In the mid-1990s, one of the biggest waves of technological innovation in business was the use of data warehousing and data mining to gain insights from the mounds of data produced by customers in the course of their interactions with a company, such as their purchases or responses to direct marketing. The most famous result was a grocery/convenience store chain that used data mining to discover that it had a segment of customers who came to the store to purchase diapers and beer.
It was a fun story and an easy one to discuss with marketers as an example of the kind of thing one can discover using interesting technology combing through data to find insights. Like many insights, it makes you wonder: Who are these people? And how can I as a marketer get them to buy more?
Ultimately the theory was that new fathers were frequently being sent on diaper runs, treating themselves while they were at it. CIO magazines of the era had data-mining software ads showing a fat, hairy guy wearing a diaper and drinking a beer.
While this story got a lot of play in spite of its questionable heritage, the insight wasn’t directly actionable. If you’re in charge of lifting sales, what do you do? Put the diapers and beer next to each other, so that even more dads can reach for the beer as an impulse buy? Or put them farther away with potato chips and hardware in between so that the dads have to travel through the store and maybe buy more things on the way?
Less fun to talk about at cocktail parties but more actionable were the models that companies used behind the scenes to score customers based on propensity to respond to marketing offers. These were often somewhat inscrutable, but they were directly actionable since their output was used to qualify or disqualify a customer from, say, an extra catalog mailing. This is one of those things where everybody wins: the company saves marketing costs, consumers’ mailboxes are less stuffed with junk, and a few hundred trees live for another day.
Actionable Insights in Display Advertising
Recently we discovered that people who listen to electronica and alternative music were most likely to respond to a particular campaign. Just like grocery store merchandisers, savvy display advertising media buyers are always looking for insights from campaign results. But they have similar struggles in getting actionable insights from their partners. While it’s nice to know that behavior, gender, or geography influences response rate, it’s not like we can target just female mobile phone shoppers in Nebraska who listen to Enigma for the next wave of a campaign; the audience becomes too small.
Let’s say you have a campaign to sell mobile phone service, and you’ve delivered 3,000 conversions after 180 million impressions at an average $2 CPM (define). If the client asked if you could deliver another 150 conversions, you’d likely say yes, and it should take around 90 million impressions or $180,000 merely by running the campaign with the same tactics and prices. This seems achievable (except for fatigue effects, which are interesting but a topic for another day). Essentially, according to overall statistics, you can deliver a second wave at any level of spend between $0 and $360,000 and expecting the same average CPA (define):
But can you be smarter? Say you’ve run the campaign with three tactics: retargeting, a behavioral segment of mobile phone shoppers, and a run of network (RON) buy. The results look like this:
|Tactic||Impressions (M)||eCPA ($)||CPM ($)||Conversions||Spend ($)|
|Behavioral targeting: mobile||40||40||2||2,000||80,000|
|Run of network||100||400||2||500||200,000|
If you wanted to run the campaign again to optimize CPA, you’d first target just the mobile segment, then add in retargeting and RON to add volume. You can draw a chart that represents how you can most efficiently spend marketing dollars based on this strategy:
This chart represents a set of attainable operating points that you can choose from for the next wave of the campaign. Want to spend $100,000? No problem, you can expect 2,130 conversions at an effective CPA (eCPA) of $47. Want to spend $200,000? Expect 2,600 conversions at an average eCPA of $77. Just slicing results by tactic generates a lot of improvement over the previous operating curve based only on the average. It makes you wonder: are there other attributes to slice by, and how far can you go before it becomes unwieldy to manage this by hand?
If time of day, geography, the page the ad is being served on, the user’s response history, and even the user’s music tastes all matter and can influence the likelihood of a response, then a statistical model incorporating all of these factors will have an operating curve superior to manual segmentation, because the model is looking at more data and evaluating more possible relevant correlations.
It’s like the episode of “The Office” where Dwight tries to outsell the company’s Web site. There’s no point in fighting tools; you have to learn to use them. Dwight only wins on TV.