Attribution models compared

Last click gives all credit to the final touchpoint before the conversion. Its near-universal variant, last non-direct click, adds one amendment: if that final visit was direct, step back to the last visit with a known source. This is the default logic behind most acquisition reporting everywhere, because it's simple, explainable and only needs one session's context.

First click is the mirror image - all credit to the touchpoint that started the journey, on the theory that discovery is the hard part. Between the two extremes sit the split-credit models: linear spreads credit evenly across every touchpoint, time decay weights credit toward the touches closest to the conversion, and position-based gives 40 percent each to the first and last touch with the remaining 20 spread across the middle.

Data-driven attribution replaces all of these fixed rules with a computed one. Google's version compares the paths of people who converted against the paths of people who didn't, and assigns each touchpoint credit in proportion to how much its presence raises the odds of converting, factoring in things like touch order, timing and device. The credit split is different for every account and recomputed as data changes.

In November 2023 Google removed first click, linear, time decay and position-based from GA4 entirely, migrating everything to data-driven attribution. What remains is a three-item menu: data-driven, which is the default, paid-and-organic last click, which is the last non-direct rule described above, and a Google-paid-channels last click that exists mainly for ads reporting.

So the classic blog-post comparison of six models is now mostly history lesson. If your measurement plan involves position-based attribution, it involves a tool other than GA4. The retired models still matter for one reason: they name the assumptions - discovery matters, closing matters, everything matters equally - that the data-driven black box now arbitrates invisibly.

The least-known fact about GA4 attribution is how little of the interface the setting controls. The attribution model applies to event-scoped traffic dimensions - the source, medium and campaign attached to a key event in reports like Advertising's attribution views. Session-scoped and user-scoped dimensions, which is what the standard Traffic acquisition and User acquisition reports are built on, stay on last-non-direct-click and first-touch logic respectively no matter what you pick.

Practical translation: switching the model will quietly change the Advertising section's numbers and leave the reports most people live in untouched. If two teammates cite different conversion counts for the same channel, the first thing to check is which report each was reading, not who miscounted - a cousin of the tool-vs-tool discrepancies problem.

Every model runs inside a lookback window - the maximum age of a touchpoint that can still earn credit. In GA4 the window for acquisition events is 30 days by default and can be shortened to 7, while other key events look back 90 days by default, adjustable to 30 or 60. A touchpoint older than the window gets nothing, however important it was.

Windows are the quiet reason long-consideration businesses undercount their top of funnel: a blog visit 4 months before a B2B purchase is outside every default window and simply doesn't exist to the model. They're also a fence worth knowing when comparing tools, since ad platforms run their own windows - counting view-throughs and using different lookbacks - which is a large part of why platform-reported conversions never match analytics.

Last-click credit goes to whatever channel catches people on their final approach, and the final approach is usually navigational: someone who already decided searches your name and clicks the top result, or types the URL. So branded search - paid or organic - hoovers up credit for demand that something else created. The blog post that introduced you, the podcast mention, the friend's recommendation all show up earlier in the path or not at all, and earn nothing.

The "not at all" case is bigger than it looks. SparkToro's research on dark social found most content sharing happens in private channels - messages, email, chat - where no referrer survives, so the recommending touch arrives labeled as direct traffic. Last non-direct click then reassigns that person's conversion to whatever tagged channel they touched last, often the brand search that merely finished the job. The model isn't broken, it's answering a narrow question - who closed - and the mistake is spending as if it answered who created demand.

The appeal of data-driven attribution is that it stops arguing about philosophy: instead of decreeing that first or last touches matter, it measures which touchpoints actually distinguish converting paths from non-converting ones in your data. Upweighting a display campaign that keeps appearing in successful paths is exactly the correction the fixed models can't make.

What you give up is auditability. No one can recompute the credit split by hand, the split shifts as the model retrains, and thin data makes the estimates wobbly - the machinery needs conversion volume to find patterns, so small sites get something closer to a dressed-up last click. Treat its output as a better default, not an oracle: when it reallocates meaningful budget, that's a hypothesis to test, not a verdict to obey.

Universal Analytics veterans keep looking for the assisted conversions report, and it's gone: the concept - counting how often a channel appeared in a path without closing it - was replaced in GA4 by the attribution paths report. It splits paths into early, middle and late touchpoints and shows fractional credit per channel under the selected model, along with days and touch counts to conversion.

It's a genuinely better report for the same question. A channel that keeps showing up in the early segment of converting paths is your prospecting engine even if it never closes, and that pattern is exactly what a last-click view of the same data hides.

Every attribution model, data-driven included, splits credit among the touchpoints that happened. None can say whether the conversion needed them. The person who searched your brand and clicked the ad might have clicked the free organic result below it - attribution gives the ad full credit either way. The only way to measure caused conversions is an incrementality test: hold out a group who don't see the campaign, compare conversion rates, and read the lift. Ad platforms offer these as conversion lift studies, and branded search is the classic first candidate.

The cheap companion is self-reported attribution - asking new customers how they heard of you. Rand Fishkin's writing on the question documents its limits (people misremember, and recent touches crowd out first ones), and the standard upgrade is asking two questions: how did you first hear about us, and what prompted you today. The answers are directional, and they're the only instrument that sees podcasts, communities and word of mouth at all. Disagreement between the survey and the model is not an error - it's the map of what your tracking can't see.

For most sites the honest setup is boring: leave the default, know that acquisition reports run on last non-direct click, and read channel numbers as "who closed" rather than "who caused." Consistency beats sophistication - a model held steady turns into trend data, while a model switched quarterly turns every comparison into archaeology. Layer a two-question survey on top, and when one channel's budget is large enough to matter, buy the answer properly with a lift test.

Whatever the model, it can only re-slice the source data it's given, so the unglamorous work is upstream: tag campaigns consistently so touches carry labels, and keep an eye on the size of your direct bucket. In Bigdelta, every visit is attributed to its source and campaign and UTM breakdowns come standard - which is the session-level ground truth any attribution logic, simple or learned, is built on.