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.