E-Commerce Attribution Models: What They Actually Measure (and What They Don't)

Pull up Meta, Google Ads and GA4 on the same afternoon and ask each one which channel drove last month's revenue. You will get three different answers, and none of them is lying — they are each counting a different thing. Attribution modeling is the set of rules that decides which touchpoint gets credit for a sale, and the model you pick quietly decides which channels look profitable and which get cut. This guide covers the models actually in use in 2026, what changed when Google gutted four of them, and why the harder problem for e-commerce is not which model to pick but what "credit" should even mean once returns are in the picture.

What Attribution Modeling Actually Decides

A customer sees a paid social ad, ignores it, finds you two weeks later through an organic search, signs up for the newsletter, clicks a retargeting ad a few days after that, and finally buys after typing your brand name into Google. That is one order and five touchpoints. Attribution is the rule that decides how much of that order's value each touchpoint is allowed to claim — and every reporting tool downstream, from Meta's dashboard to your own revenue report, is built on top of that decision.

There is no ground truth here. The customer's actual reasoning for buying is not observable, so every model is an assumption dressed up as a measurement. The question worth asking about any model is not "is this correct" but "what decision does this push me toward, and is that the decision I want to be making".

Single-Touch Models: Simple, and Wrong in a Predictable Direction

• Last-click. 100% of the credit goes to whatever touchpoint immediately preceded the order. It is the default almost every analytics tool ships with, because it needs no modeling and no lookback logic beyond "what was the last thing before checkout". It systematically overvalues branded search, retargeting and email — the channels that catch a purchase decision already made — and undervalues everything that built the intent in the first place.
• First-click. The mirror image: full credit to whatever introduced the customer, nothing to what closed them. Useful for judging pure awareness spend, useless for budgeting a performance channel, because it will tell you your prospecting campaigns are printing money while your retargeting looks like a waste — even when the retargeting is what actually converts the traffic the prospecting brought in.

Single-touch models still make sense in one situation: a genuinely short, low-consideration purchase where most customers convert within one or two sessions. Outside that, they are not simplifications of the truth — they are a specific, biased answer that happens to be cheap to produce.

Rule-Based Multi-Touch Models

These spread credit across the whole recorded path instead of picking one touchpoint. They are auditable — anyone can recompute the split by hand — and they need no minimum volume, which is exactly why they matter for stores that will never hit the conversion thresholds algorithmic models require.

• Linear. Every touchpoint gets an equal share. Five interactions before the sale means each gets 20%. It is the honest admission that you don't know which touchpoint mattered more — which is also its weakness: it treats an accidental ad impression the same as the email that made someone click "buy".
• Time-decay. Touchpoints closer to the sale get more credit, older ones less, on a defined half-life. This fits short e-commerce cycles reasonably well, but it can quietly overvalue retargeting simply for showing up last in the sequence — the same distortion as last-click, just softened.
• Position-based (U-shaped). 40% to the first touch, 40% to the last, the remaining 20% split across whatever happened in between. It is popular because it acknowledges both jobs — introducing the customer and closing the sale — without pretending the middle doesn't exist. The 40/40/20 split is still arbitrary; it rewards the bookends of the journey whether or not they were the touchpoints that actually mattered.

The shared flaw across all three: the weighting is fixed before you look at a single conversion path. A U-shaped model gives 40% to the first touch whether that touch was a considered demo request or someone's thumb slipping on an ad.

Data-Driven Attribution, and Why GA4 Only Offers Two Models Now

Data-driven attribution (DDA) does not fix a weighting in advance. It trains on your own converting and non-converting paths and assigns credit based on which touchpoints statistically correlate with a sale actually happening. In November 2023, Google removed first-click, linear, time-decay and position-based attribution from GA4 entirely — not deprecated behind a flag, removed. Two models remain: data-driven, and last-click. Google Ads imports from GA4 default to data-driven as well, falling back to last-click only when there isn't enough data.

That "enough data" threshold is where most small and mid-sized stores quietly lose the model they think they're using. DDA requires at least 400 conversions for the specific event and 20,000 total conversions across all events within the lookback window. Fall short, and GA4 does not warn you — it silently reverts the property to last-click while the interface still shows "Data-driven" as the reporting model. Many stores are making budget decisions on last-click attribution while genuinely believing they've graduated past it.

Even when DDA does run, it is a black box by design. You cannot recompute why a touchpoint got 23% instead of 30%, which makes it hard to defend a budget decision to someone who wants the reasoning, not just the number.

Why the Platforms Never Agree With Each Other

Switching models inside one tool is only half the problem. The bigger distortion is that Meta, Google Ads and GA4 are not describing the same event in the first place.

