Performance Marketing + Mobile RTB + Machine Learning = Opportunity

Humans shouldn’t be buying ads any more than they should be driving cars.

Mobile RTB is great for marketers: it enables retargeting; fairer pricing; more targeting data; more control.

But marketers don’t want control. They want performance.

Buying mobile RTB ads exposes hundreds of variables to the performance marketer — variables like publisher placement, device type, time of day. These variables have fundamental impact on user acquisition rates and the cost of the media. Media buyers adjust buying to maximize conversions and minimize ad spend. People’s entire jobs— and an entire industry, you might say — exists to do this planning.

But noticing the patterns behind these variables, and learning how they affect conversion rates, for billions of individual users? And then optimizing spending with the goal of maximizing conversions and minimizing cost? Computers should do this.

RTB’s real opportunity is that it allows a new kind of ad buying, where machines figure out the types of users who are likeliest to convert on your offer, and buys more users like them. That’s the idea behind AdWill.

AdWill = performance marketing + mobile RTB + machine learning. Let’s go into each of these themes.

Why it matters: Performance marketing

AdWill = performance marketing + mobile RTB + machine learning

If marketers know what they’re willing to pay for a new user, software can to go and find them.

AdWill includes a mobile demand-side platform, or DSP. We integrate with mobile ad exchanges, and buy individual ad impressions from them.

But unlike most DSPs, human marketers don’t optimize AdWill campaigns at all. They just give us creatives, offer links, and how much they want to spend per conversion.

A conversion could theoretically be any event the user takes after seeing an ad, as long as it generates a notification to our servers (just a simple postback URL through something like HasOffers/Tune). Conversions could be anything from clicks, email submits, to in-app actions. But most of our clients are mobile apps who are looking for new users to install.

(If a DSP is charging percentage of spend, they are incentivized to bid higher and waste your money, especially if they don’t have enough inventory to sell you. We try to maximize conversions for you, and minimize our ad spend cost, which means we bid lower, and target the users likeliest to convert.)

Why it works: Mobile RTB

AdWill = performance marketing + mobile RTB + machine learning

RTB gives any individual DSP access to to huge amounts of inventory. We process a few billion ad impressions every day, across the world.

Scale is perfect for our model. For example, users’ responses to ads are unique for each publisher, and there are hundreds of thousands of publishers out there. We can compute, for each publisher, its exact effect on conversion and media price, along with how it interacts with other variables.

RTB also allows user-level precision. Every single impression we see is either tagged with a unique ID, or unique IDs can be approximated with fingerprinting. We can keep track of individual users for things like session frequency.

Why it fits: Machine learning

AdWill = Mobile RTB + performance marketing + machine learning

Machine learning is a field of study that gives computers the ability to learn without being explicitly programmed.- Arthur Samuel, 1959

The coolest part of AdWill is the decision engine. For each offer that we run, and for each creative, the system learns how different users respond. It’s constantly updating its buying algorithms with new data from how users respond. With software integrated into the bidding algorithms, it can learn patterns from the data, and bid higher on impressions that it thinks are likelier to convert. All of this is made possibly by machine learning.

Machine learning can’t tell you definitely whether an individual user will convert, but it can tell you that this user is 7x more likely than normal to convert. Hence applications like spam filters, fraud detection, and even predicting seizures. It’s about learning patterns from large amounts of data.

How do you know how much data to collect before you make a prediction? Actually your predictions aren’t single numbers. You learn estimates around a guess, and make it narrower and more confident as you get more data. Estimates are bayesian, and in probability distributions. We can sample from the probability distributions on each impression.

The beta distribution. Helps answer questions like, “If we’ve seen 5 true samples and 1000 false samples, what is our estimate for the rate of true samples?”

I (Aarlo) have spent the past 2 years trying tens of machine learning algorithms. Finally, after a few different breakthroughs, the current algorithm generates tens of thousands of conversions every day and looks at a few billion impressions. And there’s a lot more to do ☺