Analytics and Tracking

Markov Model

Also called Markov chain attribution, removal effect model

A probabilistic attribution method that credits each channel by how far conversion odds fall when it is removed.

Quick facts: Markov Model

Category
Analytics and Tracking
Also called
Markov chain attribution, removal effect model
Level
Advanced
Affects
Credit allocation, upper-funnel budgets, channel cutting decisions
Where to see it
Custom modelling on exported GA4 journey data, open-source attribution libraries
In this article4
  1. How a Markov model works
  2. Why Markov models matter
  3. Where Markov models go wrong
  4. How to act on it

How a Markov model works

Start with every recorded customer journey: organic search, then email, then paid social, then either a purchase or nothing. A Markov model turns those journeys into a map of states and the probabilities of moving between them. From any given channel, it knows how often the next step was another channel, a conversion, or leaving altogether.

Credit then comes from what is called the removal effect. Delete one channel from the map, rewire the journeys as though it never existed, and recalculate the overall probability of reaching a conversion. The amount that probability drops is that channel’s contribution. Repeat for each channel and share the credit in proportion to the drops. A channel that nothing depends on can be removed with almost no effect, and it is credited accordingly.

Why Markov models matter

They value a channel by the role it plays in a path rather than by its position in it. Last-click attribution is blind to anything that started a journey; a Markov model can show that removing an early discovery channel collapses conversions further down, even though it never appears at the end.

The removal effect is also easy to explain to a client, which matters more than it sounds. “If we take this channel out of the map, this share of journeys never reaches a purchase” is a sentence a business owner can act on, unlike an unexplained algorithmic score.

Where Markov models go wrong

The mathematics assumes the next step depends only on the current one, not on everything that came before. Real buying does not behave that way — someone who has already read three articles is in a different frame of mind from a first-time visitor at the same page. Higher-order versions ease this, at the cost of needing far more data.

The larger problem is the same one every model-based method shares: removal in a spreadsheet is not removal in the market. Simulating the deletion of branded search will show a large drop, because most journeys pass through it, yet those customers would mostly have found you anyway. Only an incrementality test distinguishes a channel that carries traffic from one that creates it.

Data coverage is the practical limit. The map is built only from journeys you managed to record, so consent refusal, cookie expiry, device switching and closed platform reporting all leave gaps, and the missing steps quietly hand their credit to whatever was measurable.

How to act on it

Use it as a corrective, not a budget rule. If a channel receives almost no credit under last click but a large removal effect under a Markov model, that is a reason to look again before cutting it — and a reason to test it properly rather than trusting either number.

Confirm you have enough complete journeys before commissioning one; on small volumes the estimates move every month and nothing can be concluded. For most advertisers the sensible order is to get tracking and reporting consistent first, judge results at business level, and reach for path modelling only when there is enough clean journey data to make it meaningful.

Do and do not

Do

  • Check you have enough complete journeys before modelling
  • Explain results as a removal effect, not a score
  • Test any channel the model recommends cutting

Do not

  • Treat a simulated removal as a real-world result
  • Ignore journeys missing through consent or cookie loss
  • Set budgets from one model run

Questions people ask about this

What is the removal effect in simple terms?

It is the answer to a what-if question asked of your recorded data. You take one channel out of the map of customer journeys, redraw the paths as though it had never existed, and see how much less likely a conversion becomes. The size of that fall becomes the channel's share of the credit.

Is a Markov model better than data-driven attribution?

They belong to the same family and often reach similar conclusions. The advantage of a Markov model is transparency: the removal effect can be explained and audited, whereas platform attribution is largely a black box. The disadvantage is that you must build and maintain it, and it can only use the journey data you were able to export.

Can it prove which channels to cut?

It can shortlist them, not prove them. A low removal effect means the recorded journeys did not depend on that channel, which is a good reason to investigate. Before cutting, run a holdout or geo test on the channel in question, because a model can only describe the paths it saw, never what buyers would have done instead.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.