Understanding Attribution: The Markov Model

15 Jul 2019

Data Analyst

Understanding Attribution: The Markov Model


Humans are constantly, and often instinctively, attributing value to people, objects and events. But as with so many of our instinctive judgements, deeper investigation reveals them to be shockingly inaccurate.

In football, how often have you seen the goal scorer be given all the credit for a goal?

In politics, how often have you seen a leader take all the blame for a disastrous policy?

In everyday life, we find heroes and scapegoats and attribute vast and often inaccurate proportions of value to them. We do this mainly because it’s the easiest, simplest way to think about the world. If every time someone scored a goal in a football match the pundits analysed how much value should be attributed to not only every member of the football team, but also every member of the physio team, or even every fan at the game, then the show would go on all night…

In the world of marketing, instinctive attribution or the reliance on overly simplistic attribution models can cause misconceptions about what’s really working in your marketing strategy, leading to budget being spent in the wrong ways and brands failing to capitalise on huge opportunities.

On the other hand, complex attribution modelling projects can be very difficult, time consuming and costly.

So how can you find a good middle ground?


Markov Attribution Model

We recently undertook an attribution project for a client using the Markov model.

The project aimed to advise the client on how their spend could be distributed better amongst their different marketing channels, and sparked conversations about how attribution modelling can be built into future strategies.

The most common attribution models only involve a single touchpoint e.g Last Touch, however in the data world we live in we need to move towards a data-driven attribution (DDA) model, such as Markov. A DDA model will fairly assign credit to all of the  channels involved in the journey to a conversion.

Example: Organic Search > Paid Search > Direct – (1 conversion), under a last touch model Direct would get the credit for the conversion, however a DDA model will look at all the different conversion paths to see the overall effect of the different channels and realise that Organic Search appears towards the beginning very often so it should deserve more credit, the overall result could look something like this:


We shall now see how we calculated these weightings.

Markov aims to highlight the assisting channels towards a conversion; we’ve explained the methodology using a dating/marriage model.

Imagine we had all the data possible from a dating app (e.g Tinder) and examined all the different journeys each person had, their relationship milestones and if they got married.

We would look at how likely people were to move on to the next step, e.g swipe right and go on a coffee date (20%) or from dating to a relationship (15%) etc. A simplified journey can be seen below:

Once we have our network set-up, we can calculate the chance of marriage considering all the different routes. The probability of marriage comes out as 0.027%.

To measure the importance of each milestone we use the removal effect, this is based on the premise that you don’t realise how important something is until it’s gone. We remove each milestone one by one. Let’s start with removing a coffee date:



We notice that some pathways are no longer possible without a coffee date, and after calculating the new probability of marriage we get 0.011%. We can see that the probability of marriage has reduced by 59%.

We repeat this process for all the different milestones to get dinner date (70%), relationship, living together and proposal at 100%. You can start to see which steps have more impact than others, however, to see how they all contribute to the marriage we need to normalise them so they can add up to 100%.



Under the different attribution models the credit is distributed differently, Markov Attribution enables all the different milestones to gain the credit they deserve.

In practice this is applied to digital marketing using Google Analytics multi-channel funnel data and the statistical software R.

Our project helped highlight that paid search was performing more than x2 as better than had been recorded. Findings like this help shape future budgets for the client and invite further questions which we can help solve.



Proving the worth of a channel within a user journey is paramount to the success of any digital strategy. Businesses need to understand the roles of every channel and how they work together in order to measure true performance; The Markov attribution model helps reveal this.

Working with Google Analytics data as a starting point will help you make the first step into attribution modelling without the breaking the bank and still getting some valuable insights. When it comes to data, it is important to look at the whole journey, where am I trying to get to and what are the steps that will allow me to get there. The journey to clean and insightful data can be long and tedious, therefore starting small and constantly improving your process will make it easier.



Data Analyst

As Data Analyst, Milan is dedicated to applying his sharp maths and stats skills to drawing keen insights from all kinds of data, which are fundamental to the planning, buying and optimisation strategies for our performance campaigns.