• Different definitions of a touchpoint. Meta counts likes, video views and image expansions as engagement and includes view-through conversions — someone who saw the ad but never clicked. GA4 only credits clicks that produced a tracked session, the most conservative count of the three.
• Every platform is self-attributing. Meta's algorithm has every incentive to credit Meta; Google Ads has every incentive to credit Google. Neither is being dishonest — each is measuring inside its own walled garden and has no visibility into what happened on the others.
• Dark social and stripped parameters inflate "direct". A link shared in a DM, a Slack message, or a mobile app handoff usually arrives with no UTM parameters attached. Analytics tools have nothing to attribute it to, so it lands in "direct" — which is not a channel, it is the label for traffic your tracking failed to explain.
• Cross-device journeys break the identity chain. Someone who sees an ad on their phone and buys on a laptop looks like two different people to a tool that has no way to link the sessions. The first touch simply vanishes from the recorded path.

Add these together and a model comparison that assumes clean, complete paths is comparing three incomplete pictures, not one true picture from three angles. Better identity resolution — the kind server-side collection and hashed customer IDs provide, covered in our guide to server-side tracking for e-commerce — narrows this gap, but it does not close it. No amount of modeling recovers a touchpoint nobody recorded.

Attribution Windows Are a Separate Decision From the Model

The model decides how credit is split between touchpoints; the window decides how far back you are willing to look for touchpoints at all. A 7-day click window and a 30-day click window applied to the same position-based model will produce different reports from the same raw data, because the 30-day version simply sees more of the journey.

Shorter windows suit impulse categories and understate the influence of upper-funnel content. Longer windows suit considered purchases and start rewarding touchpoints that happened to occur near a conversion by coincidence rather than causation. There is a second-order trap here too: if you pre-filter which sales count for a channel by your own attribution window — an affiliate program is the clearest example — you have made yourself the sole decision-maker on what that partner gets paid for, and your window needs to mirror the partner's program terms exactly, or you are quietly rewriting the agreement every time you report.

The E-Commerce Problem No Attribution Model Solves

Every model above, however credit is split, is splitting the same number: gross order value at the moment of checkout. None of them ask whether that order stayed sold.

A campaign that drives orders with a 55% return rate can look identical, on any attribution model, to a campaign that drives orders almost nobody sends back — right up until someone checks the bank account. Fashion and apparel campaigns are the clearest case: paid social ads with generous try-before-you-decide framing routinely post the best last-click ROAS in the account and the worst net contribution after returns, because the model measuring them was never designed to see a return happen three weeks later. Attribution answers "which touchpoint gets credit"; it was never built to answer "was this order worth acquiring", and treating the first question as a proxy for the second is where attribution-driven budget decisions quietly go wrong.

The fix is not a better attribution model. It is feeding the model a better number: contribution margin after returns and cost of goods, rather than gross revenue at checkout, as detailed in our guide to setting up the Meta Conversions API to feed accurate order values into your ad platforms in the first place.

How to Choose a Model Without Overthinking It

• New store, low volume. Start with last-click. It is honest about its own limitation — everyone knows it overvalues the close — and you don't have the conversion volume for anything more sophisticated to mean anything.
• Growing, multiple active channels. Move to position-based or time-decay. Both are auditable by hand, need no volume threshold, and stop crediting only the last click for journeys you can now see are longer than one touchpoint.
• High volume, clean tracking. Data-driven becomes genuinely useful once you clear GA4's thresholds and your server-side setup is feeding it complete, deduplicated events. Below that bar, it is last-click with a more expensive-sounding name.
• Any stage, before trusting the output. Run an incrementality test — a geo holdout or a platform-native conversion lift study — on your top one or two channels at least once a year. Attribution tells you where credit landed under a set of rules; incrementality tells you what happened to sales when you actually turned the channel off. They answer different questions, and a channel can win every attribution model while adding almost nothing incremental.

Frequently Asked Questions

Where Tugus Fits

Tugus is built around a problem the models above never solve on their own: the channels that introduce a customer and the channels that close them are structurally different jobs, and a single credit split hides that. Tugus separates the two explicitly, labelling each touchpoint's role in the path — opener or closer — rather than collapsing the journey into one weighted number. YouTube and organic social, which are nearly invisible under last-click, show up clearly as openers instead of disappearing from the report entirely. Full conversion paths stay inspectable per order, so a number in a dashboard can always be traced back to the actual sequence of touchpoints behind it.

That still leaves the deeper issue this article raised: attribution splits gross revenue, and gross revenue at checkout is not the number that should drive budget. Because Tugus captures cost of goods at purchase time and feeds returns and cancellations back from the shop and ERP as a structural part of the data model, credit can be evaluated against contribution margin rather than the order total on the day it was placed — so a channel with strong attributed revenue and a high return rate stops looking identical to one that actually keeps the money. Collection is server-side and consent-compliant throughout, hosted in the EU, feeding every connected destination from one pipeline.

If your current reporting can tell you which channel gets attribution credit but not which channel actually made you money after the returns came back, that gap is what a decision platform is for.