How Broken Is Your Attribution Model?
Indeed, fellow search geeks, it’s not a matter of if your attribution model is broken, but how badly it’s broken. There simply is no such thing as a perfect model. Sure, some are better than others, and that will be the focus of this column.
But, first, let’s take a brief tour of the history of attribution…
In the beginning, there was the last click
The blessing and the curse of digital marketing has always been its measurability. When online ads first burst on the scene, the ability to track interactions via click-throughs was a key differentiator. We’ve been digging ourselves out of that hole ever since.
In the early days, successful interactive media placements were those that drove relatively high volumes of clicks and subsequent conversions. Little value was placed on those ads that did not lead directly to a purchase. And, by directly, I mean within the same browsing session. Extended cookie windows had not yet been, er, baked, so the recipe for digital success was click, convert, click, convert. Any variation simply did not compute.
As a result, many ad networks began introducing new ad units (my favorite was the L90 PowerAd, which launched the destination URL in a pop-under) in an effort to cookie-bomb the Internet and claim last-click-prior-to-conversion status.
Branding, we don’t need no stinking branding
But what about all those ads that did not click and/or immediately convert? Alas, they were thrown away with the bathwater, optimized out of campaigns and discarded into the series of tubes. Many agencies and publishers tried to get marketers to consider the “branding impact” of all those unclicked ads, but we had not been conditioned to think that the Web could be a branding vehicle. That was the domain of TV. Silly rabbit, you can’t create an emotional response in a banner. And certainly not in a text search ad!
Hey, a little assist here
Then some crafty publishers set out to prove that ads on their site should have value even if they didn’t elicit an immediate action. New metrics like “post-impression” and “post-view” were minted to give credit to ads exposed to consumers who ended up converting. Agencies leveraged Dynamic Logic studies (you’re definitely an old-school digital marketer if you remember those!) to determine brand impact through surveying test and control groups.
From a search perspective, Yahoo was among the first to really quantify the impact of the “non-converting” keywords when it introduced the concept of assists. By delivering reports that showed which keywords were clicked by consumers who eventually converted, search marketers could give credit to “upper-funnel” activity -- and bids on non-brand terms have increased ever since.
Are you a model?
Once the industry reached consensus that ad interactions not immediately producing clicks and conversions had some value, the race was on to create attribution models that determined just how much value to place on each interaction in the path to conversion.
There was the distribute-evenly model in which every ad clicked got equal credit. There was the time-decay model, which weighted clicks toward the end of the process more heavily than clicks at the beginning. There was the U-shaped model, which gave 40% of the credit to the last click, 40% to the first click, and evenly split all the remaining clicks in between. There was even (gasp!) the first-click model, which gave all the attribution credit to the first ad that captured the consumer’s interest.
What you don’t know will hurt you
Applying attribution models to reward every ad that contributed to a conversion was a huge step forward, but gaps still remained. Marketers were still using one-size-fits all models and applying them retroactively to their campaigns. In other words, you had to pick the model you felt best reflected how your customers behaved -- and then look back at all previous conversions and carve up the credit.
The challenge is that every customer, and every path to conversion, is unique. In some cases, the first impression makes all the difference. In others, it’s the last click that truly seals the deal. But how’s a marketer to know?
How can you tell if a conversion would have happened anyway, even if a particular ad was not served or clicked? Was the last click merely a navigation brand keyword search? Was the conversion from an existing customer -- and, if so, shouldn’t more credit be given to the ads that led to the original conversion?
A smarter path to conversion
Welcome to the new era of attribution. Gone are the days of static models applied retroactively. Gone are the days of setting bids based on assumptions of each ad placement’s actual value.
Today, the most sophisticated marketers are using dynamic attribution solutions that leverage machine-learning and algorithmic decisioning to drive advanced digital marketing bid optimization. (Bingo!) Dynamic attribution looks at each conversion path individually and determines the proper allocation of credit among the interactions. This methodology considers position in conversion funnel, causality and synergy, as well as customer loyalty. From there, bid updates are automatically triggered based on the true value of each keyword.
The result is a more informed marketer and a higher-performing campaign. Insights derived from dynamic attribution can be tremendously helpful in determining media mix allocation. Tests have shown that email and paid search brand terms often get too much credit for conversions, while display and paid search generics don’t get enough. In one example, campaign optimization that took into account the actual value of each interaction boosted overall revenue by 28% while lifting ROI by 17%. And performance only improves over time as the algorithms calibrate to business and market conditions.
Where do we go from here?
The attribution game is only getting trickier with the proliferation of new-media channels and device types. There’s no doubt that today’s most advanced attribution methodologies will eventually become obsolete. The one immutable truth is that no single attribution model will work in every instance. The first step to fixing your attribution is admitting that you have a problem.