Multi-Touch Attribution for Ecommerce: How the Models Work and When to Stop Trusting Them

Multi-touch attribution splits the credit for a sale across every touchpoint that led to it, instead of handing all of it to one. That is the definition. The part nobody mentions is that Google deleted four of the six models you are about to read about. First click, linear, time decay and position based stopped being available in Google Analytics 4 in November 2023, and most guides on this topic still present them as live options.

So the question is not what the models are called. It is which ones you can still run, what each one does to your reported numbers, and how far you should trust the answer before you move budget. This article walks the same $200 order through six models, shows the credit landing in six different places, and covers what breaks once real customers stop behaving like a clean click path.

What is multi-touch attribution?

Multi-touch attribution assigns fractional credit for one conversion to several touchpoints in the customer journey. A shopper sees a paid social ad, searches your brand a week later, opens an email, then buys. Multi-touch attribution says all four mattered and divides the sale between them.

Single-touch attribution does the opposite. It picks one touchpoint and gives it everything. Nielsen describes the rules-based versions bluntly: by arbitrarily applying rules, these methods fail to measure the contribution of every touchpoint in the consumer journey accurately, causing marketers to make decisions based on skewed data.

That is the case for multi-touch in one sentence. Spreading credit across the path gets you closer to what actually happened than crowning a single winner.

Two things decide how a multi-touch model behaves, and they are easy to confuse:

• Which touchpoints are eligible. The lookback window sets how far back before a conversion a touchpoint can sit and still qualify. It changes which touchpoints enter the path. It does not change how credit is split among them.
• How credit gets divided among the touchpoints that qualify. That is the model.

Change the window and the path gets longer or shorter. Change the model and the same path pays out differently. Keep them separate in your head, because when a number moves you need to know which one you touched.

One order, six models, six different answers

Here is the whole problem in one table. A customer buys once, for $200. Five touchpoints, in this order: a paid social ad on day 1, an organic search visit on day 3, an email on day 6, a brand search click on day 9, a direct visit on day 10 when they buy. Paid social spend against this customer was $50.

Nothing about the customer changes across the rows. Only the model changes.

Every row sums to $200, because the order was $200. One sale, one revenue figure, six defensible ways to split it.

Look at the ROAS column. Same customer, same $50 of spend, same purchase. Paid social reports anywhere from 0.0x to 4.0x depending only on which model you picked. If your rule is that you cut a channel under 1x and scale it over 2x, then three of these six models cut paid social, one scales it, and two leave it alone. The customer did the same thing in every case.

This is why "what is our ROAS" is an incomplete question. The answer is a function of a setting.

The six multi-touch attribution models, and which ones still exist

Last click

The last touchpoint before the order takes 100%. It is the default in most reporting, including Shopify's marketing reports, where last click is selected by default when the metric and dimension conditions are met.

Last click is not a multi-touch model. It is worth naming because it is the number most operators are actually looking at, and it systematically overpays whatever sits closest to checkout. Brand search, direct traffic and email look excellent under last click, because they are near the purchase, not because they created the demand.

First click

The first touchpoint takes 100%. It flips the bias: now the channels that introduce people get everything and the channels that close get nothing.

Removed from Google Analytics 4 in November 2023.

Linear

Every touchpoint gets an equal share. Five touchpoints, 20% each.

Linear is honest about its own ignorance, which is its appeal. It also flattens real differences. A scroll-past impression and a hand-typed brand search count the same, and they are not the same.

Removed from Google Analytics 4 in November 2023.

Time decay

Credit rises the closer a touchpoint sits to the order. In the table above, day 10 earns five times what day 1 earns.

Time decay suits short consideration cycles, and it punishes upper funnel work by design. If you run awareness campaigns and judge them on a time decay model, you will conclude they do not work.

Removed from Google Analytics 4 in November 2023.

U-shaped, also called position based

40% to the first touchpoint, 40% to the last, and the remaining 20% divided across everything in the middle. It encodes a specific belief: introduction and conversion are the hard parts, the middle is maintenance.

That belief is often reasonable for ecommerce, which is why U-shaped is a sensible starting point when you have no strong prior. Whether it is true for your brand is testable, and mostly untested.

Removed from Google Analytics 4 in November 2023.

Full impact, the algorithmic models

Instead of a fixed rule, an algorithmic model learns credit from your own data. The common approach borrows Shapley values from cooperative game theory: it asks what each touchpoint added across all the paths where it appeared, and pays it proportionally to that marginal contribution.

The consequence matters more than the maths. A rules-based model gives the same answer for a given path forever. An algorithmic model gives an answer that depends on your whole dataset, so the same path can be valued differently as your mix changes. The numbers in the table for full impact are illustrative for that reason: on your data they would come out somewhere else.

Google Analytics 4 kept its version of this, data-driven attribution, and made it the default. Nielsen's term for the family is fractional attribution, which employs machine learning to calculate and assign fractional credit to the influential marketing touchpoints.

Also worth knowing: the useful variants restrict the field rather than change the split. A paid-only setting makes only paid touchpoints eligible for credit, which answers "how should I divide my ad budget" rather than "what drove this sale". Both are valid questions. They are different questions.

Why the model you pick changes your budget

Attribution feels like a reporting choice. It behaves like a budgeting decision, because the reported number is what people act on.

Run the table forward. Under time decay, paid social returns 0.3x and gets cut. Under first click it returns 4.0x and gets scaled. Cut it, and the day-1 touchpoint disappears from future paths, so brand search and direct start converting less, and the channels that looked strongest under last click quietly weaken. The model did not just describe the funnel. It reorganised it.

There is one number that does not move, and it is the most useful thing on this page. Total marketing efficiency, revenue divided by total spend, is identical under all six models. Attribution reshuffles credit between channels. It cannot create or destroy revenue, and it cannot change what you spent. So blended efficiency stays put at the total level and only moves within a channel.

That gives you a stable reference. Blended efficiency tells you whether the whole machine is working. Attribution tells you where inside the machine to look. Treating attribution as the scoreboard, rather than the map, is how brands talk themselves into cutting the channel that was feeding everything else.

What breaks multi-touch attribution for DTC brands

Every model above assumes you can observe the path. That assumption has been degrading for years.

Signal loss

App tracking permission changes and browser cookie restrictions removed a large share of the identifiers that used to stitch a journey together. What survives is a partial path. A model applied to a partial path returns a confident number about incomplete data, and looks exactly as confident as it would on complete data.

Double counting across platforms

Ask each platform what it drove and the answers sum to more than you sold. Paid social claims the sale, paid search claims the same sale, email claims it too. Each is telling the truth inside its own walls and none of them can see the others.

The check is arithmetic. Your attributed revenue should reconcile to your actual order revenue. If your channel reports add up to more than your store took, you are looking at claims, not a split.

View-through inflation

Some platforms count an impression as a contributing touchpoint. A shopper who scrolled past an ad without stopping enters the path as an influence. Meta's attribution setting documents which click and view windows apply, and view-through credit is the single biggest reason platform-reported conversions run ahead of anything reconciled to orders.

Include views if you believe an unclicked impression moved someone. Just do not compare a view-inclusive number against a click-only number and read the gap as performance.

Channels no pixel can see

Podcast reads, connected TV, direct mail, word of mouth and practitioner recommendation leave no click. For brands where demand comes largely from these, a click-based model will report that most sales came from nowhere, and route budget accordingly. Post-purchase survey responses and ingested view-through touchpoints can enter the path as real touchpoints, which is the only way these channels appear at all.

Windows that are too short

A 10-day window on a considered purchase with a 6-week research cycle will drop the touchpoint that started it. The path looks short because the window was short.

Multi-touch attribution vs marketing mix modeling vs incrementality

These three get discussed as competitors. They answer different questions and the mature setup runs more than one.

Attribution tells you who was there. Only an experiment tells you what was actually caused. If you have ever suspected that brand search is being credited for demand it did not create, no attribution model can settle it, because brand search genuinely was the last click. A geographic holdout can settle it, by turning the channel off in matched regions and measuring what happened to sales.

The practical order: use attribution daily, use marketing mix modeling for ecommerce for long-range planning if your spend justifies the history it needs, and use incrementality tests on the two or three decisions that actually matter each quarter.

How to set up multi-touch attribution on Shopify

Tie every touchpoint to an order. Your source of truth is your order table. If attributed revenue does not reconcile to orders, you have a tracking problem, and no model choice fixes it. Attributed conversions should tie back to actual orders before you interpret anything.

Pick one reference model and leave it alone. Choose a model, write down why, and report on it consistently. U-shaped is a reasonable default for ecommerce. An algorithmic model is better if you have the path volume to support it. What matters more than the choice is that you stop changing it, because a model change and a performance change look identical in a chart.

Keep platform numbers in a separate column. Do not reconcile them and do not average them. Show the platform-claimed figure next to your deduplicated figure. The gap between them is information: it tells you how much each platform is overclaiming, and it is stable enough to learn.

Set the window from your real purchase cycle. Look at time from first touch to order in your own data, take the point that covers most orders, and use that. Do not inherit a default.

Feed clean events back to the ad platforms. The platforms optimise on the conversions you send them. Send deduplicated, server-side events and their targeting improves. Send inflated ones and you are paying them to chase noise.

Then test the thing you most believe. Whatever your attribution says is your best channel, that is the one worth a holdout test. It is also the one you would least like to be wrong about, which is the point.

For a fuller comparison of how each model behaves on Shopify data, see our guide to attribution models for Shopify brands.

Reading attribution numbers without being fooled by them

Attribution is a lens, not a measurement. Used well it is the fastest way to find where to look. Used as truth it will move real money on the strength of a setting.

Three habits keep it honest.

Read every channel number as conditional. Not "email drove $40,000" but "email is credited with $40,000 under a U-shaped model on a 30-day window". Anyone who cannot state the model and the window is not reporting a result.

Watch blended efficiency for the verdict and attribution for the diagnosis. Blended tells you if the machine works. Attribution tells you which part to inspect.

Test before you commit. When a model change would move six figures, an experiment costs less than being wrong. A comparison of two attribution approaches on one brand's own data showed how differently the same spend can be credited, and what first-touch attribution did to one brand's reported numbers is a useful reminder that the model is doing more work than the channel.

Multi-touch attribution is worth running. It is a better description of reality than crowning a single click. Just hold it as the best available description, and keep one experiment and one blended number nearby to check it against.

FAQ