It was august 2009 and I was invited to a meeting with Marketing managers of a multinational company. End of year was coming and all of them where working on their budgets for the next year.
Each director was supposed to present their proposal and then distribute the marketing budget based on the potential revenue each area could generate for the company (paid search, SEO, TV, email marketing, etc).
Each director made his presentation trying to convince all the rest that his/her area had the best potential and, hopefully, get the biggest budget. It was really interesting that each of them used for that propose a different attribution model. They basically checked which attribution model was showing the best result for their area and came to the meeting with that information and a very “solid” argument on why that attribution model was the best one.
But, why they did that? The answer is simple. Because they can. All the standard attribution models are equally the best and equally the worst because they are not base on information. And when I say “all of them” I’m referring to both, single and multi touch attribution models. What are they? Let’s go through all the standard attribution models:
Single touch attribution:
- First click: This model attributes the conversion 100% to only one channel. The one that brought that user to the site for the first time.
- Last click: It’s the opposite of the first click. It attributes the conversion 100% to the channel that brought the user when it converted (purchase, registration, etc).
Multitouch attribution:
- Time Decay: This attribution model takes into consideration all the measured touch points giving the touch points that are closer to the conversion more weight than to the older ones.
- Position Based: In the Position Based attribution model, both the first and the last touch points gets more weight (40% in the case of Google Analytics) and the remaining (20% in the case of Google Analytics) is distributed evenly to the middle interactions.
- Linear: As all the other multi touch points attribution models it considers all the measured touch points but in the case of this model, it weights all the touch points evenly.
All of the above attribution models share something in common. They are not necessarily supposed to represent the reality but instead to allow you to have a homogeneous way of measuring activities in a comparable way within the company.
Attribution models became more popular in 2005 when Google purchased Urchin and developed and launched Google Analytics. At that time Google strategy was visionary, if Google was able to help their Adwords and Adsense users become more effective, Google was able to attack the (very famous at that time) long tail, generating a huge business. So that’s what they did, and by launching Google Analytics they democratized data and created the Data Analytics market as we know it today. See how the [analytics] keyword stared to generate searches from that time on.
So in 2005 Google Analytics was not a solution for corporations, was never meant to be one at that time. It was supposed to be simple and very intuitive so bloggers could generate more Adsense inventory and small businesses could easily invest in Adwords.
By April 2008, in an illogical move, even for 2008 standards, Yahoo! acquired Indextools (a corporate Analytics solution) and launched it as free solution. At that point you had free solutions for SMBs and paid solutions for corporate use. That April everything changed for ever. Competition begun and some months later Google Analytics accelerated their pace on launching new features to convert their SMB solution in a powerful solution for the big guys. In 2008 I was invited by Google to present 11 new features, all corporate oriented.
In 2008 the term [conversion attribution model] begun to gain interest among both SMBs and Corporations as we can see in the following Google Trends.
Some companies kept working and measuring themselves with standard attribution models while other companies with a more sophisticated vision realized they needed something else. Why? The reality is very complex. Data allows us to approach reality with some acceptable level of risk. Companies are systems, a set of parts that interact together with a common objective (making money, today and tomorrow). Companies are live systems because everything is changing all the time, and every time something change, other parts of the system modified their behavior as well.
So once we understand that companies are complex systems with several parts interacting together we can have an idea on why we should never use a single touch model. How can we attribute 100% of a conversion to a single action? Is just not possible and not representative of the reality at all. It can definitely allow you to compare things, but what’s the point anyway?
Now let’s move to the multi touch attribution world. MTAs is more effective since it measures more interactions from the “system” (the company’s marketing efforts). But that’s not enough. Most MTAs today use either Shapley Value or Markov model. Let’s take a closer look at them.
Shapley Value: In game theory, the Shapley value is a solution concept of fairly distributing both gains and costs to several actors working in coalition. Game theory is when two or more players or factors are involved in a strategy to achieve a desired outcome or payoff. The Shapley value applies primarily in situations when the contributions of each actor are unequal, but each player works in cooperation with each other to obtain the gain or payoff. So basically, the Shapley value is the average expected marginal contribution of one player after all possible combinations have been considered. While not perfect, this has proven a fair approach to allocating value.
Markov Model: The Markov model is normally used to model randomly changing systems, so it makes total sense to use it for Conversion Attribution. Is a probabilistic model, which focuses on specific calculations of the chance that an interaction in one channel will transition to a different state, in this case a conversion.
Those two are models to approach the reality (that’s impossible to know entirely) in a simplified way. Even when they are still not the reality they are based on SOME events from the reality and, ergo, are decision making models with acceptable level of risk. In reality a conversion journey would be little more complex, something like (with probabilities in between touch point):
Now, how can we reduce more risk? These two models are MTAs (or multitouch attribution models) that can still be standard models. Which means that they will take the same touch points from the same information platforms for all companies using it. So, what’s the problem with that?
Well, the problem is that in the digital world visits are temporary cookies, visitors are browsers and pageviews are loaded javascript tags (just to mention some examples). So in a world were touchpoints are completely configurable the changes that other company’s model works for ours is rather odd.
So what’s the best of all worlds? The best is having a model that’ve been tested with several companies and at the same time allows a fine tuning based on our own data. The attribution model should represent the reality with the lower level of risk possible. What reality? Your own reality.