Sramana Mitra: Before we get to use cases, let’s set the context of how you are applying big data issues in your business.
Martin Smith: In our business, every time we serve a piece of media or every time we measure a side activity, we create log records of that activity. On a weekly or monthly basis, we are pulling back billions of records of data for our clients. From a data management perspective, we have been moving away from how we used to do it, which was highly aggregated summarized data, to a much lower aggregation of the data. This means we can handle and develop much more significant insights. What big data has done for us is open up opportunities at a very granular level of analysis for interpretation and recommendation to provide significant new insights into how to look at media. That has been phenomenally exciting.
SM: Let’s do some use cases of how your customers are using your big data capabilities.
MS: Our customers are using it to answer four principal questions around media analytics. The first is, “What is happening?” the second, “Why is it happening?” the third, “What is the best that can happen?” and finally the fourth, “What is going to happen next?”
What we are able to do because of our first-party framework is get better insight into not just what is being consumed, but also into the nature of the relationship between advertisers and consumers at any particular point. We understand what we call state and what is the most appropriate message to say to that person.
If I go into the use cases, we have evolved a model that breaks down into three areas. The first is measurement involves understanding efficiency and the impact on standard third-party measurement. We spend a lot of time analyzing at a granular level the different measurement paradigms that are available in the marketplace and looking at them in different types of ad campaigns: mobile campaigns versus desktop campaigns, re-targeting campaigns versus premium awareness campaigns, etc.
What we understand is that there isn’t a one-size-fits-all measurement model anymore. There isn’t a single model you can apply and expect to give you the correct metrics. What find that because of campaigns and because of the way the browsers are handling data, especially between mobile browsers, you have to eventually provide a custom set of measurement for every single advertiser, both for frequency and measures of behavior. That is the first step. There are very large campaigns where we see the measurement is off by a factor of 12. So, we are talking about significant and highly impactful metrics. What that kind of measurement tells our clients is that the media investments they are making need to be tuned and validated to ensure they have the right media strategies. This is part of the framework we call big measurement, which is our take on how to leverage big data in a meaningful way. Big data itself is phenomenal. We now have essentially open source software and hardware that allows us to manage data in new ways.
However, this ties back to what happened in the CRM stage in the last decade. In 2003 Gartner came out with a study which said that they find it hard to show the value of a lot of the CRM implementation. The risk of big data is that the advertising marketing and media entertainment companies will rush to build data structures without understanding how to handle the measurement aspects of large-scale data operations. TruEffect published a paper on big measurement early this year. We regard measurement as a fundamental and highly impactful area.
The second area of the model is understanding the efficiency in the media and dealing with the reason things are happening. One [way to do this] is to look at audience behavior, and [another] is to look at a whole set of different data and values. This is where we are starting to understand not only how effectively are people buying, but how much better can it be. We ask ourselves: How can we optimize media to increase the outcomes? If you are spending a million dollars, how can you get three million dollars of value out of it? Finally, we are going to the predictive area by giving an effective, well-tuned model for media. How can you do better, and how can you improve the outcomes once you have that well tuned?