Markov Chain Attribution: Channel Paths, Removal Effects & Fair Credit

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Turn 14 customer-journey paths into the transition count table of a Markov chain: first touches from Start, channel-to-channel steps, and journeys ending in a conversion or in Null.
• Divide the transition counts by their row totals to get the probability matrix P, add the absorbing Conversion and Null rows, and use MMULT to get two- and three-step probabilities.
• Walk a visitor through the chain ten steps with MMULT, watch the conversion probability converge, then get the exact value with the fundamental matrix (I − Q)⁻¹ and compare it with the observed conversion rate.
• Delete one channel from the chain, recompute P(conversion), and turn the drop into a removal effect for every channel with a single masked MINVERSE formula.
• See why raw removal effects double-count, normalise them into shares that add up to 100%, and credit each channel with conversions and revenue.
• Calculate last-click and linear credit from the same journeys, set them next to the Markov credit, find the channel last click undervalues most and chart the three models.
• Price each channel's conversions under last click and under Markov, compare CPA and ROAS, and see what a budget split by Markov credit would do, and why that is only a starting point.
• Re-run the attribution on three months, measure how much each channel's share moves with the coefficient of variation, flag unstable credit and size the uncertainty in orders.