Multi-Touch vs Last-Click Attribution in the AI Era

The attribution model debate that marketers spent a decade having was settled by force, not by argument. In GA4, the first click, linear, time decay and position-based models are no longer available as of November 2023. What remains is a two-horse race: last click, and Google’s machine-learning model, data-driven attribution. Every hand-tuned compromise in between - the models that let you argue about whether the first touch deserved 30% or 40% - is gone.

I think the deletion clarified something the debate obscured. The real choice was never between five ways of splitting credit. It is between a model you can explain and a model that might be right, and that trade reads differently depending on your data volume, your channel mix and who has to trust the number. This guide is my read on when each one is the honest choice, and on the limits both share that matter more than the difference between them.

What you can actually choose now

GA4’s attribution reports offer three models: data-driven attribution, paid-and-organic last click, and a Google-paid-channels last click variant. Data-driven attribution is the default, and on the ads side the positioning is the same - Google Ads makes it the default attribution model for most conversion actions.

So “multi-touch attribution” in 2026, for most marketers, does not mean a model you configure. It means DDA: a model that distributes credit across touchpoints by machine learning, with weights you never see. The configurable multi-touch era - assign 40% here, 20% there - existed because nobody knew the right split. Google’s answer was not to find the right split but to stop letting you ask, and I have some sympathy for that: the old models were arbitrary, and arbitrary numbers argued about confidently are worse than a black box labeled as one.

What data-driven attribution actually does

It is worth reading Google’s own description of the mechanism rather than the marketing name. DDA looks at path data from converting and non-converting users, incorporates factors like time from the key event, device type, number of ad interactions and the order of exposure, and then uses a counterfactual approach: it contrasts what happened with what could have occurred, to estimate which touchpoints actually shifted the probability of conversion.

The counterfactual part is the genuinely good idea. Instead of decorating a conversion path with credit after the fact, it asks the only question attribution was ever supposed to answer: did this touch make the conversion more likely than it would have been without it? The non-converting paths are what make that askable - a channel that appears in every path, converting or not, stops looking heroic.

My respect for the mechanism comes with two caveats I apply every time I read a DDA report. First, the weights are invisible. When credit moves from search to display in this month’s report, I cannot see why, cannot audit it, and cannot explain it to a CFO beyond “the model reweighted.” Second, the model is built, trained and defaulted by the largest seller of the ads it evaluates. I don’t treat that as a conspiracy. I treat it the way I treat every number a platform reports about its own product: as a claim from an interested party, to be checked against something the party does not control.

The case for last click, still

Last click is wrong in a way everyone understands, and that is a real virtue. It answers one narrow question - what was the final recorded step before conversion - with total transparency. Nothing reweights overnight. Two people looking at the same report see the same causality story. When a number has to survive a budget meeting, legibility is not a luxury.

Where I still reach for last click deliberately:

• Thin data. DDA is a statistical model, and on a few dozen conversions a month its confidence is costume. Below meaningful volume, I would rather have a simple number I know is biased than a sophisticated number that is mostly noise.
• Auditing and debugging. When conversions move, last click tells me literally what happened last. It is the raw view I check the modeled view against.
• Short, single-channel funnels. Google’s docs note that last click and DDA can produce the same results in some situations. If nearly all your paths are one touch long, the models converge and the simple one wins on explainability.

What I no longer accept is last click as an unexamined default, because the market has moved past defending it: in eMarketer and Snap’s survey of 282 US marketers at companies spending $500K+ on digital ads, only 21.5% were confident last-click attribution reasonably reflects a platform’s long-term business impact, and 74.5% were moving away from it or wanted to. Most people using it know it is a simplification. The discipline is saying so out loud in the report.

How I’d choose, case by case

My working rules, in the order I apply them:

1. Enough volume and a multi-channel paid mix: run DDA as the primary lens. This is what it is for, and the counterfactual logic beats any manual split you would have invented. If DDA also feeds your bidding, this decision was effectively made when you chose a Smart Bidding strategy - the bidder is already optimizing on modeled credit.
2. Low volume, or a mostly single-channel program: last click as primary, labeled as what it is: “final recorded step,” not “cause.”
3. Either way, look at both. GA4 keeps a comparison report under Advertising, then Attribution, then Attribution models, showing the same conversions valued under each model side by side. I read the gap as the finding: channels whose credit swings hardest between models are the ones doing early-funnel work that last click hides, and the ones worth an incrementality test before a budget shift.
4. Check the settings before trusting any of it. Lookback windows quietly shape every number: GA4’s defaults are 30 days for acquisition events and 90 days for everything else. A 90-day window in a 7-day sales cycle, or the reverse, distorts both models equally.

One rule sits above the others: model choice changes how credit is divided, never how much value exists. If the program didn’t work, no attribution model makes it work. That is why I would spend an hour choosing the model and a week pressure-testing whether the totals are real.

The limit both models share

Here is the part that has changed most since the old debate, and the reason I hold the whole question more loosely than I used to. Both models divide credit among recorded touches - clicks and visits your analytics saw. A growing share of buying research now happens where nothing is recorded: inside AI assistants, in answers that cite you without a click. When the only recorded touch is the branded search at the end, last click gives it 100% of the credit and DDA gives it 100% of the credit, because you cannot reweight touches that were never logged.

I’ve written a full piece on what attribution can and can’t see in the zero-click era, and the practical bridge is pulling AI referral traffic into its own GA4 channel so at least the clicks that do arrive from assistants are visible. But the honest summary is this: multi-touch vs last click is an argument about how to divide the visible slice, and the visible slice is shrinking. Choose the better model, then spend the saved energy on measurement that doesn’t depend on per-touch tracking at all.

FAQ

Is data-driven attribution the same as multi-touch attribution? DDA is a multi-touch model - it spreads credit across recorded touchpoints - but it is not the configurable multi-touch of the old debate. You do not set the weights. Machine learning estimates them from your converting and non-converting paths, and they change as the data changes.

What happened to linear and position-based attribution? GA4 removed first click, linear, time decay and position-based models in November 2023. Properties using them were moved to data-driven attribution. If a dashboard or an agency deck still shows “linear attribution,” it is being computed outside Google’s tools.

Does DDA need a minimum amount of data? Google has loosened its formal eligibility rules over time, but my practical bar hasn’t moved: a statistical model on tiny data produces confident-looking noise. If your monthly conversions are in the low dozens, read DDA as directional at best and keep last click beside it.

Which model should I report to leadership? Whichever you can defend, with its bias stated in the report. My format is one primary number, one sentence on what the model can’t see, and the model-comparison gap flagged when it is large. What erodes trust is not a biased model - it is a number that changes and a marketer who cannot say why, and on that score a report that labels its own confidence beats a precise-looking one every time.

This post is part of the analytics and ROI guides. The zero-click half of the problem is covered in marketing attribution in the zero-click era, the assistant traffic you can still see in the GA4 AI referral setup, and the wider practice starts at the complete guide to AI marketing.