We’ve previously discussed in this post why an app-based company might want to automate user acquisition on Unity, Applovin, Ironsource, and Vungle.

The main challenge with these networks is the ability to bid individually on thousands of sub-publisher / geo combinations. This gives the UA manager a significant amount of flexibility but is also an extremely time consuming and financially ineffective process if attempted manually.

Goals of automation of subpub bidding:

The project of automation of user acquisition on SDK ad networks should have three goals:

  • To apply the correct bid at each geo / subpub combination where install data is available to hit the required recoup time horizon
  • Exploration of existing subpubs by increasing bids to gain incremental spend at the required target payback window
  • Exploration of new subpubs in the most cost-effective way

A framework for automation of subpub CPI bidding:

  • Sub-publisher LTV Projection — As a starting point, it’s possible to aggregate user-level LTV projections at the sub-publisher level. The CPI bid is based on LTV forecasts over a specified recoup time horizon for users that have historically installed from a specific subpub. This approach provides a base level bid but produces noisy estimates and typically, subpub sample sizes are too low for this to be an effective strategy in isolation.
  • Sample Size Techniques — To account for low sample sizes, one should use a bidding algorithm to produce subpub bids based on the predicted LTV from installs through that subpub backed up by prior data from larger groupings (e.g. app publisher, bundle ID (across networks), campaign, geo, channel, or any other relevant dimension). For subpubs with a low number of installs, these robust estimates can look much different than the “raw” average pLTV but offer a more accurate estimate of LTV for future acquisitions.
  • Optimal Bid Exploration — the addition of exploration / “bandit” logic into the algorithm allows gathering more data on publishers where no or very little data install data is available to find potentially untapped value
  • Spend Recoup — it’s necessary for the bidding algorithm to output CPI bids that on average ensure all installs recoup the marketing investment over a certain time horizon (eg: 100% recoup at d180)

Automation of bidding on SDK networks can be a challenging project, but with user-level LTV forecasts and a robust bidding algorithm, it’s possible to maximize spend whilst hitting a required payback window. This can unlock significant incremental spend and valuable returns, whilst delivering a key component in automating a portfolio of user acquisition channels.