Marketing Attribution Models Explained: One Journey, Nine Different Answers

Five people can look at the same order and disagree about which ad sold it. That is not a data quality problem. It is what happens when five different marketing attribution models score the same customer journey, and every one of them is working exactly as designed.

A marketing attribution model is a rule for splitting conversion credit across the touchpoints that came before a sale. Change the rule and the same order pays out to a different channel. Nothing about the customer changed. Only the arithmetic did.

Most guides on this subject list the models and stop there. This one runs a single ecommerce journey through nine of them and shows the working, then covers the two settings that move the answer as much as the model does: the lookback window, and whether an ad impression counts without a click. By the end you should be able to look at any channel report, name the model that produced it, and say what that model cannot see.

Start with the table. It is the whole article on one screen.

Eight of those nine rows add up to 180 dollars, because 180 dollars is what the customer paid. The ninth adds up to 720. That row is not a mistake and it is not a rounding artifact. It is the number you get when you open four dashboards and write down what each one claims, which is what most brands are actually doing when they say their attribution is broken.

What a marketing attribution model actually does

An attribution model does one job: it takes an ordered list of touchpoints and a conversion, and returns a split. That is the entire mechanism. Everything else in an attribution product, the tracking, the identity resolution, the deduplication, exists to produce a trustworthy list of touchpoints for the model to score.

This distinction matters more than it sounds. The model is a policy decision. The touchpoint list is a data problem. Teams routinely spend a quarter arguing about whether to run first click or position based, when the thing costing them money is that a third of their orders arrive with an empty touchpoint list and land in a bucket called Undefined.

Attribution also gets confused with two neighbours that answer different questions. Attribution asks which touchpoints preceded the sales you got. Incrementality asks how many of those sales you would have got anyway. Media mix modelling asks how total spend by channel correlates with total revenue over time, without needing user-level tracking at all. They are not competitors, they are three instruments with different resolutions, and the section near the end covers when to reach for which.

Nielsen surveyed global marketers managing budgets of a million dollars or more for its 2024 Annual Marketing Report and found 84 percent say they are confident in their ROI measurement, while only 38 percent measure traditional and digital marketing together. Confidence and coverage are moving in opposite directions, and attribution modelling is where that gap becomes visible.

The nine attribution models, and what each one is blind to

Every model below is defined by what it credits and, more usefully, by what it refuses to see. Read the second half of each description first.

First click

The first recorded touchpoint takes 100 percent of the credit. In the journey above, TikTok gets all 180 dollars.

First click answers one question well: what introduces people to the brand. It is the right lens for prospecting budget, and it is why it tends to be the default starting point in ecommerce tools rather than last click. It is blind to everything that closed the sale. Run your whole business on first click and your retargeting and email programmes look worthless, because by construction they can never earn a dollar.

Last click

The final touchpoint before the order takes everything. Here that is a direct visit, so 180 dollars goes to a channel you cannot buy more of.

That failure is the most common one in ecommerce. A shopper who already knows you types the domain, or opens a bookmark, or taps a saved tab. Last click reads that as the reason they bought. It is blind to the entire top of funnel, and it systematically overpays branded search and retargeting, which are the two places where the customer had already decided.

EMARKETER surveyed US digital advertisers spending 500,000 dollars or more on digital ads in October 2024 and reported that only about 22 percent are confident last-click measurement reflects long term business impact, with roughly 75 percent moving away from it. Most teams already know last click is wrong. Fewer have replaced it, because replacing it means picking one of the models below and defending the choice.

Last non-direct click

Identical to last click, except direct visits are skipped. Credit falls to the last touchpoint you can actually attribute, which here is the Klaviyo email.

This is the single highest-value change most brands can make, and it costs nothing but a setting. It removes the worst artifact of last click without changing anything else. It is still blind to the first four fifths of the journey, and it now overpays email and SMS specifically, because those are usually the last non-direct thing a returning customer touches.

Linear

Every touchpoint gets an equal share. Five touchpoints, 36 dollars each.

Linear is honest about its own ignorance, which is a real virtue. It makes no claim about which touch mattered. The cost is that it treats a three second scroll past a TikTok ad as exactly equal to a click on a discount email, and it rewards channels that generate many cheap touchpoints. A brand running heavy retargeting will see retargeting inflate under linear purely on volume.

Time decay

Credit grows the closer a touchpoint sits to the order, usually on a half life. The table uses a seven day half life, so a touch fourteen days out is worth a quarter of a touch on the day of purchase. That is where 19, 25, 34, 46 and 56 come from.

Time decay suits short consideration windows and promotional calendars, where the last week genuinely does most of the work. It is blind to the fact that recency is not influence. A brand with a ninety day consideration cycle running a seven day half life is telling itself that nothing before the final fortnight exists.

