Attribution's Recommendation Engine: Windshield Vs. Rear-View Mirror

In "A Chevy Convertible & Attribution: Both Offer Top-Down Technology," I used the analogy of a convertible car to describe "top-down" attribution.

This month let's extend the car analogy -- this time right from the driver's seat -- and discuss the need to be equally skilled at using both the rear-view mirror AND the windshield to be an effective driver. 

In the case of attribution management there has traditionally been a significant focus of time and attention on applying algorithmic processes to historical data in order to look back in the rear view mirror at the cross channel metrics produced by your past efforts.  But just as looking in your car's rear view mirror doesn't tell you which way to turn the wheel next, looking back at your historical attribution metrics falls short of accurately informing a more effective allocation of your marketing dollars moving forward.

Validating Past Performance for Future Predictability

The attribution management process done correctly is exactly that - a process.  It is not a one-time project (aka media/marketing mix modeling) or a single snapshot in time of your cross channel marketing performance.  It is iterative, self-refining and self validating.  It is an oval racetrack where results from each successive lap of the track perpetually inform the algorithms which build the models to help predict your future performance.

As was intimated by the clue provided at the end of "Can The Accuracy Of Attribution Be Validated?" with each attribution cycle the model that predicted the outcome of the previous cycle is validated by new results and is adjusted to compensate for the delta between predicted and actual results.  The more the measure->analyze->predict->optimize->validate->measure cycle takes place, the more refined and more accurate the model becomes.  Attribution solutions that are "productized" into software form - rather than executed as a custom professional services project - enable these cycles to be repeated as frequently and rapidly as you can load your marketing performance data.  This is why attribution management software kicks the butt of media mix modeling or marketing mix modeling service engagements when it comes to accuracy and long-term viability of its results.

Models to Predictions to Recommendations:  Rubber Meets Road

All that said, attribution models do not show or tell you which way to turn the steering wheel or when to apply the brakes.  To obtain that information you need to introduce one or more values into that set of rules, onto which those rules can get applied.  For example, if you invest $1000 on creative "C1" on publisher "P1" and bid $1.00 on keyword "K1" on Google, what will the outcome be?  The response that attribution provides to this question is a prediction, but in this example, it is but a single, isolated prediction that addresses only a couple dimensions across a couple channels.  And as the poser of this question, you had to pick the criteria you were going to ask - and do so in a fairly hit-or-miss fashion.

At its best, what attribution allows you to do is clearly define what your objective or objectives are. Attribution also asks all the possible questions of the model for you, and then delivers recommendations to you on how to achieve those goals.  For example, if you have $100,000 to spend, what's the optimal combination of marketing channels, campaigns and attribute criteria in which to invest to produce the greatest return?  Or if you need to produce 20,000 new loan applications, what's the optimal mix needed to achieve that objective? What will your CPA be?  The recommendation engine allows you to pose "what-if?" scenarios, and based on the models that have been built, validated and refined, this engine delivers specific instructions along these parameters:

·       Channel Mix (in which channels should you invest)

·       Tactics/Attributes (publisher, size, placement, keyword, creative, etc.)

·       Affinities (run specific tactics in unison to produce 1+1=3 effect)

·       Timing (when to run tactics)

·       Sequence (in which order to run which tactics)

·       Frequency (with what frequency to run tactics)

Armed with these instructions, you'll see as clearly and accurately looking backward through your rear-view mirror as you do looking forward through your windshield. This will enable you to keep your eyes more clearly on the road, and on the bottom line.

Tags: metrics
Recommend (18)
1 comment about "Attribution's Recommendation Engine: Windshield Vs. Rear-View Mirror".
  1. Damon Ragusa from idio , September 16, 2011 at 2:08 p.m.

    I couldn't agree more with you on the need for attribution modeling to be focused on the future. And agree that it is an on-going process. And your criticism of traditional marketing/media mix modeling is dead on. But here is the rub as I see it (and the gap). Attribution models are great but limited to those things you can count well which is primarily focused on digital marketing vehicles. This limitation is substantial the more an advertiser spends across all platforms, both measurable or not. Marketing/media mix models have historically been at too strategic a level to be of executional value.

    So there is a big gap between the broad, strategic value of marketing/media mix and the granular, execution-orientation of digital attribution models. Focus needs to be placed on methods that can fill this gap. That require a blend of both counting and modeling…a blend of audience metrics and consumer behavior…replace static models with dynamic systems that (once rooted in consumer behavior) have greatest context for future change.