Build an Insight Engine: Behavior Value Algorithms

In my last post, I quickly skimmed over the process of defining customer “behaviors.” While coming up with an initial list isn’t generally all that challenging, assigning the right metrics and value algorithms (which link the behaviors to the other parts of our Insight Engine) can sometimes get tricky. I’ve seen a large consulting company (one of the “Big Six,” when there used to be that many) struggle with effectively codifying a consistent methodology for doing this. The reason, I believe, is that it’s not fundamentally a plug-and-play endeavor… it’s more like fitting together the pieces of a puzzle until they all snuggly mesh into a meaningful alignment.  It takes some iteration, and a healthy dose of creativity, but I’ve never seen a business or an industry where it wasn’t in the end possible to achieve (prior successes have included such diverse businesses as retail banking, consumer entertainment, commercial-sector energy, P&C insurance, pharmaceuticals, automobile retailing, networking hardware, consumer telecom, and non-profit community services).

The key is to define the behaviors and their associated metrics and algorithms in a way that ensures they are:

1.) Clearly measurable (ideally they are already being measured in some form)

2.) Indicative of direct, individual customer action (i.e., something a customer does or does not decide to do).  This also typically ensures they have traceable impact on financial value.

3.) Driven by an isolated set of decision criteria (more on this in a minute…)

4.) MECE (that is, “mutually exclusive, collectively exhaustive,” so that adding together the individual impacts of every behavior gives you a sum that is exactly equal to the total financial impact of all consumer behavior).  This is where the puzzle comes in!

Essentially, we’re trying to link how “value” is measured (e.g., revenue, profit, market share) with how individual customers behave (e.g., “buy product,” “use service,” “refer a friend.”)Now as you might be able to guess, this can become an onerously complex exercise if you let it.  Linking traditional accounting metrics to this ‘behavior perspective’ can become a headache if you take it so far as to parse all the components involved in margin calculations and try and map to every metric in a product-oriented accounting system.  My perspective on this one is somewhat like the Web 2.0 philosophy of perpetual beta.  Build it basic but good enough to show its value, put it in play, and start a cycle of feedback and iteration until it “evolves” into what you need it be.  Start with behaviors based on existing measures of product margins, make the puzzle fit as best you can with an understood margin of error, and leave second and third order considerations out until you have broad buy-in and clear justification in place for addressing them.  Trust me, the level of insight you’ll see generated from even a ‘bubble gum and bailing wire’ Insight Engine will be highly useful.

Let’s look at a few simple examples of behaviors.  Let’s pretend we’re a company that sells widgets over the internet.  You might start your list of desirable customer behaviors with “buy widgets.”  This meets our first criteria above… almost certainly our company is already tracking how many widgets it sells.  And it meets our second criteria as well… customers decide on whether or not to purchase widgets and then act on that decision.  It’s actually not that strong on the 3^rd criteria, that the decision be based on just one set of decision criteria, for reasons that are somewhat subtle.  Typically, marketers break out the “trial” stage from the “repeat purchase” stage, largely because they are often driven by different decision criteria (for example, ‘reputation’ may matter more to the first purchase decision and ‘satisfaction’ more to the second).  For the same reason, we’ll be better served when we go to link these behaviors with customer decision criteria if we break “buy widgets” into the distinct behaviors of “Try” and “Buy Again.”  Let’s take it a step further…. say that widgets are something a typical customer buys every few months.  Instead of “Buy Again,” we may want to define a “typical customer profile” and have our post-trial purchase behavior be “Become a Regular Customer.”  Doing this let’s us then establish other customer behaviors relative to our “typical” baseline, such as “Buy More Widgets [than average customers do],” “Buy More Often,” “Stay a Regular Customer Longer.”  This then lets us focus on what differentiates a typical customer from an avid customer, which can be a source of invaluable insight (think 80/20 rule). 

For each of these behaviors, we need a way to measure it and determine its financial impact.  For example, the behavior of “Try” can be measured by the number of initial purchases by first time customers during a specified time interval, such as in a month.  (Note that this definition requires us to keep track of who’s buying what, which might not be the right way to go for all businesses; another example of the need for customization).  Our value algorithm may be as simple as the net margin on a widget (if we’ve chosen profit as our financial impact measure).  “Buy More Often” might be measured as the number of purchases per year relative to the typical (baseline) customer.  We might need two value algorithms here, based on whether we’re using our model to predict future impacts or account for things that have already happened.  The prospective measure might be linked to the margin on an average purchase order, while the retrospective measure might take into account whether increased frequency dilutes purchase order size by linking to an average purchase order amount for purchases made above the baseline.  As you can see, this is inevitably an iterative process (although really not as complicated as I’m making it sound.)

And the MECE challenge… As I implied before, there really is no formula for this; it just takes practice and iteration, communication and vetting.  The behaviors have to be measurable and meaningful and their links to financial value mappable.  Defining an average customer baseline can be very helpful, as it allows you to build behavior algorithms related to customer quantity (e.g., “Become a New Customer,” “Stay Longer as a Customer”) using an “average customer value” parameter.  You then don’t have to contend with multiplicative impacts when also including incremental existing customer behaviors (e.g., “Buy More,” “Buy Different Product Lines,” etc.)

While assigning value algorithms to behaviors such as “Buy a Widget for the First Time” are pretty straightforward, more complex behaviors can also certainly be incorporated into our model.  In the marketing-cluttered world we live in, one of the most powerful value drivers today is customer referral.  Here, the metric may be number (or percent) of new customer trials resulting from referrals.  Again, there’s a data tracking component required for measuring retrospective impacts, but more and more businesses seemed to have already incorporated this one in a standard upfront question of, “Where did you hear about us?”  The value algorithm might weight the value of an average customer with the likelihood of becoming an average customer and credit all customers generated by referral to that behavior. 

Sometimes value comes from a source not directly related to the customer actions that are a focus of your business.  For example, if you have a website that primarily generates revenues from on-site advertising, the customer behaviors you are most interested in aren’t necessarily those of “click ad,” which is easy to link to revenue, but those of “visit site” and behaviors that lead to deeper levels of engagement with your content so as to promote loyalty and return visits.  In these cases, it’s often helpful to calculate and assign a dollar value to an intermediate metric, such as “revenue per page visit,” (which can be determined based on an estimated likelihood of clicking ads) and which provides a base for behaviors such as “visit more pages,” “visit more often,” etc.

My apologies for going into the weeds in a few places in this post.  Perhaps in a later post I’ll pick an example and work it through as that may better demonstrate it’s not nearly as complicated as I’ve made it sound.