I’m always surprised that you don’t see more direct marketers hired into Web analytics roles — they’re expert in a number of skills such as predictive modeling, experimental design, and statistical analysis that really turn the Web analytics role into a strategic source of business revenue. Along those lines, I wanted to give an example of how a tried-and-true direct marketing technique — using regression modeling to predict how likely each user is to convert — can allow media buyers and SEM managers to significantly reduce the amount of time it takes to determine which campaigns are winners and campaigns are losers.
Companies rely on pay-per-click and online ad buys to drive growth. While buying clicks is easy, ensuring that individual campaigns are profitable – before significant losses have been incurred – presents a host of challenges to media buyers. Many online businesses – such as the travel, mortgage, real estate, and online dating industries – face an additional challenge in that leads may take many days or even weeks to convert. As a result, visibility into how well new campaigns are performing early-on can be very limited. Put simply: Is there any way media buyers can get an idea of how individual campaigns are converting before waiting days or weeks for the conversion numbers to come it?
Absolutely! The direct mail industry thrives on predictive modeling techniques that estimate how likely individual consumers or businesses are to convert. This same technology has been successfully applied in online contexts. Predictive modeling works by quantifying how likely each user is to become a customer. Analyzed at the campaign level, this provides media buyers with an early read on what conversion for any particular campaign will look like – without waiting days or even weeks for a report on eventual conversion rates.
What do I need to get started?
Customer data — ideally, lots of it. Here’s a good list to get started. You’ll need to have this data available for each user:
- Campaign Identifier (so you can summarize results by campaign)
- Campaign metadata (e.g., is this a Facebook campaign, AdWords, etc.)
- Market-level summary data, such as number of businesses in the ZIP code, your penetration in the ZIP or CBSA, average conversion rate in the market area, etc. Claritas Pop-Facts is a terrific source for ZIP-level and CBSA-level data in the US.
- Source metadata, such as conversion rate by email domain, landing page, or source URL
- User data, e.g., age, sex, ethnicity, income. Almost any demographic information you have is worth trying.
- Behavioral data. Has the user ever logged back in to the site? Did their email bounce? Are they engaged with the site? Any site behavior you can measure may make a big difference.
- A flag variable, 0 or 1, that operationally defines conversion. For example, “this user converted within 3 days” or “this user converted within 7 days” — whatever that makes sense for your business needs.
What do I do next?
Use logistic regression to develop a predictive model. You’re going to have more data than something like Excel can handle — typically you’d use a tool like SAS, SPSS, or R to do the modeling. If your Web analytics team doesn’t have this skillset and this software, they should.
The output of this model is very simple: It is an equation, based on the data you fed it, that tells you how likely each user is to convert. If you just add up all these marginal probabilities at the Campaign level, voila — your predicted conversion rate by campaign.
As a bonus, if you multiply this predicted value by average revenue, you have a great estimate of how much every single user is worth. The possibilities are almost endless for discounting, segmentation, remarketing, even A/B testing… the list goes on and on. If you’re in a long time-to-convert business, this kind of model is something you can’t afford to be without.
I’m going to go ahead and throw out a shameless plug for a mentor of mine over at Greater Good Analytics who does this kind of predictive modeling for a living (if you need more help getting started.) Also, over the next couple of weeks I’ll be supplementing this with post examples of how to implement regression models in a Web-based world and the pros and cons of using SQL Server versus MySQL for your analytics environment. Until then, enjoy!