First thing first, let's start by defining quickly what marketing attribution is, what it's trying to solve and why it's not that straightforward.
As you know, a sale is usually the result of several marketing campaigns executed across many different channels. Each campaign has a certain associated cost and so does every communication with your customer. An email will be less expensive than a SMS, a SMS less expensive than a phone call, etc.
That cost, that investment, is supposedly producing an outcome that needs to be tied back to the campaign. Ultimately, the outcome is the sale itself. That means that all campaigns/channels "touching" the customer on his/her path to purchase, will want to claim a certain percentage of the generated revenue.
So, who is to decide who will receive the chunkiest part of the cake? That's where it's getting interesting because it's impossible to accurately define it. Imagine if you were to ask each of your customers: "Mr customer, between the display ad you've clicked on, the commercial insight you've read on the blog and the email you've received, can you please tell me precisely the influence that each one of those had on your purchase?". At best, you'd get a "guess-estimate" from them.
The reality is that the science of marketing attribution modeling is about "guess-estimating" that split.
But why is it so important to be able to accurately split the merit back to the right campaign/touch point? Because you might want to optimise the way you allocate budget across all those campaigns/touch points to deliver more bang for your bucks.
Now, this can seem obvious but before being able to do any kind of attribution, you first need to be able to connect the dots of your customer's journey. That means you need to deploy your marketing analytics technology across all your online channels and stitch on top your offline data coming from your in-person interactions with your customers if you have brick'n mortar agencies/stores for example. Depending on your technology provider, you might need additional tools to achieve that.
That part is not to be underestimated. Enterprises have many technology silos involved in delivering the customer experience and stitching the customer journey together can be an perilious exercise, hence the important of selecting the right technology partner to assist you with that.
But let's assume you've done it and let's have a look at what are the different approaches commonly used for attribution.
"Let's picture it" First, I'd like you to think of a timeline on which will be logged all interactions between your company and a customer who just purchased/transacted with you. The beginning of this timeline is the first interaction (=touch) and the end of the timeline is the transaction itself which generated a certain revenue for your company.
To illustrate more easily the different types of models, let's consider the following example: a customer, Michael, is starting his journey searching on google about a problem he's experiencing.
The first touch model, as you can guess, is attributing all the sale's revenue to the first interaction, so in our example, to the blog post discovered via the organic search.
The last touch model, on the other end, is attributing the entire revenue to the last interaction prior to conversion, meaning the remarketing email. Some variations of the last touch model exists, like the non-direct last touch model which excludes all direct traffic from the consideration. Considering direct traffic is important in my opinion as it is a manifestation of your brand's reputation and strength. The problem is that it's sometimes the default bucket used by web analytics tools when no referral is found which can skew the model.
The last touch model is probably the most widely used attribution model. "Why?" you're going to ask... "it's too simplistic and it obliterates the important work done upstream by the content marketing team and the offsite social targeted ad". I guess the answer lies within its actual simplicity. It is usually available in all analytics tools and is simple to implement... The real question is more: "Is it good enough?".
Obviously not, because if we follow the data from that attribution model, the next thing to do is to cut down the content marketing team's budget. The reality is that the sale would have never happened without the right content building progressively the customer's awareness and consideration towards your brand and your product. So let's have a look at the other options.
The linear model is splitting equally the merit/revenue from the sale across the different touch points. That seems like a more "fair" approach compared to the last touch. However, if it is not going to put at risk the fragile balance of your conversion path, it will also not be of great help to optimise your budget's allocation across touch points. Note that this model is also called "even" by some experts while others associate it a weighted model...
The time decay attribution model is in the same vein as the even/linear model but it adds another dimension: a recency weighting. That means that the most recent interaction will receive more credit than the last but one and so on. The interesting part is that the further away the interaction happens, the less credit it will get. So if for example the blog post was read months prior to the conversion, it will get less credit than if it had been read a couple of days before.
This is a top favourite amongst marketers but would be questionable for long and variable sales cycles like in the B2B space IMHO, unless you can tweak the algorithm accordingly to match your average sales cycle.
So the manually weighted model is going to allow you to pour your business knowledge into the attribution model. As the name suggests, you'll be defining manually how much % of the revenue is allocated to each and every interaction.
A possible implementation of this model is the "starter player closer" model. In such case you allocate a certain credit to the first interaction, let's say 20%, a certain credit to the last interaction, let's say 30%, and the other interactions get an equal part of the remaining 50%.
Even though determining the actual percentages still feels a bit like a gut-feeling exercise, it will be as good as the knowledge and the understanding that you have of your customers.
The "W" Model is also relatively popular for B2B use case and consists in attributing like 30% to the first and last touch, 20% a "bridging" touch point in the middle of the journey and split the rest across the other touch points.
Many data-driven attribution models available within existing analytics softwares are implemented on well known algorithms like the Shapley Value method for example. You can find much more details on this algorithm as well as on the Markov Chain method in this very good article from Trevor Paulsen, Sr Product Manager for Analytics at Adobe.
The Shapley based algorithm typically uses previous conversion rates from each campaign to calculate the weight of each campaign within a sequence of campaigns. If you have potentially a high number of different channels and campaigns, then it will obviously deliver superior results than any manually weighted models.
Now, where is the AI (Artificial Intelligence) and ML (Machine Learning) in all of that? Good question. Even though I am persuaded a lot of tech companies are currently working on it, I haven't heard of the silver bullet being released yet.
But imagining the art of the possible, you'd expect such model to be:
What is most likely though is that those AI driven models will need a lot of data, and more precisely clean and connected data across all the different silos. So that's certainly something you can get started with if you haven't connected the dots yet.