Stop Choosing an Attribution Model. Compare Them All.

Every planning season, some team relitigates which attribution model to use, and the debate never closes, because every side is defending something defensible. The question has no correct answer. An attribution model is an opinion about credit: last click is an opinion, data driven a subtler opinion with a confidence interval. The mature move is to stop litigating and run the models side by side. The disagreements between them, rather than any single model's output, are where the information lives.

Every model is an opinion about credit

An attribution model is a rule for distributing credit for a conversion across the marketing touchpoints that preceded it. The definition ends there: nothing in it promises causation, because causation is unknowable from click paths alone: you cannot see from a path whether the buyer would have purchased anyway. Only controlled experiments approach that question, a point we make in Platform ROAS Is Fiction.

So every model encodes a behavioral assumption about how influence works, and every assumption systematically favors certain channels. Choosing a model is choosing whose work gets to look good this year. That is why the debate feels political. It is political.

The six models and who they flatter

Last click

Assumption: the final touch did the work. Flatters branded search, retargeting, and bottom-of-funnel email, the channels standing nearest the register when the sale rings up. Punishes everything that created the demand in the first place. Still the default lens in most platform reporting, which explains a decade of overfunded retargeting.

First click

Assumption: the introduction is everything. Flatters prospecting, top-of-funnel content, and influencer traffic that starts journeys. Ignores everything that nursed a hesitant buyer over the line. The mirror image of last click, and exactly as extreme.

Linear

Assumption: every touchpoint contributed equally, so a seven-touch journey pays out one seventh per touch. Diplomatic and rarely believed. Its practical effect is flattering channels that show up often in the middle of journeys, like email flows and paid social feeds.

Time decay

Assumption: recency equals influence. Credit grows toward the conversion, typically on a seven-day half life, so a touch yesterday earns double a touch eight days ago. A softer last click that still tilts toward closers.

Position based

Assumption: introductions and closes matter most. The common variant gives 40 percent to the first touch, 40 percent to the last, and splits the remaining 20 across the middle. Flatters both ends of the funnel and squeezes the consideration work in between.

Data driven

Assumption: let the math decide. Conversion and non-conversion paths are compared to estimate each touchpoint's contribution, which is what GA4 does by default. It is the most sophisticated of the click-based models and it deserves neither the reverence nor the suspicion it gets. It learns only from touches it can observe, and you cannot open the hood to see why a channel's credit moved this month. A learned opinion is still an opinion.

The delta is the insight

Run all six over the same conversions and each channel gets six credited revenue numbers.

A channel whose credit collapses as you move from last click toward first click is a closer. It harvests intent that already existed. A channel whose credit jumps in the same move is an opener: it creates demand that other channels later collect. And a channel whose credit barely moves across models is the safest line in your budget, because every defensible opinion about credit agrees on it.

That last case is the practical payoff. Budget decisions stop depending on winning the model argument when a channel looks good under every model. Scale those with confidence. Channels that only look good under one model are where you spend your skepticism, and your next incrementality test.

The same month under four models

One month, €63,000 of spend, €210,000 of attributed revenue. Each column redistributes credit for the same orders, so every column sums to the same total.

What the deltas say:

• Meta prospecting looks like a failure under last click (€38,000 against €40,000 spend) and like the engine of the business under first click. It opens journeys. Cutting it on last-click evidence would starve every channel downstream of it, slowly enough that nobody would connect the two.
• Meta retargeting swings the other way, from €55,000 at last click to €9,000 at first click. It is a closer, and a €12,000 line item earning most of its credit at the register deserves a spend-pause test before it earns another budget increase.
• Branded search is high under every model, but €52,000 of first-click credit for people typing your brand name should raise an eyebrow. Nobody's journey truly starts at a branded search. Something upstream and invisible put the name in their head.
• Email barely moves across models. Whatever you believe about attribution, €3,000 of spend returning roughly €35,000 is defensible under all of it.

What it takes to run models side by side

One table, in one warehouse, holding deduplicated touchpoint paths: every ad click, session, and email open for a given buyer, stitched to one identity and joined to the order of record. GA4's BigQuery export supplies the behavioral spine, ad platform exports supply spend and click identifiers, and your store supplies the orders, with revenue defined once along the way, as laid out in why Shopify, GA4, and Meta report different revenue.

Platform dashboards cannot do this at all. Each platform sees only its own touches, applies one model of its own choosing, and deduplicates only against itself. Model comparison is a warehouse exercise, which is exactly how an attribution layer on your own data should be built.

What no click model can see

Every model above shares one blind spot: journeys without clicks. Podcast mentions, a recommendation in a WhatsApp group, an influencer story with no tagged link, a rejected consent banner. These paths are missing from all six models equally, so even the deltas stay silent about them. The cheapest instrument that observes this territory is a one-question survey at checkout, which we cover in Post-Purchase Surveys Catch What Your Click Data Misses. Read survey shares next to model outputs and the invisible channels start showing up in the gaps.

Where Verity fits

Verity provides the foundation this work sits on: managed pipelines land touchpoints, spend, and orders in your own BigQuery, SQLMesh models build the deduplicated path table with tests and lineage, and the semantic layer keeps channel and revenue definitions consistent across every model you run. The attribution page shows how the side-by-side view comes together.

Frequently asked questions

Is data driven attribution not already the answer?

It is a good input and a poor referee. GA4's data driven model learns from the clicks it can see, cannot be audited channel by channel, and produces one opinion where you need the spread of several. Include it in the comparison. Do not let it end the comparison.

Which model should I show leadership?

None of them as "the number". Report revenue and MER from your order of record as truth, then present the model spread when proposing budget shifts: this channel holds up under every model, that one only under last click. Leaders make better calls on a range with a recommendation than on a false point estimate.

How much data do I need before the deltas mean anything?

Enough conversions per channel for the shares to stabilize, as a rule of thumb a few hundred conversions per channel per period, which for many brands means reading monthly rather than weekly. Below that, model deltas wobble for sample-size reasons and you will read noise as narrative.