Ulti­mate Fris­bee is one of the favorite lunch-break past times of many Adobe Con­sul­tants.  If you’re not famil­iar with the game, it’s actu­ally pretty sim­ple.  The object of the game is to score points by throw­ing a Fris­bee down­field to a player in the oppos­ing end zone, while any player in pos­ses­sion of the Fris­bee is not allowed to run or move.  The team with the most points wins.  While we were play­ing dur­ing a lunch break a few days ago, one of my fel­low con­sul­tants pointed out to me how attri­bu­tion mod­el­ing can be applied to this game the same way we would apply it to dig­i­tal mar­ket­ing cam­paigns.  Let me elaborate:

Think of a dig­i­tal mar­ket­ing cam­paign as a player on an ulti­mate Fris­bee team.  Each player is work­ing together to move the Fris­bee (or site vis­i­tor) to the end goal.  Often­times it takes mul­ti­ple play­ers (or mul­ti­ple cam­paign touch points) before the con­ver­sion suc­ceeds, and each con­tribut­ing player has an impact on the end conversion.

So the $100 ques­tion is, when your team scores a point, who gets the credit?

Last touch attri­bu­tion would give all the credit to the player who caught the Fris­bee in the end zone.  First touch would give all the credit to the player who made the ini­tial kick-off pass.  These are the most com­mon attri­bu­tion mod­els in web ana­lyt­ics, but nei­ther model incor­po­rates the per­for­mance of all the play­ers who helped move the Fris­bee downfield!

Per­haps a lin­ear model would be bet­ter since it would give credit to all the play­ers who helped move the Fris­bee, but it doesn’t give a whole lot of insight into which player has the best throw­ing abil­ity, or which player is the best catcher, or the fastest run­ner.  Per­haps there are even com­bi­na­tions of play­ers that seem to work very well together, but are not as effec­tive separately.

In my opin­ion, a good model would give the most credit to the play­ers who par­tic­i­pate in scor­ing runs most often.  In other words, when my amaz­ing team­mate Jes­sica Olsen touches the Fris­bee, we seem to score more often than when she does not.  As a result, I want a model that is sure to give Jes­sica the credit she deserves.  With that in mind, let me explain how we would deter­mine the best cam­paign per­form­ers using sta­tis­ti­cal mod­el­ing tech­niques, and most impor­tantly, how you can use this for your web ana­lyt­ics data.

First, let’s make the big assump­tion that dig­i­tal mar­ket­ing cam­paigns are caus­ing peo­ple to con­vert (hope­fully your mar­ket­ing cam­paigns aren’t turn­ing peo­ple away from your prod­ucts!).  With that assump­tion in place, it’s ok to say that the cam­paigns that are most cor­re­lated with con­ver­sion are the most successful.

Sec­ond, we need to look at the expe­ri­ence of each vis­i­tor over their life­time and cap­ture all of the cam­paign touch points they expe­ri­enced before con­vert­ing within a look-back time­frame.  In order to do this, Adobe Con­sult­ing ser­vices has con­structed a spe­cial­ized data pull algo­rithm that sum­ma­rizes all users’ behav­ior over any given date range.  Essen­tially, the out­put dataset con­sists of a sin­gle row for every user that pur­chased and each cam­paign they touched on the path to their con­ver­sion.  With this dataset, the sta­tis­ti­cal mod­el­ing can begin.

One sim­ple and effec­tive model is to cre­ate a cor­re­la­tion matrix.  This is essen­tially a grid that shows how related a row and col­umn item are, or how they tend to move together.  It’s also impor­tant to note that dif­fer­ences in aggre­gate num­bers won’t affect the cor­re­la­tion score, or in other words, if you have a bad player on your team that hogs the Fris­bee, they won’t get any extra credit.

Click to Enlarge Image

(Click to enlarge image)

In this exam­ple, we’re inter­ested in the two far right columns because it shows how cor­re­lated each mar­ket­ing cam­paign was with orders and rev­enue (if your mar­ket­ing cam­paigns have other key goals or met­rics such as ad rev­enue or lead gen­er­a­tion, those met­rics can be sub­sti­tuted here as well).  You can see that Paid Search is a bet­ter Fris­bee player than Social Media because Paid Search led to higher rev­enue more effec­tively than Social Media did.

So based on these scores, I can assign a weighted aver­age of our rev­enue so that the rev­enue gets dis­trib­uted accord­ing to each cam­paigns’ com­par­a­tive per­for­mance.  The for­mula for that would look some­thing like this:

Now this is a fairly sim­ple approach, but is almost cer­tainly bet­ter than first or last touch since the best Fris­bee play­ers will get the most credit for their skills.  Of course there are more advanced mod­els that can be explored from this spe­cial dataset that incor­po­rate addi­tional infor­ma­tion such as which chan­nels seem to work well together, how a channel’s influ­ence decays over time, or how a channel/campaign is affected by the order in which it appears for users (i.e. some play­ers are bet­ter catch­ers or throw­ers), but these are top­ics for another day.

Under­stand­ing which cam­paigns are actu­ally boost­ing your site per­for­mance is as crit­i­cal to dig­i­tal ana­lyt­ics as know­ing who you can rely on to catch your game-winning Fris­bee pass.  If you’re inter­ested in build­ing an advanced attri­bu­tion model for your own com­pany, please reach out to your Adobe sales rep, account man­ager, or con­sul­tant who can guide you to the next steps.