Markov Chain Attribution: Crediting Channels by What Breaks Without Them

Most attribution models start by deciding a rule for splitting credit, then apply it. Markov chain attribution does the opposite. It starts with your actual conversion paths and asks a simple, almost surgical question of each channel: how many conversions would we lose if this one vanished? That question is called the removal effect, and it is the heart of the model.

What is Markov chain attribution?

Markov chain attribution is a data-driven model that maps every customer journey as a sequence of states (channels) and calculates the probability of moving from one to the next. It assigns credit to each channel using the “removal effect”: how much the overall probability of conversion drops when that channel is taken out of the system.

A Markov chain is, formally, “a model describing a sequence of possible events in which the probability of each event depends only on the current state”. In attribution terms, the “current state” is the channel a buyer is on, and the model only cares about where they go next, not the entire history behind them.

How it works

There are three steps:

Build the graph. Turn all your conversion paths into a map of states, from a “start” node through each channel to either “conversion” or “null” (no conversion). The transition probability between two channels is the number of paths using that step divided by all paths leaving that node.
Calculate the removal effect. Remove one channel from the graph and recalculate the probability of reaching a conversion. The percentage drop is that channel’s removal effect. Repeat for every channel.
Normalise. Because removal effects do not naturally add up to 100%, you scale them proportionally so they sum to one, then multiply by your actual conversions to get each channel’s credit.

A worked example

Imagine four observed paths, two of which converted. Removing Remarketing drops the conversion probability to zero (every winning path went through it), while removing Facebook only reduces it to one in nine. After normalising the removal effects across the channels, the credit might land at roughly Remarketing 41%, Facebook 32%, and Google 27% of the two conversions. Notice how the model rewards the channel the journeys genuinely depended on, rather than whichever happened to be first or last.

Strengths and weaknesses

Markov chains are a workhorse behind a lot of data-driven attribution, and for good reason:

They cope better with smaller datasets and more granular channel splits than heavier methods like Shapley value attribution.
They need less computing power, and the results tend to align with linear models as a sanity check.
They are less rattled by statistically insignificant random data.

The weaknesses are real, though. The approach carries “an embedded error resulting from the use of the removal effect” that can unfairly shift credit toward longer paths. And the deeper caveat applies to every model of this type: “any algorithm that use conversion paths as input data is based on the analysis of correlations”, which means it can describe what tends to happen but cannot prove what caused it. For that, you need incrementality testing.

Questions to ask yourself

As a modern growth or agile marketing professional, ask yourself the following about Markov chain attribution:

Do I have enough clean conversion-path data to make a probabilistic model meaningful?
Am I treating the removal effect as a useful estimate, or mistaking it for proven causation?
Does the model’s credit roughly agree with a simple linear model, and if not, why?
Would a Markov model actually change a budget decision, or am I modelling for its own sake?
Have I paired it with a holdout test to check the story against reality?

About Growth Method

Growth Method is the growth platform designed for experiment-led and data-driven marketers.

Learn more on our homepage, connect with me on LinkedIn or Twitter, or book a call here.