A Pragmatic Approach to Attribution

by Feb 14, 2018

Attribution is a hot topic in marketing that keeps just getting hotter. The problem is that, all too often, it gets treated as an all-or-nothing proposition.

As John Wana­maker observed almost a century ago, much of our adver­tis­ing dollars are wasted. If only…if…ONLY!…we could figure out which dollars those are, we could finally stop dread­ing those meet­ings with the CFO! Surely, with the explo­sion of data and data plat­forms that have come with the shift of so much consumer behav­ior into the digital realm, we should finally be able to perfect our media mix. Right? Right?!!

The unfor­tu­nate reality is that, as long as we’re market­ing to human beings in a chaotic, message-over­loaded world, there will be some degree of unknow­able waste in our market­ing. Step #1 in an appro­pri­ate perspec­tive regard­ing attri­bu­tion is to treat any vendor’s claim to have “solved” attri­bu­tion with a level of skep­ti­cism that matches the cost of their offer­ing. That perspec­tive leads us directly to those three magic words: return on invest­ment.

ROI calcu­la­tions have two compo­nents: the cost (invest­ment) and the return (value deliv­ered). In the world of market­ing invest­ments, there are some impor­tant subtleties:

  • The return is the incre­men­tal increase in busi­ness results (revenue or others) that we are able to achieve as a result of the invest­ment.
  • The cost of using attri­bu­tion as a tool for guiding our invest­ment can vary wildly — from almost nothing to six-figure invest­ments in complex plat­forms with highly involved data collec­tion require­ments

This means we are perpet­u­ally perform­ing a balanc­ing act — trading off the incre­men­tal value we will get from each improve­ment in accu­racy or complete­ness of our attri­bu­tion, with the cost of that improve­ment. With attri­bu­tion, it’s easy to hit a point of dimin­ish­ing returns — substan­tial increases in invest­ment may yield only nominal improve­ments in value.

It is entirely too easy to get wrapped up in seeing “attri­bu­tion” itself as the goal, rather than recog­niz­ing it as a means to an end: effec­tively and effi­ciently invest­ing your market­ing dollars. Or, even better, maxi­miz­ing the return from your market­ing invest­ments.

What Is Attribution And Why Is It Messy?

Skip this section if you know the answer to this ques­tion already. There is no earth-shaking defi­n­i­tion here, but some of the termi­nol­ogy dropped later in this post assumes a basic level of famil­iar­ity.

Attri­bu­tion, put simply, means accu­rately and effec­tively assign­ing value to a market­ing channel, campaign, or tactic. That way, we can compare the cost of that channel, campaign, or tactic to the value assigned to it, and derive the ROI described above.

So, why is this messy?

In a simplis­tic scenario, it’s not. Imagine that no one has ever heard of your brand: Super Awesome Widgets. And, imagine that your only promo­tion of the brand is through Google Adwords. And, imagine that consumers click through on your paid search terms, navi­gate the site, and place an order on the site during the same visit. Every time.

In that case, it’s easy and clean: 100% of your site’s revenue can be attrib­uted to Google Adwords.

But, exactly zero brands operate in a scenario like that.

Another (still) fairly simple scenario would be:

  1. A user searches for “widgets.” She sees a paid search ad and clicks through on it, navi­gates the site a bit, and then leaves without order­ing. But, now she has a good idea of what you offer widget-wise and and at price.
  2. A few days later — real­iz­ing she really needs to get around to buying a widget — she searches for widgets again. This time, though, she clicks through on an organic search result. She then navi­gates the site and makes a purchase.

In this case, which channel should get the “credit” for the revenue? Is it paid search, since that is how she first came to the site for her initial research? If paid search receives all of the credit for the revenue, then that is consid­ered first touch attri­bu­tion.

Or, rather, should organic search get the credit, since that was the last channel she used before making the purchase? That would be consid­ered last touch attri­bu­tion.

Or, should both chan­nels get 50% of the credit (linear attri­bu­tion)?

Perhaps both chan­nels should get full credit (wildly irre­spon­si­ble attri­bu­tion)?

This scenario illus­trates why attri­bu­tion is compli­cated. And it can be very expen­sive. The rest of this post breaks attri­bu­tion down a bit and lays out how, by keeping an eye on ROI prize, it’s possi­ble to incre­men­tally improve the sophis­ti­ca­tion and complex­ity of an attri­bu­tion manage­ment program.

Three Different Dimensions of Attribution

One way to think about attri­bu­tion is that there are three dimen­sions of the problem. Each dimen­sion ranges from “basic” to “advanced,” and, to some extent, we can progress along each dimen­sion inde­pen­dently from the others:

Keeping the ROI discus­sion above in mind, we can now think about which dimension(s) make for the most sensi­ble incre­men­tal invest­ment. There is no one-size-fits-all!

