By Paul Sobel, CEO, Dataline
The April 20 edition of Marketecture is perfect. It’s the Data, Stupid! is beyond reproach, especially since it was authored by a guy who admits that the DSP he started, ran and sold for millions had no unique inventory.
I accept Ari’s three key premises:
1: Unique data is the only real way to compete;
2: Other stuff is nice, but losing importance; and
3: The actual DSPs in the market have somewhat weird and idiosyncratic datasets.
Best of all is his conclusion: DSPs should be bidding as much as they can afford for any data asset that becomes available.
95 Percent of Buyers are Wrong
So, why does anyone on the buy side ever use syndicated segments? Literally 95 percent of all programmatic media buyers rely on syndicated segments over modeled audiences and other data choices, despite the fact that these alternatives outperform against every metric imaginable. How do I know this?
I run a data company, of course. And yes, we sell modeled audiences and syndicated segments. So do other data providers. So, without revealing anything proprietary or seeming to advantage what we provide, let’s stick to what marketers should be asking their data providers.
How to Get Better Results Immediately
Everyone on the buy side knows how to conduct a test. But, how many of you know how to conduct a test between syndicated segments and modeled segments? Here’s how to do it:
1: Divide the test into two groups: Modeled Audience and Handpicked Syndicated Segments
- Create your General Audience: select random reference sample of 50K records from 230 MM database and a target of 10K records from seed list.
- Run a model and create a General Audience
- Create a Syndicated Audience: select the 50K random records from handpicked syndicated segments.
- Create a model from the syndicated audience using the 50K reference sample and 10k target records from seed list.
- Create your Modeled Audience: select a reference sample of 50K records from an audience built using any modelling process that uses a different seed list as its target. Then take a random sample of 50K records from the modeled audience as a reference and build a new model with a 10K seed list.
- Run that model and create your Modeled Audience.
- Compare the efficacy of both approaches by dividing the number of client records in the top two bins by the number of total records and calculating the ratio between the test group’s performance and the general audience’s performance.
Your Modeled Audiences will outperform any syndicated segment on average by 2.5X or more.
The bottom line is that modeled audiences will outperform syndicated segments by an average of 2.5 times. Considering that 95% of all digital advertisers use syndicated segments, having them move to modeled audiences would be a big win by increasing ROI and ROAS.
What do you have to lose – besides a bunch of the data providers you’ve been working with?
Give it a try!
