The problem with mobile user acquisition, Part 1 of 2: The Law of Large Numbers

Part Two

When a marketing manager buys a user from a mobile ad network, he knows exactly three things about his purchase:

  • The purchased user has a mobile phone and has installed at least one app;
  • The user’s mobile phone model and geographic location;
  • The developer of the app from which that user came was willing to sell that user for $X.

And, depending on the network:

  • Demographic information about the user, such as gender.

That’s not a lot of information; considering the role subjective taste plays in whether or not a person likes an app (especially a game), mobile acquisition purchases are essentially made blind.

Which says nothing about a user’s predilection to make in-app purchases: if the odds of a person selected randomly from a huge population enjoying an app are small, the odds of a user enjoying an app and purchasing items within that app are infinitesimally small.

I believe these odds are effectively unpredictable for a single user: when a mobile developer acquires a user, he cannot predict how much money that user will spend in his app. It’s random – or, put more precisely, the determinants of a user’s lifetime in-game spend are independent stochastic variables:

  • That user’s current financial status (and expected near-term future financial status);
  • That user’s preferences in terms of game genre / app category;
  • The amount of free time that user currently has (and expects to have in the future) to play games;
  • The extent to which that user may find purchasing virtual goods socially acceptable and financially responsible.

No mobile networks filter for these variables because these variables are not measurable (the last variable can be measured by proxy of past spend, but this data isn’t available when a user is purchased from a network).

These properties of mobile user acquisition beg the questions: can user acquisition be optimized and can LCV by channel be predicted?

And most importantly: Is LCV by channel a knowable metric?

It is, and I do believe certain channels provide appreciably higher-LCV users than others – but no individual developer has the capacity to calculate LCV by channel with any precision. Why? Three reasons:

  • Lack of Data. The sample size of users acquired by most developers is too small to draw any meaningful conclusions about the properties of the total population of users sold by that network. And we can’t use statistical methods (such as the t-test) to evaluate the randomness of a sample because we know nothing about the general population.
  • Constant industry evolution. Mobile ad networks are in constant flux in terms of the users they sell, and those fluctuations are independent of past states. Even if a mobile network could authoritatively say, “The average LCV of the users in our network is X”, that number would mutate over time in a matter independent of / unrelated to that current value.
  • Lack of expertise and motivation. Only the very largest developers command the economy of scale necessary to build a system that could even approach the full utilization of acquisition data in predicting LCV by network. To achieve positive ROI on such a system, a developer would need to be purchasing tens of millions of users per month across a large portfolio of apps and games. That developer would also need a large data science and BI team to maintain and usefully implement the insights produced by that system. The development time of such a system would be years and the development and maintenance costs would be monumental. Perhaps as some of the Asian gaming behemoths grow, they’ll implement such a system – but of the public mobile gaming companies, I don’t believe any has the mobile revenues or DAU numbers to justify having shouldered such a massive expense in the past year.

The mobile networks could ameliorate this problem with mobile user acquisition by publishing more information about their pools of users; they won’t, because this information asymmetry fosters a Heads I win, Tails you lose dynamic with developers.

If a developer buys traffic from an acquisition network and it’s bad, that developer has no choice but to buy more traffic (or their game languishes); if the developer buys traffic from an acquisition network and it’s good, the developer is incentivized to buy more. When the quality of a cohort of users is randomly distributed – not only within one network but across all networks – and not evaluable, a developer’s own micro-perspective of the marketplace doesn’t afford many options.