Shapley and Markov Attribution Models, Explained for Marketers

Last-click attribution has one rule: whoever touched the customer last gets all the credit. It’s simple, and it’s wrong. A customer who saw three ads, read a comparison page, and then clicked a branded search result didn’t get bought by branded search alone, but that’s exactly what last-click tells your dashboard. Shapley and Markov are two honest ways to fix this. Not by guessing, by actually calculating who did the work.

Neither needs a statistics degree to understand. They need two questions.

Shapley: what did each channel actually contribute?

Shapley value comes from game theory, originally built to answer a fairness question that has nothing to do with marketing: if a group of people work together and earn a payout, how much did each person actually earn?

Apply that to a customer journey. Say a customer touches Meta, then email, then Google Search before buying. Shapley doesn’t just ask “what was the order?” It runs the maths on every possible ordering of those three channels and asks, for each one, how much extra value that channel added at that point in the sequence. Then it averages across all of them.

The result is a number that reflects genuine cooperation. If Meta and email tend to show up together and both matter, Shapley gives both real credit, not just whichever one happened to be closest to the sale.

The decision it informs: how to split budget fairly across channels that work together.

Markov: what happens if you remove a channel?

Markov chain attribution asks a different question, and it’s a more practical one for budget decisions. It models the customer journey as a path, a sequence of steps between channels, and then does something almost destructive: it removes one channel entirely and recalculates how many conversions the remaining channels could still produce.

The difference between “conversions with the channel” and “conversions without it” is that channel’s removal effect, its credit.

Think of it as pulling a stepping stone out of a path and checking whether people can still cross. Pull out a channel that was doing real work, and the conversion rate drops. Pull out one that was just along for the ride, and almost nothing changes.

The decision it informs: which channels are load-bearing, and which ones you could cut without losing much.

Shapley or Markov, when to use which

They’re not competing answers to the same question, they’re answers to two different questions.

Question you’re asking	Model	What it’s built for
”How should I split budget fairly across channels working together?”	Shapley	Fair credit for cooperation
”What breaks if I cut this channel?”	Markov	Finding load-bearing vs. replaceable spend
”Who technically touched the customer last?”	Last‑click	Good for tuning the closer, blind to the path that got them there

Most teams don’t need to pick one. They need both, asked at different moments. Shapley when planning next quarter’s budget split. Markov when deciding whether a channel earns its keep.

Why this matters more than the maths

Here’s the trust problem with most attribution tools: they’ll show you a number, and you’re expected to believe it. NodeNarrative runs seven attribution models, including Shapley and Markov, on the same underlying knowledge graph of a customer’s real touchpoints, and every one of them is inspectable. You can see how the credit was assigned, which touchpoints contributed, and why. That’s not a small detail. A number you can’t audit is a number you’re just trusting a vendor about, and vendors have an obvious incentive to make their own channel look good.

Transparency isn’t a feature bullet here. It’s the whole point of running a real model instead of a black box.

The honest caveat

Shapley and Markov are only as good as the journey data underneath them. A store with a handful of weekly orders and three tracked touchpoints will get a noisier answer than one with thousands of monthly conversions and a clean, deduplicated identity graph. These models don’t invent signal that isn’t there. They’re a better way of reading the signal you actually have, which is why the identity resolution underneath them matters as much as the model itself. That’s not something you need to bring yourself: NodeNarrative’s knowledge graph does the deduplication and identity resolution automatically from your first-party pixel data, so the clean journeys these models need are already there before you look at a single number.

They’re also not a crystal ball. Shapley and Markov explain what happened in the journeys you’ve already tracked. Neither one predicts a channel that hasn’t been tried yet.

See it on your own data

The real test of an attribution model isn’t reading about it, it’s watching it run against your actual channel mix. NodeNarrative processes attribution decisions in under 100ms at the 95th percentile, measured on our own infrastructure, so you’re not waiting on a report, you’re looking at your own journeys with Shapley and Markov (and five other models) applied live, on every plan, not gated to an enterprise tier.

See your true ROAS: every model, every plan.