• by Mark Torrance , Op-Ed Contributor, September 15, 2015

First-party data has never been more important to marketers, and there's a rush to leverage technology to make it actionable. This data includes online signals from tags on a website or app SDK, as well as data from CRM systems or offline marketing lists activated through use of a DMP. But too many marketers are using this data for retargeting in naïve manual ways, missing out on most of its potential.

Marketers rely a lot on retargeting — it drives great performance even when used in a simplistic way. However, in my opinion a lot of retargeting spend is wasted on prospects without substantive interest or who show behavior that suggests purchase intent where none actually exists (e.g., visited the site in error or stayed for a short length of time).

I think there’s a lot of variability in the quality of retargeting data and have found that the highest performing retargeting inventory may be 10,000 times more effective on conversion rate, and hundreds of times more effective on cost per action, than the lowest performing sets of retargeting inventory. But most companies are blindly buying it all, and paying the same price for it, regardless of its value.

While retargeting usually performs well, for most campaigns, good machine learning models can identify lots of prospecting traffic that performs just as well as all but the very best retargeting. I can see this in the overlap between the distributions of retargeting and prospecting traffic when measured on performance. This is really exciting: It means marketers can get great retargeting-like performance combined with the expanded reach of new customer acquisition.

Of course, the best results come from combining first-party data with detailed third-party data. But it is frankly impractical to do so in the usual way, by manually constructing "composite audiences" using Boolean combinations. Unsurprisingly, I think marketers need a technology partner — one that can leverage machine learning models to automatically identify high-performing combinations of data, and to determine their relative value.

With machine learning, non-intuitive links can be uncovered automatically. For instance, people who have just booked a vacation are more likely to buy other products in the four to six weeks leading up to the trip. By simply adding the behavior of consumers who have recently booked travel online to a long list of other attributes used by machine learning models, I believe retailers can drive higher campaign performance.

Many DSPs build models to help identify and extend valuable audiences, and use those models to place people into segments. But this is just the first step in extending the value of your audience. The same customer may be much more or less valuable depending on the context — such things as recent activity, current website or app, time of day, weather, and myriad other contextual factors. By blending this contextual data together with behavioral, demographic and first-party audience information in a single model, you can effectively score every potential moment of engagement, moment to moment constantly. This means that every subsequent moment has the benefit of the learning from every previous moment. In my view, this is the only approach to leveraging the combined power of a DSP and a DMP, of first-party data combined with third-party data.

If you have not yet taken the leap, the best move is to jump all the way to combining that first-party data with machine learning and behavioral, contextual and third-party data for optimal marketing performance.