All About Attribution Models

Data Analytics

One of the most challenging areas in digital marketing today is how we can account conversions to different marketing channels fairly. Why does it have to be fair? Businesses spend a lot of marketing budget on different channels and by allocating the conversions fairly, businesses can focus budget on channels that really brought them returns. 

Attribution Modelling is a term that is brought up a lot and is related to this problem. Google and a lot of the other marketing platforms offers similar analytic tools on their platforms, however, we often found these platforms don’t agree with each other and offer different numbers. So which is the most accurate? In this post we’ll try to provide a quick outlook on what the attribution model is, how it can be applied, the challenges involved, and how they are useful.

So firstly, what are attribution models?

Attribution Modelling is a set of rules that are used for attributing the value of a desired outcome to some prior events in marketing. The idea is we treat an outcome as a value, (e.g a value of 1). Then we split this value to multiple particular past events (e.g. event A accounts for 0.5 of the outcome). It is often used in multi-touchpoint analysis or multi-channel analysis.

Attribution Models come in a lot of forms. The simplest form is the single source attribution. For example, if Joe studies before the night of the exam, he passes. Otherwise, he fails. Hence, loosely speaking, we say Joe passing the exam depends on whether he studies. (In a proper probability definition, event[Joe studied] and event[Joe passed] are just occurring together. That doesn’t justify the causality! )

Things get complicated when we try to consider more events. In our example,  Joe could be smart so he doesn’t have to study to pass. Or the exam was easy. All these other factors may have a hand in Joe’s exam result. So can we split the marks that Joe gets on his exam to, say 50% because he is smart and 50% because he studied? Attribution models are the rules for splitting the result and attributing it across the factors that “contribute” to this result.

Why does it matter?

With businesses shifting to digital, data is more readily available for analysis than ever. You are able to track the source of a visitor, what page they have read and even the actions they took on your website. The issue is internet is completely open. And often it takes more than one interaction with the visitor before they make any conversion. How we allocate the contribution will have a strong impact as to whether we will find a channel or a touchpoint.

As mentioned earlier, attribution models are the rules for allocating the conversion to channels or touchpoints. We try to allocate the conversion to the touch point. This provides a reference for where we should put our resources in and, in theory, we can calculate which investment is bringing us more return. We will discuss this more in a later section.

Types of attribution models

There are several common strategies that you can use to split the contributions. You should choose based on which fits the business scenario you are in best. These strategies can be divided into three main categories: Single Source Attribution, Fractional Attribution and Data Driven Attribution. Each of them have their pro and cons. There is no one model that is perfect – it all depends on the question we are trying to answer.

Single Source Attribution

A Single Source Attribution model assign all credits to one event/source. For example, event[Joe Pass] only occurs because of event[Joe Study] occurred. In digital marketing, the last or first click model belongs to this type of attribution. The benefit of this model is it’s simple to compute and understand. It is useful for customer journeys that are direct and short. 

However, this simplicity can hurt for more complicated customer journeys. It assumes the other events/sources have zero contribution to the result. For example, for a new in-market product, a customer probably doesn’t know what to search on Google without seeing the ad on social media first. So if the customer learnt about the product from social media ads and buys the product through a search session on Google, we can argue that the social media ad contributed more to the product sale in this case.  But a last click model will account this conversion to Google Search and ignore the impact from the real driver, which is the social media ad.  It can be disastrous if we relocate all our resources from social media to Google Search just based on this.

Fractional Attribution

Instead of assigning all credits to one event, fractional attribution distributes credit to multiple events based on rules. A Linear attribution model allocates equal share to all accountable events. A Time Decay attribution model allocates credit based on the time difference between the occurrence of an event and a result. Position based attribution allocates credits based on the order of the events. If you are implementing these models for your digital marketing campaigns, Google Analytics provides the implementation. Avinash Kaushik has a really good overview on what models Google Analytics offers.

The benefit of a Fractional attribution model is every event get credited in some way. However, it requires strong knowledge and planning about the customer journey to apply the right rules. Different rules can hugely impact on how an event is being attributed. For instance, before you can apply a position based model that allocates credits to the first and last click, you must first know that you need social media to grow your brand awareness before a customer can search your brand on Google.

Data-Driven Attribution

In the last two categories, the rules are set before the conversion is allocated. It is rather subjective as it required a certain understanding about the customer to decide which model is the fair one. Data Driven attribution takes another approach. Instead of deciding how the weight of each channel is allocated, it uses data and creates a statistical model to allocate credits.

The most popular methods is using shapley value and markov model. The details of these models are too much for this blog, so for now all you need to understand is the idea that these models are trying to come up with an attribution that is fairer through a machine learning algorithm.

