Attribution models: what they are, and why we only deliver one

By the HonestTag team · Published July 12, 2026

Every attribution number you have ever seen was produced by a rule. Change the rule and the number changes, without a single dollar of spend or a single order changing underneath it. That is why the model matters more than the dashboard it lives in, and why a tool that lets you swap models is not measuring anything.

What an attribution model actually is

A customer's path to checkout usually has more than one touch: a prospecting ad, a search, a retargeting ad, an email. An attribution model is the rule deciding which of those touches gets credit for the order.

• First-click credits the click that first brought the customer to your store.
• Last-click credits the touch nearest checkout.
• Any-click (participation) counts every platform that touched the journey.
• Linear, position-based, and time-decay split credit across an ordered sequence of touches by some weighting scheme.
• View-through credits ad impressions the customer saw but never clicked.

None of these is physics. Each one is an answer to a different question, and each one produces a different number from the same orders.

Why HonestTag reports exactly one

Every attributed figure HonestTag reports (First-Click NCAC, the Gap's first-click column, every verdict) runs on one model: first-click new-customer. Store-level MER, NMER and NCAC use no attribution model at all. Delivery is separate: every verified purchase is sent, new and returning customers both. That is a deliberate fit, not a default we shipped by accident.

The acquisition question is "which click first brought this customer in?" First-click on verified new customers answers exactly that. Last-click answers a different question, "what was the customer touching on the way to checkout?", and it systematically over-credits retargeting and branded search, the touches that harvest buyers who were already coming back. Grade acquisition spend with last-click and your harvest campaigns look like heroes while the prospecting that first brought those customers in looks expendable.

One model, stated openly and applied consistently, is what makes numbers comparable across weeks and across platforms. The model choice, like every threshold and formula in the app, is published in numbers on the in-app methodology page.

The comparison table: see the difference without a switch

On the Truth tier and up, HonestTag shows a read-only model comparison table: first-click new customers, last-click new customers, and any-click participation, side by side, per platform. A delta column quantifies exactly what the model choice does to each platform's number: for example, how many new customers a platform "gains" the moment last-click is doing the grading.

The table's fixed header states plainly that first-click new-customer is the model behind First-Click NCAC, the Gap's first-click column and the verdicts; delivery sends every verified purchase, new and returning customers both. Next to it, a read-only last-paid-click table covers every order, labeled as reporting only and never sent to ad platforms. There is no model switch anywhere, on any surface. You get the comparison; you do not get a dial that changes what the app tells Google Ads or your team.

One honest footnote sits under the table instead of hidden in documentation: any-click columns overlap by design. An order counts once per platform that touched it, so the columns do not sum to total orders. That is what participation means, and we say so.

What we can compute honestly, and what we refuse to fake

The comparison has stated limits, because computing it honestly imposes them.

• Last-click and any-click are go-forward. They are computed from the day collection began and are never reconstructed backward. History is not restated to fill in a view that did not exist yet.
• Linear, position-based, and time-decay are absent. They would require storing every visitor's full ordered touch history. HonestTag deliberately does not collect that, because expanding data collection to feed model theater is backwards.
• View-through is impossible by design. It needs impression data, which only the ad platform has, the same platform whose performance is being graded. HonestTag never receives impression data, and treats that as a feature.

What the dishonest version looks like

Most attribution dashboards handle models one of three ways, and all three flatter someone.

First, the switchable model: a dropdown that lets whoever runs the dashboard shop for the flattering number, then present it as measurement. Second, the silent model change: the vendor adjusts the rule and history quietly restates itself, so last quarter's numbers no longer mean what they meant when you acted on them. Third, "multi-touch AI attribution": weighted credit whose weights nobody outside the vendor can audit, sold as sophistication.

In July 2026, we reviewed 80 tracking and attribution apps in the Shopify App Store. Zero published how their numbers are computed. When the model is hidden, switchable, or unauditable, the number is a negotiation, not a measurement.

What the comparison shows

1. It translates platform claims. A platform's last-click brag maps to what it means in acquisition terms: when it claims 200 conversions and the first-click new-customer column shows 60, the gap says what kind of work that budget is really doing.
2. The delta is itself a signal. Channels whose numbers depend most on model choice are the least stable claims. A channel that looks similar under every model is doing real acquisition; a channel that only shines under last-click is harvesting.
3. One model is the reported one. Verdicts and every attributed count run on first-click new-customer, and delivery sends every verified purchase, new and returning customers both; the comparison exists to explain disagreements between models, not to create new ones.

If any of the terms above are unfamiliar, the glossary defines each one in a sentence or two.

Frequently asked questions