Markov Chain Attribution Explained: How It Works and Where It Misleads

What Markov Chain Attribution Actually Is

Markov chain attribution borrows a tool from probability theory. It models the buyer journey as a series of transitions between states, where each state is a marketing channel (Paid Search, Email, Organic, Direct) plus two special states: Start and the absorbing states Conversion and Null (no conversion).

The model's defining assumption is the Markov property: the probability of the next step depends only on the current state, not the full history that preceded it. A first-order Markov chain "remembers" only where the customer is right now, not the winding route they took to get there.

From thousands of real journeys, the model builds a transition matrix: the probability of moving from any channel to any other, or to Conversion. Then it computes credit using the removal effect: remove one channel from the graph, recalculate the overall conversion probability, and measure the drop. The bigger the drop, the more essential that channel was to the network of paths. This makes it a form of algorithmic attribution and a close cousin of Shapley value attribution.

How the Removal Effect Works (Step by Step)

Here is the workflow a Markov model runs, and the one you would replicate if you built it yourself in Python or R's ChannelAttribution package.

1. Collect path data. Export every converting and non-converting journey as an ordered sequence of channels. In practice this comes from a GA4 export or a data warehouse.
2. Normalize the channel names. "facebook", "Facebook", and "FB / paid" must collapse into one state. Inconsistent UTM parameters silently fracture a channel and corrupt every downstream number.
3. Build the transition matrix. Count how often each channel leads to each next state, then convert counts to probabilities.
4. Compute the baseline conversion probability of the whole graph from Start to Conversion.
5. Remove one channel (route its traffic to Null) and recompute the conversion probability.
6. Measure the drop. The percentage decrease is that channel's raw removal effect.
7. Allocate revenue to each channel in proportion to its normalized removal effect.

The output looks reasonable, scientific, and far better than last-click attribution.

The Markov Trust Ladder

A Markov result is only as trustworthy as the data behind it. Use this four-rung ladder to decide how much weight to put on the numbers before you move budget. No published threshold says how many conversions is enough, because the data you need grows with the number of channels in your paths, so the ladder tests stability instead: fit the model on two consecutive months, and if the channel ranking flips while spend barely moved, you're on rung 1.

The uncomfortable part: Markov hands every channel a precise-looking percentage on every rung, and nothing on the screen says which rung you're on. Precision is easy to mistake for accuracy; the ladder exists to break that illusion.

Markov vs Other Attribution Models

Markov sits with data-driven attribution and machine-learning attribution: a major upgrade over rules-based models, but living on the wrong side of the correlation-versus-causation line.

Where Markov Quietly Misleads DTC Brands

1. It is correlational, not causal. The removal effect describes how channels co-occur on paths to purchase. It cannot tell you whether a channel caused incremental revenue or simply appeared on journeys that would have converted anyway. This is the same trap that makes multi-touch attribution fail: observing a touchpoint is not proving its lift.

2. It cannot see demand it did not capture. If a customer saw a YouTube ad, never clicked, and bought a week later via Direct, Markov often credits Direct: the un-clicked impression rarely enters the path data. Channels measured only by view-through conversions are systematically distorted, just as they are in view-through ROAS.

3. It rewards channels that platforms already over-claim. Feed Markov path data assembled from platform pixels and you inherit self-attribution bias: every platform grades its own homework, and Meta, Google and TikTok each insist they were on the path.

4. It cannot distinguish incremental from cannibalized sales. Brand search and retargeting harvest demand other channels created. Markov happily hands them credit, exactly the cannibalization blind spot every path-based model shares.

5. It is hungry and fragile. With too few conversions the transition probabilities are mostly noise, and short attribution windows truncate the very journeys the model exists to analyze.

A €-Denominated Worked Example

For illustration: a Shopify skincare brand spends €40,000 a month and records 2,000 conversions, and its Markov model returns:

Branded search tops the table, so the team shifts another €6,000/month into it. Three months later, total revenue is flat.

Why? Branded search mostly harvests demand created upstream by Meta prospecting and email. Its high removal effect reflects that almost every path passes through it near the finish line: a correlational artifact, not incremental value.

A causal attribution analysis, validated with a branded-search holdout, tells a different story:

Same data, opposite budget decision. Markov ranked the harvester first; causal inference ranked the creator of demand first. (Figures are illustrative.)

Common Mistakes

• Treating removal-effect percentages as incremental lift. They are not the same quantity.
• Running Markov on too little data and reading three-decimal precision into noise.
• Skipping channel normalization, so one channel is split across three misspelled states.
• Ignoring un-clicked impressions, then wondering why upper-funnel channels look weak.
• Never validating against a holdout or geo test. Markov should generate hypotheses; experiments confirm them.
• Using a 7-day window on a brand with a 30-day consideration cycle.

Checklist: Before You Trust a Markov Result

• The channel ranking holds when you fit the model on two consecutive months
• Channel names normalized and UTMs audited
• Cross-device paths stitched, not double-counted
• Attribution window matches your real buying cycle
• View-through impressions accounted for, not silently dropped
• Top channels flagged for an incrementality test
• Result cross-checked against new-customer CAC
• You know which Trust Ladder rung you are on (see the ladder above)

Key Takeaways

• Markov chain attribution credits channels by removal effect: the drop in conversion probability when a channel is deleted from the path graph.
• It is a genuine upgrade over rules-based models and a sibling of Shapley and data-driven attribution.
• It remains correlational: it explains observed customer journeys, it does not prove incrementality.
• It is data-hungry, sensitive to tagging, blind to un-clicked demand, and prone to crowning demand-harvesters.
• Use Markov to rank and hypothesize; use causal inference and incrementality testing to decide. See also the unified measurement framework and MMM vs MTA vs causal inference.

Before you move budget on a Markov ranking, test the channel it credits most: cut its spend in one region for two weeks, keep everything else the same, and compare revenue with the regions where it kept running. If revenue barely moves, the removal effect was measuring presence, not cause, and that gap is what causal attribution exists to close.