Machine Learning Lessons from Google

Machine Learning Lessons from Google

Love it or hate it? Google’s Universal App Campaign (UAC) product is nearing a year since its launch and growth marketers are still trying to decide if it’s worth a try.

I’ve personally spoken to many devotees of the product who have shifted significant portions of their Facebook budgets over to UAC. Many, many more conversations have revolved around horror stories or fear of testing it at all because of such cautionary tales.

At the core of the UAC product is Google’s algorithm – the machine learning that collects data over time and uses it in decisioning where to place your next ad. As a machine learning company ourselves, we have an opinion on UAC, its strengths and shortcomings, and why it may not work for every app (yet).

When it comes to machine learning, Google is top of class. They have more data and manpower behind their product than anyone else on the market. So why can’t all apps succeed there? It boils down to the data you give them. The algorithm needs a significant chunk of data (read: budget) in order to learn what works for a specific case. Though it is marketed as a one-size-fits-all solution – the word ‘universal’ is right there in the name – it is actually not well suited to smaller apps with a modest budget to spend.

UAC promises to optimize toward in app events. The potential pitfall here is selecting an event that doesn’t occur often enough to gather sufficient data. A phrase that has been repeated at growth conferences is that “an install is a proxy” meaning to say an install represents a user in the best way we can, but an install is not a user. True enough, but the fact is we need a proxy. You may want to optimize toward users registering, reaching level three, taking a photo or adding to their cart. Putting an item in the cart, for example, is an excellent proxy for purchase intent. If your app has hundreds of people per day adding items to the cart, that’s a good goal to optimize to. If you have ten people a day doing this, you need to pick something broader, or further up funnel as a proxy.

According to Google, an event that 40 or more users complete per day is a good baseline. So even if it isn’t your end goal, choose an event that occurs more commonly, and optimize to that. Consider starting with the install.

Another related pitfall is giving up too quickly on your UAC campaign. Even if you want to spend slowly, keep things live long enough for the machine to learn. Your CR will not start out stellar, but give it the chance to get there. Pause if you aren’t seeing any improvement day over day. If you don’t have the budget to do this, UAC might not be right for you.

Finally, outside the realm of machine learning, a common complaint about UAC is lack of transparency. For a company that open sources even its ML algorithms, transparent reporting seems like it should be a given. Concerns of brand safety, especially considering Youtube and other UGC are a real issue. Of course, you can count on full transparency from LucentBid! We’ll optimize for your event goals on a risk-free, performance basis with complete reporting of your placements.

Leave a Reply