically been solved by mobile measurement partners (MMPs) in partnership with ad networks. The Identifier for Advertisers (IDFA) provided a persistent ID that enabled MMPs to reference time/date stamps on clicks and impressions. They could then assert which ad network was responsible for the final ad engagement before the user installed an app.

As users upgrade to iOS14, access to user-level attribution, downstream revenue, and in-app engagement events will be significantly degraded. This poses a data science problem of probabilistically attributing installs back to a campaign. AlgoLift is perfectly positioned to solve this problem.

AlgoLift has data science at its core. We’ve been using statistical models over the past four years to accurately predict user-level lifetime value (LTV), and to understand the future behavior of user acquisition campaigns on leading ad networks. Data science is in our DNA. Probabilistic attribution to power automated user acquisition is a problem squarely within our capability and scope of responsibility in the ecosystem.

User-level LTV models

Determining the correct campaign membership probability to an install is the key problem probabilistic attribution can solve. We want to calculate the probability that an install came from one or more campaigns.

An example distribution of an app user over multiple user acquisition campaigns based on probabilistic attribution.

In the context of probabilistic attribution, user-level LTV models are superior to cohort models for accurately attributing predicted revenue to the correct campaign. Cohort models aggregate valuable signals like individual users’ in-app spend and engagement behavior. There is no way to accurately break down a cohort LTV prediction to understand the distribution of revenue or in-app engagements within the cohort.

A cohort LTV model presents a significant challenge for probabilistic attribution: It’s impossible to accurately apply campaign membership probability to whole cohortsr and use it to distribute projected LTV across campaigns in any meaningful way. Ultimately, this would produce inaccurate campaign return on ad spend (ROAS) predictions as well as suboptimal user acquisition. User-level models offer a flexible solution to this problem, powering accurate campaign predictions based on probabilistically attributing individual users to multiple campaigns.

Access to data

AlgoLift is fully aligned with app developers in their efforts to achieve long-term ROAS targets. We have no bias towards any ad network or platform, and this neutrality is reflected in contractual agreements with our clients.

To help train our user-level LTV model, we ask clients to share all their attribution, revenue and in-app engagement data with us. Our Software as a Service (SaaS) business model doesn’t penalize clients when sharing non-personally identifiable in-app engagement data. This ensures our models achieve the best level of accuracy. Our clients are already sending in-app engagement data, which we will use to power the SKAdnetwork conversionValue.

Current clients

Lastly and most importantly, we have current clients who depend on us for a solution to the problem posed by iOS14 and SKAdnetwork. We optimize hundreds of millions of dollars of user acquisition spend per year and our clients rely on our partnership to navigate this turbulent time.

Next steps

Today, businesses throughout the world depend on us to accurately predict revenue from their advertising campaigns. We continue to test our probabilistic attribution solution using historical deterministic attribution data from our clients. This provides us with an accurate benchmark for our probabilistic solution.

We are well-poised to leverage our existing data science capabilities, user-level models, and data integrations to sustain revenue attribution, ensuring our clients transition seamlessly into this new paradigm. This will allow each to maintain and grow their businesses in the most informed way possible, while less sophisticated competitors pull back spend due to an inability to adapt to the new world.