Position based, also called U-shaped

Adobe publishes the standard split as 40 percent to the first touch, 40 percent to the last, and 20 percent spread evenly across everything between. In the journey above that is 72 dollars to TikTok, 72 to the direct visit, and 12 dollars each to Google, Meta and Klaviyo.

Position based encodes a defensible belief: discovery and conversion are the two hard moments, and the middle is maintenance. It is a good default for a brand that has never picked a model deliberately, because it credits both ends of the funnel and is easy to explain to a finance team. It is blind to journeys where the middle is the whole story, which is most considered purchases above a couple of hundred dollars.

W-shaped

Adobe also publishes the W-shaped split as 30 percent each to three named milestones, with 10 percent shared across the rest. In business to business those milestones are first touch, lead creation and opportunity creation. Ecommerce has no lead creation event, so the usual mapping is first touch, the touch preceding the first add to cart, and the last touch. The table assumes the add to cart followed the Meta retargeting click.

W-shaped is the model most often copied into ecommerce from a business to business article without the mapping being thought through. If you cannot name the middle milestone and point at the event that fires it, you do not have a W-shaped model. You have a position based model with extra steps.

Data-driven, or Shapley value

Rather than applying a fixed rule, this family estimates each channel's marginal contribution across your entire order history: how much does the conversion rate change when this channel is present versus absent, holding the rest constant. The mathematics comes from cooperative game theory, and the Shapley value is the standard method for allocating credit fairly among cooperating players.

The data-driven row in the table above is illustrative, and it has to be. No data-driven model can score a single journey, because the model is fitted to the whole dataset. That is the honest caveat every vendor table leaves out. Given enough orders it is the most defensible option available, since it derives the weights from your data instead of importing someone else's assumption. Given a few hundred orders a month it is a confident-looking number with very little behind it.

Amplitude groups all of the above into single-touch and multi-touch families and adds a J-shaped variant that gives 60 percent to the last touch, 20 percent to the first, and 20 percent to the middle. The taxonomy is useful. The proliferation of named shapes is mostly noise once you understand that every one of them is a weight vector.

Platform overlap

The last row is not a model you would choose. It is a reconstruction of what you already have. Each ad platform runs its own last-touch logic inside its own window on its own data, and each one reports the full order value. Meta claims it. Google claims it. TikTok claims it. Klaviyo claims it. Nobody deducts anything for the others.

Reproducing that overlap deliberately is genuinely useful, because it lets you put platform-reported revenue and deduplicated revenue side by side and quantify the gap instead of arguing about it. The gap is the number worth reporting to your board, not either figure on its own.

Single-touch versus multi-touch matters less than you think

The standard framing splits these nine into single-touch models, which give everything to one touchpoint, and multi-touch models, which spread credit. It is a clean taxonomy and it is how most of the internet organises this topic.

It is also close to useless as a decision aid, because the meaningful difference is not how many touchpoints get paid. It is which direction the model errs. First click and last click are opposites that happen to share a category. Position based and time decay are both multi-touch and disagree violently about the same journey: in the table, position based pays TikTok 72 dollars and time decay pays it 19.

A more useful question than single or multi is this one. If this model is wrong, which channel gets defunded? Every model has a systematic bias, and you should choose the bias you can live with rather than the one that sounds most sophisticated.

The lookback window changes the answer as much as the model does

A lookback window is the maximum age of a touchpoint that still counts. Anything older is discarded before the model runs. It is a filter on the input, not a rule about the output, which is exactly why it gets ignored: it sits in a different settings panel from the model dropdown.

Run the same journey through first click at three different windows.

Three windows, three different channels credited with the same sale, and the model never changed. Anyone comparing two attribution reports without first checking that the windows match is comparing nothing. This is also the most common reason a brand's numbers move after switching tools: the new platform shipped with a different default.

Set the window from your actual time to purchase, not from a convention. Pull the distribution of days between first touch and order, take something near the ninetieth percentile, and use that. A brand selling 40 dollar consumables and a brand selling 900 dollar furniture should not be running the same window, and they usually are, because neither ever changed it.

Click-based versus view-through: where your numbers and Meta's diverge

The second setting that quietly rewrites your report is whether an impression counts as a touchpoint when nobody clicked.

Ad platforms say yes. A default Meta configuration credits a conversion to an impression seen within one day, alongside clicks within seven. That is a large surface on which to claim revenue, and it is applied to an audience the platform selected partly because they were already likely to buy.

Most independent attribution is click-based, because a click is an observable, first-party event and an impression on someone else's platform is not. This is the honest explanation for most of the gap between a brand's own numbers and the platform's, and it is not a bug in either system. They are counting different things, and only one of them is countable from your own site.

The practical resolution is not to pick a side. It is to keep a click-based model as the number you plan against, and to run a platform-matched view alongside it when you need to reconcile with what a media buyer sees in Ads Manager. Polar exposes both: ten attribution models in total, including click-based models like first click, last click, linear, time decay, position based and a Shapley-based data-driven model, plus paid-only variants and a platform overlap mode that folds in Meta-reported view-through so the two views can be compared directly rather than argued about.

