Which Customer Journey Attribution Model Should You Use?

First touch, last touch, linear, time decay and data-driven are five ways to divide the same credit among the same observed touchpoints. Run all five over one export and you get five different channel rankings from identical customer behaviour. The choice between them is a policy about how to share credit. It is not a finding about what worked, and picking a better rule does not convert an allocation into evidence.

The five, and who each one flatters

First touch gives everything to the earliest observed touchpoint. It flatters discovery: paid social, display, anything upstream. Useful if your explicit question is which channel introduces people to the brand, and misleading for anything else.

Last touch gives everything to the final touchpoint. It flatters whatever sits nearest the purchase, which is usually brand search, retargeting and email. It is the default in more reporting than anyone admits, and it is where the largest overstatements tend to live. Last-click attribution covers the failure mode in detail.

Linear splits credit evenly. It is the least opinionated and therefore the least wrong on average, which is not the same as being right about any particular channel.

Time decay weights recent touchpoints more heavily. It is last touch with the edges sanded off, and it flatters the same channels by a smaller margin.

Data-driven fits weights from your own observed journeys rather than imposing them. This is genuinely more sophisticated and it is still an allocation over the visible set. Why GA4's data-driven attribution is a black box covers what you can and cannot inspect.

The test that separates a policy from a measurement

Ask what the model would report if a channel had no effect at all. A causal design has an answer: an estimate near zero, with an interval that includes zero, and the design will say so. An allocation rule has no answer, because a channel that appears in observed journeys receives credit under every one of the five rules whether or not it caused anything. Retargeting is the clearest case. It appears late in journeys by construction, so it collects credit under last touch and time decay, and it does that whether the customer was going to buy anyway or not.

That is the whole reason the rule choice cannot be settled by picking harder.

How to choose anyway

You still need a number this quarter, so choose on the decision rather than on the theory.

For a question about discovery, first touch is the least distorting, with the caveat that it ignores everything after.

For operational pacing inside a channel, where you are comparing this week's campaigns against last week's under the same rule, any consistent rule works, because the rule cancels out of the comparison.

For budget reallocation between channels, none of them is adequate, and this is the decision people most often make with them. Moving money between channels on shared credit means moving money on the rule.

What to run beside it

Two numbers reframe any model's output. Coverage, attributed conversions over the orders your store shipped, tells you what share of the business the percentages describe. The claim ratio, summed platform claims over orders, tells you by how much the platforms collectively overstate; when it exceeds one, the channel table is arguing with itself. The one-hour claim ratio audit computes both from data you already hold.

Then, for the reallocation decision specifically, use a design built for it. A causal read estimates what would have happened without each channel and reports an interval rather than a point. How to diagnose over-attribution in Meta and Google Ads shows what the gap between the two usually looks like.

What to do this week

• If you own the budget: run your last quarter under two models instead of one. The spread between them is the size of the policy decision you have been treating as a measurement.
• If you have to defend the number: name the model on every slide. A ranking without its rule is not reviewable.

The interactive demo shows the causal comparison on a sample store, no signup.

Product facts as stated on causalityengine.ai on 9 September 2026. Method and measurability material is from The Price of Being Found (Edition 2.10), Chapters 12 and 15, with the book's caveats.