What Attribution Model Should You Use? (2026)

Start here, because most of what you'll read on this topic is out of date: in GA4 you can choose from three attribution models, not seven. First click, linear, time decay, and position-based were removed in November 2023 — Google's own documentation says they are "no longer available as of November 2023" (Google Analytics Help). Google Ads did the same, leaving last click and data-driven, and auto-upgrading anyone still on a deprecated model (Google Ads Help).

That was nearly three years ago. The top-ranking guides for this exact question still walk you through picking between all seven, complete with diagrams. If you've been trying to reconcile advice you read against a settings menu that doesn't match it, this is why. Below: what actually exists, how data-driven attribution really works (it isn't what everyone says), and what to use when the model you want isn't on the menu.

Key Takeaways

• GA4 offers three attribution models today: data-driven, paid and organic last click, and Google paid channels last click. Four others were removed in November 2023.
• The near-universal claim that GA4's data-driven attribution uses the Shapley value traces to a legacy Universal Analytics document for a sunset product. Current GA4 docs say only "counterfactual approach."
• Every conversion action is eligible for data-driven attribution regardless of volume. The often-quoted "200 conversions" is a Google recommendation, not a requirement.
• Across 2,226 Meta experiments, last-click explained just 19% of the variance in actual measured incrementality (R²=0.19). No attribution model is a substitute for an experiment.

What Attribution Models Still Exist

Three remain in GA4:

• Data-driven — Google's algorithmic model, now the default
• Paid and organic last click — last click across all channels, ignoring direct
• Google paid channels last click — last click among Google paid channels only

Four are gone: first click, linear, time decay, and position-based. Google Ads went further, keeping only last click and data-driven.

Why this matters beyond trivia: if a strategy deck recommends a position-based model, that deck cannot be executed in GA4. The choice you actually face is narrower and simpler than the literature suggests — data-driven, or a flavour of last click. Everything else is a conversation about tools outside Google's stack.

How Data-Driven Attribution Actually Works

Here's the claim you'll see everywhere: GA4's data-driven attribution uses the Shapley value from cooperative game theory. It's stated confidently, in a lot of otherwise-decent articles.

Go to Google's current GA4 documentation and the word Shapley does not appear. What it says is that data-driven attribution uses a "counterfactual approach" — comparing what happened against what a model predicts would have happened otherwise — with machine learning across both converting and non-converting paths, considering factors like time from conversion, device type, number of interactions, exposure order, and creative asset type (Google Analytics Help).

So where does Shapley come from? It comes from a different document, for a product Google has shut down. The Multi-Channel Funnels data-driven attribution methodology page does describe a Shapley Value approach — and it carries a banner reading that you are viewing a legacy article about Universal Analytics (Google Analytics Help, legacy). Universal Analytics stopped processing data in 2023. As far as we can establish, the Shapley explanation in circulation traces back to that page — documentation for a system that no longer runs.

The honest position: Google historically documented a Shapley-based method for Universal Analytics' MCF reports. For GA4 it documents a counterfactual machine-learning approach and has not disclosed the algorithm. Anyone telling you precisely how GA4's DDA computes credit is guessing.

The "200 Conversions" Rule Isn't a Rule

Another widely-repeated claim worth correcting, because it changes what people do.

Google's documentation states that all conversion actions are eligible for data-driven attribution, regardless of conversion or interaction volume (Google Ads Help). There is no minimum. On the same page, Google separately recommends at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period.

Eligible regardless of volume, but recommended above a threshold. Those are different statements, and collapsing them into "you need 200 conversions to use DDA" sends small accounts to last click unnecessarily. The recommendation is about how much you should trust the output, not about whether the switch turns on.

No Attribution Model Measures Causation

This is the part the model-comparison articles skip, and it's the part that matters most.

Attribution models divide credit among touchpoints that preceded a conversion. None of them tell you whether the ads caused it. That distinction has been measured, and the results are not flattering.

Researchers at Meta and Northwestern compared attribution against randomised experiments across 2,226 Meta advertising experiments. Last-click attribution had an R² of 0.19 against actual experimentally-measured incremental conversions per dollar — meaning it explained about 19% of the variance in what the experiments showed really happened. Their proposed method, Predicted Incrementality by Experimentation, reached R² = 0.88. Last click also disagreed with the experiment-based decision in 12–20% of campaigns (Gordon, Moakler & Zettelmeyer, arXiv, revised April 2026).

The same research group's earlier peer-reviewed work reached a compatible conclusion at enormous scale — 15 US experiments, 500 million user-experiment observations, 1.6 billion ad impressions — finding that observational methods "often fail to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables" (Gordon, Zettelmeyer, Bhargava & Chapsky, