Marketing Attribution Modelling – determining the optimal mix
Digital marketers have access to a wealth of data about every marketing communication sent out (Impressions shown, Videos Seen), every user action generated (Clicks, Visits) and eventual outcomes drove (Sales, Leads, Downloads, Registrations). On one hand, this is great because never before have marketers have had access to such a humongous amount of detailed information about each and every marketing message and how it affected their target customers. On the other hand, this leads to a very different problem presenting itself – that of understanding which of the messages actually influenced the outcome and which ones did not.
This is the question attribution modeling attempts to answer. The choice of attribution model will define the value that a marketer sees in a specific marketing action. Consequently, attribution models allow a marketer to determine optimal budget allocation., i.e. given a $ 100, what would the optimal mix be in order to drive a specific outcome.
This blog will take a quick look at some of the attribution models that are used in the market today, look at the pros and cons of each model and comment upon how these models are likely to evolve over the medium term.
Part I – Attribution models
There are 3 broad types of models which are used to address this question –
1) Simplistic Attribution – Also called Single Source Attribution, these type of models will generally attribute the desired outcome to a single event. This constitutes of models like –
– First Click Attribution, where the sale is attributed to the marketing channel or event that was responsible for the first click which brought the user to the site or
– Last Click Attribution, where the sale is attributed to the last channel/event that preceded the desired event.
2) Fractional Attribution – These type of models allow for the credit for an outcome to be shared across multiple channels/events. This constitutes of models like –
– Linear Attribution. where the credit for every outcome is distributed equally to all actions (Clicks or impressions) preceding it. i.e. if a user clicked on a SEM ad, on a display banner, and on an email link before buying a product, the credit for the sale is given equally to all 3 (SEM, Display & Email)
– Time Decay Based attribution – This model gives greater credit to channels/events which take place closer to the point of sale i.e. if a person clicked on a SEM ad 15 days before the sale and then clicked on an Email 30 minutes before buying, the Email Ad would be given vastly more credit than the SEM ad, though the SEM ad too would be granted some credit based on certain assumptions made about the average time to purchase that are made by the marketer.
– Position Based Attribution – This model grants a fixed amount of attribution to channels in each stage of the funnel e.g. the channel bringing in the first click would be attributed 40% of credit of a sale, the channel for the last click another 40% and all intermediate channels/events share the remaining 20%. The choice of 40%, 20%, and 40% is a marketing choice made based on the perceived optimal split between first, last and intermediate events (impressions, clicks)
3) Algorithmic modelling – These models do away with fixed weights for first and last click events that the previous models rely on. Instead, these models use proprietary algorithms to determine how to attribute credit across events/channels.
Part II – Pros and Cons
Great, so now we not only have access to a trove of data but also have a multitude of ways in which to infer value from this data. This brings us to the question of the relative merits and demerits of each model. A quick check is marked below –
Part III – Models in the market and how these are likely to evolve
The market today largely relies primarily on Last Click attribution models since these are easier to get alignment on across stakeholders. Needless to say however these leave a significant amount of value on the table for most businesses since the budget allocation decisions are likely to be skewed in favor of events/channels which work in the final stages of the customer journey. Quick explanation for the uninitiated –
Let’s say, a customer goes through 4 stages as part of a purchase funnel – Awareness, Intent, Decision, and Action.
Awareness – Customer is made aware of the offering
Intent – The customer feels the need to acquire the offering
Decision – The customer evaluates alternatives and makes a decision
Action – The customer actually takes the necessary action (to acquire the offering/product)
Different channels have different strengths and perform differently in each stage of the funnel. For e.g. Display might work great in the first phase of the funnel (awareness) but might not fare as well in the final stages. On the other hand, email and search would perform significantly better in the final stages (Decision, Action) once the customer has been made aware of the offering.
To place emphasis entirely on the last step of the funnel (Action) would be equal to saying that the Goal-Scorer is the single most important person in a Soccer Team! Read an article that takes on this question.
While we are seeing the emergence of position based models, these leave a lot to be desired since weight choices are arbitrary to a large extent. Also, the models do not factor in cross-channel effects. For instance, reports from studies by the Atlas institute[“Where can you find your customer”] have shown that users who are exposed to both Displays and Search convert 22% better than users exposed to Search only. However, this channel-specific behavior is hard to incorporate into fractional models.
Algorithmic models offer to simplify a number of the issues that are associated with their more primitive modelling counterparts, but have 2 broad areas that need to be addressed before they can really take wing –
1) Statistics related know-how within organizations – Unless organizations are in a position to understand and challenge the approach used and results that are generated by Algorithmic attribution, there will be little buy-in.
2) Ecosystem-related constraints – Some of these models rely on using clickstream data. Still, others rely on impression level information being available in order to credit an outcome to a channel/event. The ecosystem today is not transparent enough to always be able to provide this data resulting in data gaps which would be difficult to fill.
To summarize – While there is a lot of data available and there are approaches which allow for us to interpret this data, there are a number of subjective decisions which will influence how a marketer optimizes his spends. These choices form the competitive edge because a marketer who understands how best to use the budgets and channels at his disposal is likely to be the one who outlasts his competition.