The Magic of Behavioral Targeting -- Optimization Approach #3

In my past few columns, I have set out to clarify optimization, a term that is often bandied about and regularly misunderstood.

I first covered testing, the most frequently used method of improving consumer response, which includes A/B testing and multivariate testing. With the targeting article, I covered how systems based on rules can be used to create more relevant experiences with better outcomes.

The third type is perhaps the most seductive -- and misunderstood -- form of optimization, behavioral targeting. (The fourth, social optimization, I will explain in the near future.)

What Is Behavioral Targeting?

The holy grail of direct marketing has been a system that detects consumer behavior and changes offers. The first incarnation of this approach was called data mining, and was focused on using data to drive strategic planning. There is an apocryphal story about Wal-Mart: "By scanning each sale into a data warehouse, grocery stores have determined that men in their 20s who purchase beer on Fridays after work are also likely to buy a pack of diapers. Thus, a display of Pampers or another brand might be set up in the beer aisle, or merchants will put one (but not both) of the products on sale on Friday evenings."



In the online arena, it is actually possible not only to process historical data, but also to act on it instantly. In an ideal case, the Web marketer would just plug in new offers or products and the system would automatically put the perfect one in front of each visitor based on these patterns while that visitor was still on the site.

Behavioral targeting is really a refinement of targeting, or "rules based" optimization.  The difference is that while targeting uses explicit segments, predictive models seek to discover rules within the data that are counterintuitive. Instead of When the user has searched for 'TurboTax,' show this offer at this price, the predictive model can look at prior searches and visits to pick the ideal offer.

In simple terms, it's like having the haystack find the needle for you.

With behavioral targeting, you're saying that, if you have a big enough body of data around a consistent set of products and a consistent set of visitors, then it is quite likely that you will find correlations in how those visitors behave: different times of day, different types of product, different points of origin are some subsets that can be grouped together and perhaps predicted to behave in similar ways.

The classic example is the predictive models used by Amazon for product recommendations. Amazon offers recommendations based on a customer's past purchases as well as on the purchases of others who have bought similar items. As near as I can tell, no other company has demonstrated a more lasting dedication to using predictive models.

To make behavioral targeting work, the system (or consultant) essentially works to build a model based on prior behavior that can be used to predict future customer preference. With this approach, a broad range of variables (time of day, source of traffic, prior purchase) are evaluated against performance, and a model is developed that is "fitted" to these conditions. The model is both explanatory, meaning that the past data supports its assertions, and predictive, meaning that it predicts how new customers will behave under similar conditions.

Where It Works

Ok, take a breath after that last paragraph.  You have my apologies for the excruciating terminology. But there are simple places where you will run into predictive models every day.

Take search, for example. When a consumer types in a search phrase, the search engines try to find the best list of sites. To do this, they must build a model based on what they have found to be the most predictive characteristics that they can gather. In Google's case, they weigh the content of the page, but also inbound links and a factor called "PageRank" (among other factors). This model is employed every time a "SERP," or search result page, is calculated.

Ok, so Google uses it, and Amazon uses it, so when should you use it?  The best times to use behavioral targeting are either:

  1. In applications such as search and cross-selling where there are large sets of alternatives of a single type, and a large number of interactions.  This means books, music, Web pages.
  2. Direct marketing situations where you have a small number of offerings and a very large number of relatively well-profiled prospects or customers. For example, credit card offers, loan rates, and other high-volume, high-value markets.

In both cases, behavioral targeting tends to work best when there is prior purchase behavior on the part of the consumer. Such models are less effective in prospect situations, as there tends to be less profile data on the visitor to correlate.

The Good, the Bad and the Ugly

On the surface, behavioral targeting seems like magic. Why isn't all marketing done this way?

The biggest issue in targeting behavior is building the model of customer behavior.  While modern Web analytics captures a staggering quantity of data, the models tend to be built to fit the available data, not the most predictive data. 

For example, the excellent music Web site Pandora would be impossible with most current data mining or analytics solutions. Pandora has distinguished itself for the quality of its recommendations, but its predictions rely as much on quality research and testing as on the math.

The secret behind Pandora's success is the Music Genome Project, a research effort that seeks to catalog and attribute songs based on major key tonality, dynamics of the lead singer's voice, the prominence of certain musical instruments, and the number of beats per minute within the song, and creates a channel that plays songs that have similar attributes.

An ecommerce Web site that sells music might use behavioral targeting in a different way. Rather than suggesting music based on attributes of the music itself as Pandora does, a site might look at the pages a visitor has viewed and make suggestions based on what was purchased by others who followed a similar clickpath.


But Behavioral Targeting has real drawbacks:

  • For one thing, there is no such thing as a universal model. Models must be created and tuned for different applications -- a process that can take months. 
  • Behavioral targeting models and schemas require regular tune-ups to be certain that the groups or segments are still behaving in the way they were originally predicted to behave. They are opaque, meaning that a marketer has little opportunity to understand the reasons behind the correlations, and thus has little chance of learning from it.
  • Behavioral models can also make some bizarre mistakes (offering a cross-sell that is completely irrelevant, for example).
  • They tend to be "Black Box" and give much less information back to the marketer than testing or testing with targeting can provide.
  • And finally, they take the merchant or editor out of the equation, which can lead to serious "voice" issues. At its best, behavioral targeting is used to enhance the marketer, not to replace the marketer.

Behavioral targeting is not likely to replace the marketer anytime in the near future.  As I have said before -- marketing is done by marketers. Machines just help us listen and aim better.

Next story loading loading..