Multi-Touch Attribution: Why Last Click Is Lying to You

Multi-Touch Attribution: Why Last Click Is Lying to You

By Shashank Paliwal | Data Science Manager | Marketing Analytics

Last Touch attribution gave Loyalty Coupons 100% of the credit. Shapley values told a very different story. Here’s what fair attribution actually looks like — and why it changes your budget by 15–25%.

Here’s a scenario I’ve seen play out dozens of times.

A retailer’s marketing team looks at their Last Touch attribution report and concludes: Loyalty Coupons are the best performing channel — 38% of revenue attributed. TV looks weak at 4%. The obvious decision: cut TV, invest more in coupons.

Six months later, coupon redemptions are up but total sales are flat. Worse — new customer acquisition has fallen sharply. What happened?

Last Touch attribution happened.

The problem with Last Click

Last Touch attribution gives 100% of the conversion credit to the final touchpoint before purchase. It’s simple, it’s transparent, and it’s almost always wrong.

A typical grocery shopper’s journey might look like this:

• Monday: sees a TV ad during a cooking show
• Wednesday: notices an OOH billboard near their commute
• Thursday: clicks a digital display ad and browses the website
• Friday: receives a loyalty coupon in their quarterly statement
• Saturday: converts in-store using the coupon

Last Touch attributes 100% of this sale to the Loyalty Coupon. But remove the TV ad, the billboard, and the digital display — and that customer probably never reaches the coupon stage at all. The coupon closed the deal; it didn’t create the intent.

— -

What Multi-Touch Attribution does instead

MTA distributes credit across all touchpoints in the customer journey. The question is how to distribute it fairly.

There are two approaches: rule-based and data-driven.

Rule-based models apply a fixed formula:

• Linear— equal credit to every touchpoint
• Time Decay— more credit to recent touchpoints, less to early ones
• Position-based— 40% to first, 40% to last, 20% split across the middle

These are better than Last Touch but still arbitrary. They don’t use your actual data to determine what each channel is worth.

Data-driven models learn from your conversion data:

• Shapley Values— from cooperative game theory; calculates each channel’s average marginal contribution across all possible channel combinations
• Markov Chains — models the journey as a sequence of transitions; a channel’s credit = how much conversion probability drops if you remove it

— -

Shapley values: the maths behind fair attribution

The Shapley value comes from game theory. Imagine your 8 marketing channels as players in a cooperative game, and revenue as the prize they jointly win. How do you divide the prize fairly?

The Shapley answer: each channel’s credit equals its average marginal contribution — how much it adds to every possible coalition of channels it could be part of.

For channel TV in a world of {TV, Digital, Email, Loyalty}:

• What does TV add to {Digital}?
• What does TV add to {Email}?
• What does TV add to {Loyalty}?
• What does TV add to {Digital, Email}?
• …and so on for all 15 possible subsets

Average those contributions (weighted by how likely each coalition is), and you have TV’s Shapley value.

This is computationally expensive — for 8 channels you have ²⁸ = 256 coalitions to evaluate. But it’s the most theoretically defensible attribution you can do.

— -

Markov chains: the removal effect

Markov Chain attribution takes a different angle. It models the customer journey as a probabilistic path through channel states.

First, build a transition matrix from your data — the probability of moving from channel A to channel B, or from channel A to conversion, in your actual journeys.

Then, for each channel, ask: what happens to overall conversion probability if I remove this channel from the model and reroute its traffic proportionally to other channels?

That drop in conversion probability is the channel’s removal effect — its attribution weight.

Channels that sit on critical conversion paths have high removal effects. Channels that are easily bypassed have low ones.

— -

What the data actually shows

When I’ve run both models on real retail journey data, the pattern is consistent:

Awareness channels (TV, OOH, Radio) are systematically undervalued by Last Touch and correctly credited by Shapley and Markov. A TV campaign that drives a 30% increase in new visitor traffic shows up as 4% in Last Touch and 18% in Shapley.

Conversion channels (Loyalty Coupons, Email) are systematically overvalued by Last Touch. They close deals but don’t create demand.

The budget implication is large. Moving from Last Touch to Shapley typically shifts 15–25% of budget between channels. That’s not a rounding error — it’s tens of millions of pounds for a large retailer.

— -

MTA vs MMM: they’re not competing

One question I get often: if you have MMM, do you need MTA?

Yes — they answer different questions.

MMM works at the aggregate, weekly level. It’s essential for channels where individual tracking is impossible — TV, OOH, Radio. It controls for external factors (weather, macroeconomics, competitor activity) and gives you clean incremental attribution at a channel level.

MTA works at the individual journey level. It’s powerful for digital channels where you can track every click and impression. It reveals how channels interact in sequence — what a Shapley analysis of 50,000 customer journeys shows that aggregate MMM cannot.

In production, the right answer is to run both and triangulate. Where they agree, you’re confident. Where they disagree, there’s something to investigate.

Remember MMM is important for long term planning/optimization for your brand and MTA helps you for short time/real time optimization of your campaigns.

— -

The code

I’ve built a complete MTA implementation in Python — customer journey simulation across 8 channels (4 offsite, 4 onsite), all four rule-based models, Shapley value computation, and Markov Chain attribution with removal effects.

The final section shows the budget reallocation chart — the number that actually changes decisions.

GitHub: [link to notebook]

The notebook is designed to be readable as well as runnable — each section has markdown explaining what and why, not just how.

— -

What’s next

This is the second in a series of four posts on marketing analytics:

1. [How Market Mix Modeling Actually Works]
2. Multi-Touch Attribution: Why Last Click Is Lying to You (this post)
3. Price Elasticity Modelling for Margin Optimisation (coming soon)
4. Customer Segmentation That Drives Campaign ROI (coming soon)

Note:- I will be covering the challenges we face while we apply for multi-touch attribution for offsite channels like Meta where data is wall-gardened and how to deal with them as well How can we integrate MMM & MTA outputs.

— -

*Shashank Paliwal is a Data Science Manager at Accenture specialising in Marketing Effectiveness, MMM, and Pricing Strategy. Connect on LinkedIn