Channels: From One to Many

Each market­ing channel is unique. As such, a typical start­ing point is to opti­mize within each channel inde­pen­dently. For instance, you may use ROAS or some other metric within your paid search efforts to adjust keyword bids. Contrast this with deci­sions you make regard­ing how often to send promo­tional emails — likely opti­mized for a combi­na­tion of click­throughs and conver­sions while keeping an eye on the opt-out rate.

This approach isn’t inher­ently “wrong.” It’s actu­ally a fine place to start — certainly much better than simply throw­ing invest­ment at each channel and not assess­ing and tuning the invest­ment within that channel.

On the other end of the spec­trum is attempt­ing to manage the inter­play of all chan­nels with each other. This is…complex and expen­sive. So, is there some middle ground? There is!

The middle ground involves a little thought and an incre­men­tal increase in complex­ity and cost:

  • Assess­ing which chan­nels can most directly be impacted (“Our direct traffic has the highest conver­sion rate! We need to drive more direct traffic!” Um. Good luck with that.)
  • Assess­ing which chan­nels, logi­cally, we expect to be the most inter­twined. Paid search and organic search are two obvious ones here: two differ­ent chan­nels (in the marketer’s mind, if not as clearly so in the customer’s) with differ­ing ability to directly and imme­di­ately impact. But, focus­ing on just those two chan­nels, includ­ing possi­bly doing some controlled exper­i­ments, can yield a better under­stand­ing of both chan­nels and how they inter­act, which can lead to more informed and effec­tive invest­ments.

That’s one dimen­sion of attri­bu­tion, and that dimen­sion, alone, offers many oppor­tu­ni­ties for incre­men­tal improve­ment.

The Length of the Attributed Journey

Simi­larly, there is a “basic” start­ing point for the breadth of the consumer journey that we consider: the journey starts when a customer arrives on our website, and we consider it completed when the customer “converts” on the site.

This is rela­tively straight­for­ward, in that this basic level can be performed entirely within the web analyt­ics plat­form (although it has depen­den­cies on the quality of the data, includ­ing accu­rately track­ing the sources of traffic to the site).

The length of the journey can be expanded both upstream and down­stream:

  • Ad impres­sions occur upstream of the actual click­through. (For that matter, attri­bu­tion efforts always ignore “legacy brand feel­ings” — I was raised in a Colgate house­hold, so how much does that impact my tooth­paste purchas­ing deci­sion? As it turns out, I now use Crest…because my wife was raised in a Crest house­hold and gener­ally does that shop­ping. It’s going to take a lot of market­ing to over­come habit!)
  • Focus­ing on the imme­di­acy of an on-site conver­sion may or may not be the main value of the customer rela­tion­ship. Does the first purchase lead to addi­tional purchases? Do customers who look nearly iden­ti­cal at the time of the online conver­sion actu­ally have vastly differ­ent life­time values?

Like our first dimen­sion, it makes sense to start simple and then progress from there — weigh­ing the cost of the increased complex­ity with the value that that progres­sion provides in return.

Model Approach: From Heuristic to Algorithmic

This final dimen­sion may be the one that causes the most confu­sion. It’s also the dimen­sion with the fanci­est termi­nol­ogy.

Funda­men­tally, there are two differ­ent ways to approach attri­bu­tion:

  • Heuris­tics — this simply means you, the marketer, chooses how you want to assign value along a multi-touch path. 100% of the value to the last touch? 100% of the value to the first touch? Evenly spread­ing the value across all touch points? Weight­ing the the value so that later touch­points get more value? When Google Analyt­ics first rolled out their multi-channel attri­bu­tion solu­tion (prior to their acqui­si­tion of Adom­e­try), this is how they defined attri­bu­tion. It’s not inher­ently “wrong” to attribute value this way. It just means that value is being assigned based on the judg­ment call of the marketer.
  • Algo­rith­mic (also referred to “data-driven”) — this is much more compli­cated, requires more data (and more detailed data), but, in theory, is more objec­tive (and, in theory, more “right,” but “right vs. wrong” is a slip­pery slope perspec­tive that we should try to avoid!).

In both Google Analyt­ics and Adobe Analyt­ics, the pseudo-default view of traffic source is the heuris­tic approach of last touch. Both plat­forms make it very easy to compare the data from a first touch and last touch perspec­tive, and that’s a compar­i­son that it’s worth making:

  • The results will be differ­ent, BUT
  • Are they so differ­ent that you would act differ­ently if using one approach over the other? This isn’t a ques­tion with a single answer: it really depends on the nature of the busi­ness being assessed.

Google Analyt­ics makes it easy to go one step farther and eval­u­ate addi­tional models. Which, if the first- vs. last-touch compar­i­son raised an eyebrow or two, it’s worth explor­ing.

