Attribution models compared: from last-click to data-driven on one journey

Attribution models are rules for deciding which marketing touchpoints get credit for a conversion. Last-click gives everything to the final interaction, first-click to the first, and linear, time-decay, position-based and data-driven models spread credit across the path in different ways. Because each model answers a slightly different question, the same sale can make a channel look essential or useless, so you need to know what each one assumes before you move budget.

Attribution model is the rule or algorithm that assigns credit for a conversion to the marketing touchpoints a customer interacted with before converting.

What are the main attribution models?

Marketing attribution splits into two families. Single-touch models give 100% of the credit to one interaction. Multi-touch attribution models share the credit among several interactions, either by a fixed rule or by an algorithm trained on your data.

The rule-based models are simple and transparent: anyone can recompute them on paper. Data-driven attribution is different. Google describes its version as a counterfactual approach that contrasts what happened with what could have happened, and says the models are trained partly on data from randomized controlled trials (Google Analytics Help). You cannot recompute it by hand, which makes it harder to audit, but it adapts to your actual paths instead of a fixed formula.

One caveat applies to all of them: a model can only credit touchpoints that were recorded. If your campaign links are not tagged consistently, organic and paid traffic get mixed up before any model runs. A clean UTM tagging convention is the precondition for every attribution report.

How does each attribution model split credit on one customer journey?

Here is a hypothetical journey for an online store. One customer places a single order worth $1,200 after four recorded touchpoints:

1. Day 1: clicks a paid social ad on Instagram and browses a product page.
2. Day 8: finds the store’s buying guide through organic Google search.
3. Day 15: opens a promotional email and clicks through.
4. Day 22: searches the brand name, clicks a paid search ad and buys.

Measured from the purchase, the touchpoints happened 21, 14, 7 and 0 days earlier. Now apply each model.

• Last-click: paid search gets the full $1,200.
• First-click: paid social gets the full $1,200.
• Linear: $1,200 ÷ 4 = $300 per touchpoint.
• Position-based (40/20/40): first touch 40% × $1,200 = $480, last touch $480, and the remaining 20% ($240) is shared by the two middle touches, $120 each.
• Time-decay (7-day half-life): a touchpoint’s weight halves for every 7 days before the conversion. The weights are 0.125 (21 days), 0.25 (14 days), 0.5 (7 days) and 1 (same day). They sum to 1.875, so paid social gets 0.125 ÷ 1.875 × $1,200 = $80, organic search $160, email $320 and paid search $640.
• Data-driven: there is no hand formula. For illustration only, suppose the model, trained on thousands of paths, returns 30% / 15% / 10% / 45%.

Every column adds up to the same $1,200. Attribution never creates or destroys revenue; it only moves it between rows. Yet paid social ranges from $0 to $1,200 depending on the column you read. If you calculate return on ad spend per channel from these numbers, the swing passes straight into your ROAS, and from there into next month’s budget. The ROAS formula guide shows how small changes in attributed revenue move that ratio.

When does each attribution model mislead you?

Each model is accurate about its own narrow question and misleading when you treat it as the full answer.

Last-click attribution overrates channels that harvest existing demand. Brand search, retargeting and coupon sites often sit at the end of the path because the customer has already decided. In the example, last-click says paid social earned nothing, so a team reading only last-click would cut the ad that started the journey. It is still useful for one thing: it shows which channels close.

First-click makes the opposite mistake. It rewards discovery and ignores everything that persuaded the customer later. It is also fragile: the first recorded touch is only the first one your tracking saw. Cookie expiry, a new device or consent choices can make a later touch look like the first one.

Linear looks fair but assumes every touch mattered equally. A glance at a newsletter gets the same $300 as the ad that introduced the brand. It tends to reward channels that simply appear often, such as email sequences and retargeting.

Time-decay builds in a bias toward the end of the path. It suits short, promotion-driven buying cycles, and it systematically undervalues awareness work in long B2B or high-ticket journeys, where the first touch may be weeks before the sale.

Position-based encodes an opinion that the opening and closing moments matter most. Sometimes that is true. The 40/20/40 split is a convention, not something measured from your data, so the middle of the funnel, where comparison and education happen, is shortchanged by design.

Data-driven attribution fits your real paths, which is a genuine improvement, but it inherits three limits. It needs enough conversion volume to learn from. It can only weigh touchpoints that were tracked, so offline and untracked channels are invisible. And because it is learned from observed behavior, it can still mistake “people who were going to buy anyway see this channel” for “this channel makes people buy”.

All of these models also share a structural blind spot: they only see converting paths built from tracked clicks and visits. A sales funnel analysis of where non-buyers drop off often explains more than another column of attributed revenue.

Which attribution models does Google Analytics 4 still offer?

