Attribution Models Explained: First Click, Last Click and the Rest

An attribution model is a rule for sharing the credit for an order among the visits that led up to it. First click gives all the credit to the visit that introduced the shopper, last click gives it all to the visit that closed the sale, and linear, time-decay and position-based models spread it across the path in different proportions. None of them is the true answer; each is a different way of looking at the same journeys. This guide describes the five common models, shows where each one misleads, and explains why reading several side by side tells you more than choosing one.

What does each attribution model do?

Each takes the same list of visits before an order and divides the order's value differently.

• First click: all credit to the first recorded visit.
• Last click: all credit to the final visit before the purchase.
• Linear: equal credit to every visit on the path.
• Time-decay: more credit the closer a visit is to the purchase, less the further back it is.
• Position-based: most of the credit to the first and last visits, with the remainder shared among those in between.

The exact weights in time-decay and position-based models are settings chosen by whoever built the tool, so two products can use the same name and produce different numbers.

How do the models differ on a single order?

Take one example path. Say a shopper first arrives from a creator's link, comes back a week later through a paid search ad, and finally clicks an email and places a $90 order.

• First click: creator $90, search $0, email $0.
• Last click: creator $0, search $0, email $90.
• Linear: $30 each.
• Time-decay: email gets the largest share, search less, the creator the least.
• Position-based, using an example weighting of 40% to the first visit, 40% to the last and 20% to the middle: creator $36, search $18, email $36.

The order is the same $90 every time. What changes is the story about which channel earned it, and that story is what budgets get decided on.

When does each model mislead?

Every model has a blind spot, and it follows directly from its rule.

First click flatters whatever introduces people and ignores the work of bringing them back.

Last click flatters whatever stands nearest the checkout. Email, searches for your own brand name and discount-code sites can catch shoppers who had already decided, and then they collect credit for demand that something else created.

Linear treats a casual glance and a decisive visit as equals. A shopper who opens five emails on the way to an order makes email look five times as important as the ad that started it.

Time-decay is last click with softer edges. It suits short promotions, and undervalues slow discovery for products people think about for a while.

Position-based bakes in an opinion: that the introduction and the close matter most. The weights are arbitrary, and the middle of the path, where comparison and reassurance happen, is marked down by design.

Why look at several models side by side?

Because the disagreement between them is the information. Put the same channels in a table with one column per model and read across.

• A channel that scores well on first click and poorly on last click is an introducer. It starts journeys that other channels finish.
• A channel that scores poorly on first click and well on last click is a closer. It converts interest that already existed.
• A channel whose share barely moves between models is doing a bit of everything, and conclusions about it are safer.

This changes decisions. Cutting an introducer because its last-click numbers are weak can shrink the pool your closers draw from a few weeks later. Scaling a closer because its last-click numbers are strong may buy very little, since it was catching people who were coming anyway.

What are the limits of first-party visit data?

Attribution built from visits to your own store is honest about its source, and it has edges you should know.

• Only consented visits are recorded. Shoppers who decline analytics are missing, so paths are a sample. See how cookie consent affects Shopify analytics.
• Devices are not joined up. A shopper who discovers you on a phone and buys on a laptop can look like two people, with the first visit lost.
• Views are invisible. Someone who watched an ad without clicking never produced a visit, so the ad gets no credit in any click-based model.
• Credit is not cause. A model divides the value of orders that happened. It cannot say whether an order would have happened without a given channel.

That last point is the important one. For questions of cause, such as whether a channel adds sales at all, the tools are holdout tests and media mix modelling, described in our piece on attribution after cookies.

Which model should you use day to day?

Pick one as the house default for routine reporting and keep it fixed, so that this month can be compared with last. Then use the others to answer specific questions: first click when you are asking what finds new customers, last click when you are asking what closes.

Keep a blended measure next to any model as a sanity check. Total sales against total marketing spend does not depend on anyone's weighting, and if modelled channel numbers improve while the blended figure worsens, believe the blended figure. Our guide to ROAS, CAC and MER explains how those fit together.

Marka Metrics offers five attribution models and channel paths built from visits to your own store, with attribution models included on its Growth and Pro plans.

A short checklist

1. Learn what each model's rule rewards before reading its numbers.
2. Compare channels across several models, not one.
3. Label each channel as an introducer, a closer or a mix.
4. Choose one default model for reporting and do not change it casually.
5. Check modelled results against a blended measure of spend and sales.
6. Treat attribution as a description of paths, not proof of cause.

Attribution is one of the questions Shopify's standard reports only partly answer; our guide to Shopify analytics beyond the built-in reports covers the rest. If you would like to compare models on your own store's visits, Marka Metrics is where we would start.