ADOTAS – Digital attribution is hot, but it’s still a confusing concept for many. Given a lack of industry standards, and an ever-expanding list of attribution methodologies, it can be difficult for marketers to determine exactly what they need. So this piece intends to educate and enable marketers to optimize digital spend through advanced insights.
The Funnel and Attribution
Simply defined, attribution is allocating credit to each interaction that drives a desired action (visit, goal page view, conversion, etc.). Within this broad definition, there are two primary distinctions:
• Lower-funnel or click-based attribution incorporates assist clicks when allocating credit for conversions. Compared to last-click reporting, this is a step in the right direction. The limitation to lower-funnel attribution is that it severely discounts the role of display advertising while overstating the role of search, affiliate, email and other click-centric media. For online advertisers seeking a more complete picture, a full-funnel view is required.
• Full-funnel attribution builds on click-based attribution by incorporating assist impressions from display ads (video, rich media, flash and .gifs) when allocating credit for visits or conversions. Recognizing that display ads can be very effective, even in the absence of clicks, a full-Funnel attribution model is needed to quantify the true impact display ads have in creating awareness, consideration and preference.
Cross-channel attribution addresses the role of each digital channel (display, paid search, natural search, email, affiliate, etc.) plays in the customer engagement process. While conversion paths are interesting, they aren’t very actionable. To truly understand and optimize each channel, you must allocate fractional credit to each channel and placement that contribute to a measurable action. This generally results in a shifting of credit from non-paid channels (organic search, direct navigation, referring sites), back to paid media (display, paid search, email, etc.) that “fed” the non-paid channels. Before diving into weighting methodologies, let’s first look at leading approaches to attribution.
Three Approaches to Attribution
While there are many approaches to attribution, here are three you should be familiar with:
• Statistical attribution is based on traditional media-mix modeling, which relates to analysis of disparate data sets. In this approach, you would analyze three months of impression data and three months of conversion data and look for relationships between the data sets. At best, this approach can provide high-level directional signals. If you want granular insights into the impact of each channel, vendor, placement or keyword, you need a more granular approach.
• A/B testing seeks to attribute credit and validate causation by observing results from pre-defined combinations of media placements. A/B testing can be used to measure display’s impact on results from search, as well as the performance of a specific creative, vendor, market or channel. While A/B testing is a great way to observe directional insights, it’s nearly impossible to exclude or account for other factors (seasonality, competitors, macro-economics, weather, etc.) that might impact performance between the control and test groups.
• Operational attribution takes a bottom-up (visitor-based) approach to analyzing and allocating credit to impressions, clicks and visits that precede each conversion. With operational attribution, there is no need to calculate possible conversion paths – you have the actual data, which provides for a more granular and accurate data set for advanced analysis. (i.e. heuristic and/or statistical modeling). As with any approach, caution must be exercised to define the appropriate look-back window and cleanse the data set and exclude factors (e.g. wasted impressions and post-conversion visits) that might skew the results.
Manual vs. Statistical Weightings
With the framework of operational attribution, weighing assist impressions and clicks is both art and science. Here are two primary approaches:
• Subjective weighting of impressions and clicks: First-generation platforms require the marketer to define the rules for allocating credit to each interaction. While this approach is easy and flexible, it lacks statistical rigor and entails too much guesswork. Moreover, it allows the operator to influence outcomes through the assumptions that drive the model. But the biggest problem is that it allocates credit based on the actual number of assist impressions, rather than using observable data to model how many are actually needed. For example, you may find that 12 impressions preceded an average conversion, when only six impressions were actually needed. If you give credit for wasted impressions, you end up rewarding vendors for over-serving customers.
To reduce subjectivity and improve accuracy in how impressions and clicks are weighted, we must use machine learning and algorithmic modeling.
• Algorithmic weighting: To remove the guesswork in attribution can use machine learning and proven algorithms to calculate probability-based weightings for assist impressions and clicks. This removes the arbitrary nature of manual weightings and provides much higher levels of confidence and comfort for marketers.
There are numerous approaches to statistical modeling and there are plenty of vendors vying for the “best math” award. While it’s hard to say which approach is best, we have seen that most prefer transparency vs. opacity, and known algorithms to proprietary models. If you’re going to put your neck on the line to defend a new measurement standard, you should be comfortable with the approach and underlying assumptions. In general, a transparent, statistically validated approach is best.
There is an endless number of attributes you can seek to measure, including segment, format, audience demographics, creative concept, message, frequency, day-part, sequence, etc. While the idea of Metric Nirvana is appealing, it’s also very elusive if you try to get there overnight.
It’s important to recognize that the more granular you get, the more data you need and the more complex the setup, production, reporting and analysis. Many attempts to go directly from last-click to advanced micro-attribution fail due to the complexity of implementation and analysis. We all crawled before we walked, and walked before we ran. Analytics should be no different.
If you’re still using last-click metrics, start with channel-level and publisher-level attribution. Once you’ve identified your top performing vendors, delve deeper sequentially (as opposed to all at once) by looking at recency, format, creative and other variables that may have incremental impact on the end results. Just remember that with each incremental value, the scale, complexity, and risk of failure increase dramatically. So start with the low-hanging fruit, and work your way up the tree.
To sum all this up, we at Encore and MediaMind advocate the following best practices:
• Operational, full-funnel and cross-channel attribution.
• Use machine learning to weigh and allocate fractional credit to assist impressions and clicks.
• Improve your chances of success by starting with basics and getting more granular over time.
Hopefully you now have a better understanding of the attribution landscape and some of the distinctions within it.
Written in association with MediaMind