If you learned attribution in Universal Analytics, the menu has shrunk. Google’s help center states that the first-click, linear, time-decay and position-based attribution models are no longer available in Google Analytics as of November 2023 (Google Analytics Help). The attribution reports now offer data-driven attribution plus two last-click variants: paid and organic last click, and Google paid channels last click.

Google Ads made the same change. Its help page says those four models are no longer supported, that conversion actions using them were upgraded to data-driven attribution, and that last click remains available (Google Ads Help).

Two practical consequences follow. First, if you want to compare a linear or position-based view, you have to compute it yourself from exported path data, as in the table above. Second, data-driven attribution is now the default in most Google reports, so a sudden change in channel credit after a migration may be a model change rather than a performance change. Check the model before you explain the trend.

Why is attribution not the same as causation?

Attribution answers “which touchpoints were on the path to this conversion?”. Budget decisions need a different answer: “how many conversions would we lose if we stopped this channel?”. The two can be far apart.

Go back to the brand-search click on day 22. Last-click credits it with the full $1,200. But the customer typed the brand name into Google. Had there been no paid ad, they would very likely have clicked the organic result just below it and bought anyway. The ad was present at the sale, which is not the same as causing it.

This gap is not a thought experiment. Gordon, Zettelmeyer, Bhargava and Chapsky compared randomized experiments at Facebook with common observational measurement methods across 15 large US advertising studies. They found the observational methods often failed to recover the effect measured by the experiments, even after controlling for extensive demographic and behavioral data (Gordon et al., 2019, Marketing Science). Attribution reports are observational by nature.

How does an incrementality test fill the gap?

An incrementality test, also called a lift or holdout test, randomly splits an audience. One group can see the campaign; a holdout group is deliberately not shown it. Because the split is random, the only systematic difference between the groups is the ad, so the difference in conversion rate is the effect the ad caused.

A hypothetical example: you test a retargeting campaign on its 200,000 eligible users for one month, 100,000 per group. During that month, last-click credits the campaign with 1,500 of the exposed group’s conversions.

• Exposed group: 2.0% convert, which is 2,000 conversions.
• Holdout group: 1.8% convert, which is 1,800 conversions.
• Incremental conversions: 2,000 − 1,800 = 200.
• Relative lift: (2.0% − 1.8%) ÷ 1.8% = 11.1%.

The campaign caused roughly 200 extra conversions among the 100,000 people it reached, a small fraction of the 1,500 that last-click attributed to it. Most of those buyers would have converted without the ad. That does not make attribution useless. It means you calibrate it: use attribution for daily optimization and use tests to check whether the channels it favors are actually adding sales.

Tests need enough users to detect a lift of this size. Before you launch one, size it with the method in the A/B test sample size guide; a 0.2-point difference is easy to miss with a small holdout.

How should you choose an attribution model?

There is no universally correct model, but there is a sensible process.

1. Fix tracking first. Consistent UTM tags, one definition of each conversion, and deduplicated users across devices. A good model on bad data is still wrong.
2. Match the model to your buying cycle. For impulse purchases with short paths, last-click and data-driven will be close. For long, considered purchases, read data-driven alongside a first-touch or position-based view so awareness channels stay visible.
3. Always compare two models side by side. The channels whose credit changes most between models are the ones where your budget decisions are most fragile.
4. Agree on one reporting model for the team. Mixing models across dashboards produces arguments, not decisions. Write down which model each marketing KPI uses.
5. Test the disputed channels. Where models disagree, or where a channel looks suspiciously efficient, run an incrementality test before you scale or cut it. The incrementality testing guide walks through holdout, geo and time-based designs and how to read the result.
6. Revisit after big changes. New channels, a new market or a pricing change alter paths, and any model trained on old paths drifts.

How do you analyze attribution in AdPix?

AdPix gives you the inputs that attribution depends on. Acquisition reports break traffic and conversions down by channel, source / medium and UTM-tagged campaign. The conversion path and funnel reports follow each step from first visit to sign-up or purchase, so you can see which touchpoints appear before a conversion and where users drop off. AdPix also connects a user’s visits across devices and channels into one journey, which reduces the “first touch that was not really first” problem described above. Reports are not sampled, so path counts for smaller campaigns are not estimates.

AdPix runs alongside Google Analytics without interfering, and many teams run both for a while to compare. That is a useful way to see how different measurement setups credit the same traffic; the trade-offs are covered in the Google Analytics alternative guide.

For the causal side, AdPix adds a measurement layer: campaign lift measurement with controlled holdouts and marketing mix modelling are included from the Starter plan up, and the Pro plan adds true campaign impact (incrementality) measurement and deeper channel analysis. Plan details are on the AdPix pricing page.

A practical next step this week: export the conversion paths for your top campaign, rebuild the table from this article with your own numbers, and mark the channel whose credit moves the most between last-click and the other models. That channel is your first candidate for an incrementality test.

Measure your site with AdPix today

The free plan covers 100,000 events a month, no credit card. Setup takes under five minutes.

Start freeSee pricing