The whole purpose of DSPs is to wring as much efficiency out of the ad-buying system, limiting costly human input in favor of as much automation as possible. However, since each advertiser has unique goals, there is still a fair amount of work that humans must do -- such as establishing rules that define the bid/buying criteria, and adjusting campaigns to identify the best-performing creative units and/or most responsive audiences.
Contrary to popular belief, DSPs are not facilitating automated programmaticbuying. They just take some of the pain out of manual buying. At many points, they still require a significant amount of human intervention. For example, let's say one of your clients decides that targeting by Zip codes is important for their campaign. A DSP can only set up specifically selected Zip codes as a line item to buy millions of matching impressions. It will be up to humans to come up with that initial Zip code list that presumably works well -- and track the results to see that this is the case. Discovery of performance of new Zip codes would require repeating this process (and increasing the “learning” budget), as many times as necessary to build a campaign across all Zip codes.
But that is a rudimentary example. What if the clients also want to overlay other factors, like gender, age, income, prior click history, site context, graphics, etc.? All of a sudden you are tasked with analyzing hundreds of factors to help clients decide where to spend their money most effectively. Even if the DSPs say they have preprogrammed algorithms to help lighten their load, they are limited by the human front-loading of assumptions. and only optimize within that small subset of impressions matching those assumptions. Discovery of what’s driving performance is still all-too-manual.
What if DSPs could do it all -- perform deep impression and page analysis to extract numeric, graphic and semantic components, as well as Zip-code level variables, ad engagement, and rich user-level insights, AND understand the performance correlation of each of those separately and combined? These hundreds of granular impression attributes could be used to formulate a higher-resolution picture of a campaign’s success. Each learned attribute could then contribute to a more intelligent bid price for every new impression based on deep machine learning, rather than someone sitting with Excel saying to him- or herself, "Well, maybe we'll try another route."
While the current techniques might work, they are limited by scale. Your media optimization team found out that males aged 30-35 who go to TMZ.com at night convert. How hard was that to find? How many impressions and dollars did you waste? How much will you still need to spend to find out the value of female, 35-39, who read the Financial Times? A truly automated programmatic solution could factor in many attribute variables, constantly testing them and making adjustments as necessary.
This is the future of online advertising. New technology should totally support fully automated programmatic buying by determining the right value (eCPM) for every impression based on a multitude of data points, not just filtered to find the 1% that is “perfect.” It would maximize the volume of a given campaign and its creative, allow us to look at the insights of each campaign, and focus on coming up with new promotions, new creative for the audience group that is working great for us.
At the end of the day, DSPs follow rules controlled and limited by people. Instead of forcing that approach, performance marketers and brands running performance should embrace technologies born from data, and grown by performance.