The advantage of these models is that they allocated credits based on statistical knowledge rather than individual decisions. It also updates if the environment changes. However, the computations can be costly in their implementation. The quality and availability of the data will also have an significant impact on the model’s fairness.

How can Attribution Models be used?

  • Digital marketing

For digital marketing, as previous mentioned, you can use an attribution model to compare the performance between platforms. Google Analytics offers an attribution modelling tool within the platform itself. It also allows you to compare how each channel is being attributed based on different attribution models. Some marketing platforms, such as Facebook Ads, also provides an attribution model. However, different platforms sometimes do not agree with each other. 

If you have an awareness strategy with your campaigns, you can use attribution to credit campaigns that focus on the early stages of the sales funnel. Similarly for the content on your website, conversions can be attributed to what pages the customer has been through.

  • Customer Journey

For businesses that have a long sales cycle, your customer may interacted with you a lot of times before the deal is closed. Can we attribute these interactions? Yes, if these interactions are being logged in your CRM, such as Salesforce, you can attribute these interactions. This makes your EDM, Digital Marketing, Event and Sales Calls comparable to each other. 

  • What impacts the model accuracy and interpretation?

After choosing a model, is that the end of story? The answer is no. Indeed, choosing the type of model is just part of the story. There are other decisions that will affect the fairness of the model. Different platforms often shows different results, even for the same attribution model. What creates this discrepancy? This section focuses more on what details created the discrepancy and how you should change your interpretation of the model if you meet these.

  • Type of action

The simplest form of action is a transaction. It has the most direct impact to the business’s performance. But if we only look at transactions, we might skip the actions that assisted the transaction as it is usually at the end of the sales funnel. Assisted action can be the customer signing up to a newsletter or claiming coupon. Sometimes knowing what gets people to sign up is more useful than looking at what leads to purchase. For example, when a business runs successful remarking campaigns. Every customer signs up for the newsletter before making any purchase. Although these customers did not buy in the session, signing up newsletter may mean future remarketing potential and purchase.  

  • Events / Channels / Touchpoints

Similar to action, deciding which events/channels/touchpoint we are going to account for affects how we interpret the results. For example, if every customer signs up and comes back from a verification email before making any purchase, then the verification email will get a lot of credits in this scenario. But it has not done much in influencing the customer’s purchase decision. We  should take out that particular EDM from the model. Another example is deciding whether to skip all free traffic and focus only on paid channels in our model, as those are where the money is being spent.

  • Conversation window

Conversion window is the period of time prior to the occurrence of the result that we consider that an event had an impact to the result. For example, if a customer saw an ad of a restaurant a year ago and only dines in recently, are we still holding that ad accountable with this action? Probably not.

Different products and industry have a different conversion window. It usually depends on the time a customer needs to make a purchase decision. FMCG, like protein powder, can have short windows whereas buying a car or property can have a long window. Setting the conversion window too long may include events that are irrelevant. Setting the conversion window too short will overlook some early events that triggers the customer to start considering you.

  • Tracking

How data is tracked is often overlooked.  Google Analytics, for example, only collects data when the visitor has a session on the web site. If a customer saw an ad on Facebook then enters the web site through Google Search, only Google Search will be attributed. Other marketing platforms can track impressions. And if conversions are tracked differently, it is obvious if they have a different number of conversions. 

Brand impression’s impact to the customer is also untrackable or comparable. If a channel only provides impressions to the brand without getting most of its viewers to enter the website, that channel will not be credited. 

Is there a single attribution model that we can all use?

The short answer is no. Attribution models are not a magic wand. Getting the perfect tracking is hard, especially as a lot of assumptions behind the models are often not true in reality. Blindly trusting these models may lead to wrong conclusions. It requires marketers to make careful judgments on when to use and when not to use them.

However, attribution models can still provide invaluable information if they are used correctly. For example, by comparing first click model and last click models we can pinpoint which audiences are closer to the buying decision and we can adjust the bidding strategy on channels.

How should you start?

Step 1: Strategy

  • Make sure you define your marketing objective
  • Define the action/conversion you want to track and how they contribute to your marketing objective
  • KISS – “Keep it simple stupid”. You won’t want to spend more money on setting the model than your marketing spending.

Step 2: audit your tracking and data

  • Audit your existing tracking and make sure the data is answering the questions

Step 3: Apply data to the model and evaluate the option

  • Don’t just stick to one model
  • Try compare several models and look at what made the differences

Step 4: Apply it to your marketing platform

  • Decide the best fit of your strategy and try to use it for your decisions. 

Step 5: Review after certain period

  • As the environment and customer behavior changes over time, it is good to repeat step 3 after certain period of time

Like this blog post? Sign up to our email newsletter – Lab Report – and never miss a new one. Or, get it sent straight to your Messenger!

Reminder: Google UA Historical Data to be Deleted in July 2024