IF this heuris­tic compar­i­son turns up wildly dissim­i­lar results, then it may be worth explor­ing algo­rith­mic attri­bu­tion. Google’s new Google Attri­bu­tion plat­form will actu­ally include “data-driven” as a default option (but will require suffi­cient data volume in order to work). It’s too early to say for sure, but I suspect that option will gener­ally wind up just being a start­ing point for algo­rith­mic attri­bu­tion. Things get compli­cated quickly when moving to the more advanced end of that dimen­sion!

In Short: There are Many Options

Consider a mid-sized online retailer with a limited budget for digital market­ing and limited staff to manage that budget. The retailer has histor­i­cally spent all of its budget on Google Adwords, but is consid­er­ing exper­i­ment­ing with shift­ing some of that spend to Face­book adver­tis­ing. Using the approach described above, the retailer might:

  1. Start by eval­u­at­ing the spend within the single channel of paid search to iden­tify the lowest perform­ing keywords based on the cost and then the revenue gener­ated.
  2. As a quick check, assess, at a channel level, the differ­ences in the revenue attrib­uted from first click attri­bu­tion versus last click attri­bu­tion.
  3. Assum­ing that this simple check does not return two dras­ti­cally differ­ent stories, shift budget from the lowest perform­ing keywords to Face­book adver­tis­ing.
  4. If the overall results improve, then, poten­tially, simply focus on opti­miz­ing within each channel.
  5. In addi­tion, consider some controlled exper­i­ments: turning off Face­book adver­tis­ing in a few geographic regions and compare the results to regions that had performed simi­larly, but where Face­book adver­tis­ing has remained turned on. Perform this same exper­i­ment with paid search. In both cases, eval­u­ate differ­ent (heuris­tic) attri­bu­tion models: first touch, last touch, etc.

While this is a rela­tively “simple” scenario — it sticks gener­ally to the “basic” end of all three dimen­sions of attri­bu­tion — it still requires plan­ning and dili­gence, and it can yield mean­ing­ful infor­ma­tion about the value being deliv­ered by each channel, as well as how each channel can be adjusted to improve overall results.

Contrast that scenario with a large, multi-channel retailer that has signif­i­cant invest­ments in both offline and online adver­tis­ing. The team has already done quite a bit of intra-channel opti­miza­tion, and has also performed a range of exper­i­ments over time to get a good under­stand­ing of the inter­re­la­tion­ships between differ­ent pairs of chan­nels: when a TV campaign runs, they know what sort of a bump to expect in paid search, and they already have a mech­a­nism in place for attribut­ing that bump back to the TV campaign. The team has also performed exper­i­ments with turning off and on differ­ent ad groups — brand terms vs. non-brand terms — to get a good sense on the inter­play between paid search and organic search. And, they have adjusted their paid search spend accord­ingly.

The retailer is plan­ning to increase their digital adver­tis­ing spend signif­i­cantly over the next 1–2 years and wants to ensure that their media mix remains appro­pri­ate as they do so. In this case, the orga­ni­za­tion may be ready to move more to the “Advanced” end of the attri­bu­tion dimen­sions:

  • Rather than working off of the revenue gener­ated from each order, esti­mate the customer’s life­time value at the time each order is placed, and use that as the primary attri­bu­tion metric.
  • Shift from heuris­tic attri­bu­tion to algo­rith­mic attri­bu­tion (which may require invest­ment in a plat­form and/or inter­nal resources with more advanced analyt­ics capa­bil­i­ties)
  • Work to include impres­sion data — at an indi­vid­ual level for digital market­ing and at an aggre­gate level for offline chan­nels — in the inputs being used for the algo­rith­mic attri­bu­tion.

In both cases, the goal is to find the right level of complex­ity (aka, cost, invest­ment) to make effec­tive deci­sions. For one orga­ni­za­tion, an invest­ment of several hours a week — and no invest­ment in addi­tional tech­nol­ogy — may be the “right” approach to attri­bu­tion. For another orga­ni­za­tion, a poten­tial six-figure invest­ment in people, process, and tech­nol­ogy, as well as a multi-year roadmap, may be warranted.

At the end of the day, any assess­ment of traffic/order/lead/revenue sources implic­itly includes some form of attri­bu­tion. That may be very basic, but it may be good enough. It’s almost certainly worth some explo­ration of increas­ing the sophis­ti­ca­tion of the attri­bu­tion approach, but that doesn’t mean an imme­di­ate commit­ment to spend six figures on an advanced, algo­rith­mic, multi-channel, offline/online solu­tion. That may be where you ulti­mately wind up, but it’s best to wade steadily into those waters rather than taking a blind leap off the 30-meter plat­form.

At Search Discovery, we’ve ranged from the shallow end to the deep end of the pool depend­ing on the client and their needs. We’d love to suit up with you, so contact us if you’re inter­ested in learn­ing more about our approach and expe­ri­ence.