Credit Attribution: The Challenges
When a customer interacts with more than one marketing channel how should credit for the order be parsed?
Smart marketers all over the world are wrestling with this difficult question, and for good reason. Efficient marketing demands that we understand which ad buys are productive and which are less so. Perfect understanding would allow us to generate sales, new customers, and leads with maximum efficiency. Unfortunately perfect understanding is impossible, but we think reasonably good understanding is not far off.
RKG is wading into the fray with what we hope will be a game-changing solution. We think we have a system that fits the need in terms of flexibility, sophistication, and affordability. More on that in the coming weeks.
Before we talk about the range of solutions, let’s lay out the challenges all attribution systems face.
Any data driven process depends on having reasonably accurate information. Part of the challenge is collecting data on all the interactions a buyer has with different marketing initiatives. Even for an online pure play, this is not trivial.
- What is the value of a display ad impression, and can we tie them to an order number?
- How should we think about “unopened” emails?
For advertisers with significant direct mail activity more challenges arise:
- Can we match the mailing history to web orders dynamically at the time of checkout? Yes for many cases, but no for folks with very recent address changes or people on prospect mailing lists.
- How do we fold in match-back information? Since we often can’t link the person who placed a web order to our mailing list dynamically, we have to back-fill this information later. This creates a need to re-calculate the attributions for affected orders.
For advertisers with a wide brick and mortar footprint, who mail circulars, do TV and radio spots, have store signage, etc. we enter the realm of impossibility.
Clearly perfect is impossible, the challenges are material but getting closer to the truth still carries significant potential benefits.
The data is imperfect and messy, but once we have that data how we handle it is not trivial either.
We have to think about the following:
- Can we distinguish between brand and competitive/non-brand search for both natural and paid?
- Can we fold in results of lift tests for display ads, catalog mailings, email?
- How should we handle multiple interactions with the same ad? This is particularly important with display ad impressions.
- Should organic search on competitive phrases count as a marketing channel for this analysis?
- Over what time period should ads be active? Hard windows seem odd: the ad gets 100% credit if it’s within 30 days but at 31 that drops to 0? At the same time if the only ad touch was 82 days ago, should it get credit for the whole order? Part of it? If part, what gets credit for the rest of the order?
- Some channels cannibalize sales from others. With affiliates, control tests are impossible. That is, we’d like to know what fraction of folks who get to an affiliate site by typing in “[My brand] Coupon” would have bought from us at full price, with no commission, in the absence of that ad. Absent that test, how much should we depreciate those types of touches?
Perhaps the most important piece to consider: What actions can we take based on the information?
- Regardless of the credit given to email, because it is so cheap, do we even care what’s incremental and what isn’t?
- If it turns out that affiliates should only get 40% of the credit they’re currently getting, can we pay them accordingly or is it an all or nothing proposition?
The levers over which we have complete control and for which we spend significant dollars seem to be paid search, display advertising and direct mail. Shifts in the views of these programs have the greatest potential for improved resource allocation. If some of the other channels are stealing credit from these programs or vice-versa, the opportunities for impacting the P & L become meaningful.
Next week I’ll outline what might be viewed as “Good, better, and best” solutions for attribution management. A simple, basic solution may make sense for many, complex heuristic models will make sense to some, and more advanced statistical models may be worthwhile for others. Along with descriptions of each, I’ll toss out some other important considerations for folks thinking of implementing any of these.