How Multi-Touch Attribution Works

Every customer journey involves multiple interactions across multiple channels. A user might see a social ad, click a search result, open an email, and then convert through a retargeting display ad. Multi-Touch Attribution (MTA) is the methodology that determines how much credit each of those touchpoints deserves.

The Problem with Simple Attribution

Most analytics platforms default to last-click attribution — the final touchpoint before conversion gets 100% of the credit. This is convenient but deeply misleading. It systematically overvalues lower-funnel channels (paid search, retargeting) and undervalues upper-funnel channels (social, display, video) that initiated the journey.

Consider a typical customer journey:

1. Day 1: User sees a paid social ad and visits the site (awareness)
2. Day 3: User searches for the brand and clicks an organic result (consideration)
3. Day 5: User receives a promotional email and opens it (engagement)
4. Day 7: User searches for the product, clicks a paid search ad, and converts

Under last-click, paid search gets 100% of the credit. But was the conversion really caused by that search ad? Or was it the social ad that created initial awareness, the organic visit that built familiarity, and the email that reinforced intent? MTA tries to answer this more nuanced question.

Types of Attribution Models

Rule-Based Models

These models assign credit using predetermined rules. They're simple to implement but make arbitrary assumptions about how touchpoints contribute to conversion.

• Last-click: 100% credit to the final touchpoint. Overvalues bottom-funnel.
• First-click: 100% credit to the first touchpoint. Overvalues top-funnel.
• Linear: Equal credit to every touchpoint. Simple but assumes all interactions are equally valuable.
• Time-decay: More credit to touchpoints closer to conversion. Better than linear, but the decay function is arbitrary.
• Position-based (U-shaped): 40% to first, 40% to last, 20% split among the middle. A reasonable heuristic but still arbitrary.

The fundamental problem with rule-based models is that the rules are chosen by the analyst, not learned from data. They don't adapt to your specific business or customer behaviour.

Algorithmic (Data-Driven) Attribution

Algorithmic MTA models learn the credit allocation from your actual conversion data. Instead of predetermined rules, they use statistical or machine learning methods to estimate each touchpoint's contribution.

The main approaches:

Shapley Value Attribution

Borrowed from cooperative game theory, Shapley values calculate each touchpoint's marginal contribution by considering every possible combination of touchpoints. For each channel, the model asks: "On average, how much does adding this touchpoint to any combination of other touchpoints improve the conversion probability?"

This is the approach used by Google's Data-Driven Attribution (DDA). It's mathematically principled and fair — it satisfies several desirable properties like efficiency (all credit sums to 100%) and symmetry (identical channels get identical credit).

Markov Chain Models

Markov chain attribution models the customer journey as a series of transitions between states (touchpoints). It estimates the probability of moving from one touchpoint to another, and calculates each channel's "removal effect" — how much the overall conversion rate drops if you remove that channel entirely.

The removal effect provides an intuitive measure of importance: if removing paid social drops the conversion rate by 25%, while removing display only drops it by 5%, paid social is clearly more valuable to the journey.

Logistic Regression / ML Models

Some implementations use logistic regression or gradient-boosted models where the features are binary indicators of which touchpoints a user was exposed to, and the target is conversion. The model coefficients or feature importances then serve as the basis for credit allocation.

How We Implement MTA

Data Requirements

MTA requires user-level journey data. At minimum, you need:

• User identifier — A consistent ID across touchpoints (first-party cookies, logged-in user IDs, or device graphs)
• Touchpoint log — Timestamp, channel, campaign, and interaction type (impression, click) for each marketing exposure
• Conversion events — Timestamp and value of conversions, linked to user IDs
• Lookback window — How far back to look for touchpoints (typically 30-90 days)

The Privacy Challenge

MTA has always depended on user-level tracking, and the privacy landscape is changing fast. With the deprecation of third-party cookies, iOS App Tracking Transparency, and regulations like GDPR and CCPA, the data available for MTA is shrinking.

Modern MTA implementations need to account for this:

• First-party data focus — Build on your own customer data (CRM, logged-in users, first-party cookies) rather than third-party cookies
• Probabilistic matching — Use statistical models to stitch fragmented journeys without deterministic user IDs
• Aggregated measurement — Platforms like Meta and Google are moving toward aggregated reporting APIs (SKAN, Privacy Sandbox) that provide conversion data without user-level exposure
• Hybrid approaches — Combine MTA with MMM and lift tests to fill the gaps where user-level data isn't available

What MTA Can and Cannot Tell You

MTA Strengths

• Granularity — Credit at the campaign, ad group, and creative level. Useful for tactical optimisation.
• Speed — Near real-time insights. You can see attribution data within days of campaign launch.
• Journey understanding — Reveals how channels work together. Which channels assist and which convert.
• Creative insights — Identify which messages and formats are most effective at each stage of the journey.

MTA Limitations

• Correlation, not causation — MTA observes associations between touchpoints and conversions, but can't prove that a touchpoint caused the conversion. Lift tests are needed for causal evidence.
• Digital only — Traditional MTA can't measure offline channels (TV, radio, print, OOH). You need MMM for the full picture.
• Selection bias — Users who are shown more ads are often those already more likely to convert. MTA can mistake correlation (high-intent users see more ads) for causation (ads drive conversions).
• Privacy erosion — Declining signal from cookie deprecation and privacy regulations. MTA is becoming harder to implement with the same accuracy as before.

MTA Within a Unified Measurement Framework

MTA is most powerful when it's not used in isolation. The best measurement programmes combine all three methodologies:

• MTA for tactical, granular optimisation of digital campaigns
• MMM for strategic budget allocation across all channels (including offline)
• Lift tests for causal validation of the most important channels

Each method compensates for the others' weaknesses. MMM captures offline channels that MTA misses. Lift tests provide the causal ground truth that both MTA and MMM lack. And MTA provides the granularity and speed that MMM can't match.

The question isn't "MTA or MMM?" — it's "how do we use both together to make better decisions?"

Getting Started with MTA

If you're currently relying on last-click attribution, even a simple move to data-driven attribution (available natively in Google Ads and GA4) is a meaningful improvement. But for a true understanding of your customer journey, you'll want a custom MTA implementation built on your first-party data.

The key is to start with clean, consistent data collection. Invest in your tracking infrastructure, ensure consistent user identification, and define clear conversion events. The model is only as good as the data it's built on.