For anything above the click layer, such as television, connected television, podcast or direct mail, no attribution model will help, because there is no touchpoint to record. That is incrementality testing territory.

Every model is wrong when the touchpoint data is incomplete

Here is the part the model comparison articles skip. All nine models take the touchpoint list as given. If that list is missing touches, every model is wrong, and they are all wrong in the same direction: credit collapses toward whatever survived.

Touchpoints go missing for structural reasons, not fixable ones. A visitor rejects cookies. A visitor runs an ad blocker. A visitor sees the ad on a phone and buys on a laptop three days later. Client-side, cookie-based measurement loses the thread in all three cases and fills the gap with a modelled estimate. No amount of tag auditing recovers a touch that was never recorded.

The failure mode is specific and worth recognising in your own reports. When the early, cheap, high-volume touches are the ones most likely to be lost, first click and position based degrade fastest, and last click looks deceptively stable. A model that looks unaffected when your tracking degrades is usually just a model that was already ignoring the data you lost. The same brands then see a large Undefined or Unknown bucket and read it as a reporting quirk rather than as the size of their blind spot.

This is why the tracking layer, not the model dropdown, is where attribution accuracy is actually won. First-party tracking with server-side enrichment keeps recording when cookies are rejected. A persistent identity stitches sessions across devices without guessing. Revenue ties back to the real order rather than a modelled one. Get that right and the choice of model becomes what it should be, a reporting preference rather than a source of anxiety. Get it wrong and you are running a sophisticated model over a sample you cannot describe. Our write-up of the gaps in Shopify attribution walks through where those touches disappear in practice.

How to choose an attribution model for an ecommerce brand

Choose by question, not by sophistication. The model is a lens, and running two is normal.

Three rules make the choice stick.

Pick one model as the number of record and write it down. Not because it is correct, but because a channel budget argued under three different models is not an argument about performance. Publish the model and the window next to the number in every report.

Second, keep a contrasting model visible. If first click is your number of record, keep last non-direct click on the same screen. The spread between the two is a better signal than either value: a channel that scores high on both is genuinely carrying weight, and a channel that scores high on one and near zero on the other is doing one specific job.

Third, review the choice when the business changes shape, not on a calendar. A new product line at four times the price, a first television campaign, a shift from prospecting to retention: each of those changes the journey, and the model that fitted the old journey will quietly misprice the new one. For a walk through of how these models behave on Shopify order data specifically, see our breakdown of attribution models for Shopify brands.

Attribution models, incrementality and media mix modelling

Attribution models answer a narrow question: among the sales you got, which touchpoints preceded them. They cannot answer whether those sales were caused by the marketing, because every touchpoint in the dataset belongs to somebody who converted. Retargeting looks excellent under almost every model for exactly this reason. You are showing ads to people who were already going to buy, then taking credit when they do.

Incrementality testing answers the causal question by withholding. Split comparable regions into treatment and control, run the campaign in one, and measure the difference in total revenue. It requires meaningful scale and a few weeks of patience, and it is the only method that produces a number you can honestly call caused.

Media mix modelling works at the top of the stack. It regresses total revenue against total spend by channel over a long history, with controls for seasonality and promotions, and needs no user-level tracking at all. That makes it the right instrument for offline and upper-funnel media, and the wrong one for deciding which of two ad sets to pause on Thursday.

Use all three at their own resolution. Attribution for weekly channel and campaign decisions, incrementality to check whether a channel deserves its budget at all, media mix modelling for annual planning across everything including media you cannot track. Trouble starts when a team asks one of them to do another's job, most often when attribution numbers are quoted as proof of incremental impact.

Putting this into practice this week

Four steps, in order, none of which require a new tool to start.

First, write down the model and window behind every attribution number your team currently quotes. In most organisations this exercise alone surfaces two or three numbers that nobody can trace, and at least one report comparing figures produced under different settings.

Second, measure the overlap. Add up what every platform claims for a single week and compare it to what the store actually took. The ratio is your double-counting factor, and it is usually somewhere between two and four. That single number does more to change how a team reads channel reports than any model change.

Third, set the window from your own time to purchase distribution rather than the default. If you cannot pull that distribution, that gap is the more urgent problem.

Fourth, fix the input before optimising the rule. Check how many orders arrive with no recorded touchpoint. If it is more than a small minority, the model is not what is costing you money, and switching it will not help. First-party tracking is the fix, and it is the layer Polar's attribution is built on, with ten models and a configurable lookback window sitting on top of a first-party pixel that keeps recording when cookies are rejected.

A final note on defaults. Stape points out that Google deprecated first click, linear, time decay and position based in its own tools in 2023, leaving GA4 with data-driven attribution, paid and organic last click, and Google paid channels last click. If your reporting runs through GA4, your model choice has already been made for you, and it is worth knowing which one you inherited.

FAQ