RICH STALZER: When I joined Motricity about a year and a half ago, the writing was on the wall that our core business, providing portals and storefronts for companies like AT&T and Verizon, was going away. We needed to pivot the company in a different direction. Before we built anything, Dave [Castillo] and I went to agencies across the country asking them what are the issues they’re struggling with? What we kept hearing is mobile is exploding, but there’s no way to target ads on mobile devices. And so, Dave and I went back to Phoenix and built the Voltari-Connect platform, which essentially targets off four data points: the time of day, location, device type and content.
We take those four data points and we put it into our propensity modeling. We’ve got about 40,000 attributes, and then we make a prediction of what ad to serve to which person at that time.
The Motricity [restructuring] was officially done as of June 30. We’re now Voltari, a digital media and marketing company with 10 years of experience in mobile. Before Motricity, we owned a company called Generation 5, which is a data analytics firm that’s been around for 15 years for direct mail customers. We believe that having a direct mail background is more valuable in mobile than having a desktop background.
DAVID CASTILLO: We wanted to leverage some of the technologies and learnings that we have from our carrier business and a couple of acquisitions around email and online activity. And so we approached it from a very data driven perspective and built propensity models. With our direct mail background, we understand the value of first-party data, and so we integrate with first party data without revealing any PIIs [personal identifiable information].
Most people implement a segment-based approach and we believe segments are useful for describing profiles of people, but we believe that they’re too coarse-grained to take advantage of the power of mobile. The platform has been active since the end of August 2012 and we have over 150 customers.
Can you give me a use case example of how your platform works?
DC: One of the key pillars of what we do is use media efficiently. We don’t waste impressions. If we predict that the level of engagement is going to be low, we’ll save that impression for another opportunity where we can predict a higher level of engagement.
The way that we do that is when we start a campaign, we’ll consider the demographics and some of the inputs that the agency or brand will give us. If they insist on using that as a starting point, we’ll do that, but what we’ve typically found is if we can do a contextual relevancy or look-alike model, we can take the data, aggregate it from all the events and transactions that have happened in our system and create a “cold start.”
For example, we were asked to help with the targeting for a campaign for Texas Chainsaw Massacre last year. The agency told us that the demographics were 18-year-old males. We have a real-time audience profile and by the third day or so, the engine showed a shift towards females ages 24 to 35. It was interesting that after the first weekend of the 10-day campaign, which was about 4 or 5 days into the campaign, Yahoo did an exit poll on the audiences that came out of this movie in southern California and it matched what we had predicted.
The same theatre later came back and asked if we could work on The Last Stand. Even though it’s a different genre from the Texas Chainsaw Massacre, there were still some learnings and propensity models that we could leverage to make this campaign more efficient. The Texas Chainsaw Massacre campaign wound up having a click-through rate of about 0.9% and the other campaign had a CTR of over 2.0%, which were quite good.
What is the accuracy rate of your predictive modeling?
RS: When we run a predictive model, we have a 20% holdout that we grade against the model and so we can see the deviation of the real data from the predictive data. The holdout curve should track the predictive curve if it’s a good fit and almost overlay it. Usually nine out of 10 times, those curves are exactly alike.
In other cases, we’ve had bad models and that means there’s not enough data there from the sample you’re taking. To fix that you can link to another campaign for more data, even if it’s not data specific to your location, but just to have more data so you can make the model a little stronger.
The other thing is, when we look at impressions, it gives us a million impressions to target an audience but we don’t just say, let’s serve 200,000 impressions for five days and hope we reach the target. These models break down the chances for a match all the way to 20 decile points. If an impression falls within one decile point or three, that’s considered a high value asset. So we’ll serve an impression there. If it’s below the curve and around decile points 5 or 7, we don’t serve it because there’s a low propensity for the customer to engage.
What kind of criteria are you using to make sure you have a match?
RS: What constitutes a propensity model are a number of things: demographic data, your click behavior, location, the time of day you’ve responded. It’s a multivariate process that looks at location, demographic and psychographic data and what we’ve tracked in terms of earlier transactions. All that is considered in models that generate these score cards that are used in real time to figure out which ad campaigns to match to which users.
What lessons or technology have you taken from your past business?
RS: On the AT&T storefront, we built a recommendation engine, so we were looking at 100 million unique AT&T customers, and 75,000 different products. When you went onto AT&T’s site and downloaded a ringtone, we used our engine to make a recommendation to serve the right ring tone recommendations to the right person. We tracked this and saw the CTR dropped but the revenue per user went through the roof. In many ways, fewer clicks may be better because it’s going to be more efficient. That was the genesis of the [ad targeting] platform that we built.
Another thing to note is that all this is automated. What we see in other companies is an army of people on the backend, trying to move inventory around to get a better performance. We’re seeing anywhere from 10% to 25% of staff optimizing on the backend.
With us, it’s all machine learning. We don’t have anyone on the backend saying we’ve got to get the performance up, how do we do that? Our system is learning constantly. With every impression, whether you click or you don’t click, the system is learning and getting smarter. We believe that with real-time bidding and programmatic buying that margins are continuing to be compressed, so everything we’re doing is around automation and making things transparent.
So you don’t see a need for data scientists at all?
DC: All the adjustments are made by software, but we have a small team of data scientists who are responsible for building those algorithms and implementing them. We continually refine things as needed, but while we’re running an operation, it’s all hands off.
Is the purpose to keep your rates, such as your CPMs, low?
RS: We could go low if we wanted to, but we’re not competing on that. We’re competing by driving relevant advertising, and down the road there could be a lot of cost benefits to doing things this way because we’re able to take out a whole layer of the organization and that can be put into data science for building out better algorithms or putting more salespeople in front of customers.
Are there any plans to build your own ad exchange?
RS: There are a lot of different ways we could take the platform. Our guiding mission when we look at acquisitions or products is what we call our P&L vision—personalization and localization. We’re striving every day to make that ad or that marketing experience as personable as we can and localize it. Inventory is a big part of it and we’re working with exchanges, ad networks and directly with publishers. We’ve cast a wide net to evaluate and test different pools of inventory, but right now that’s enough for us.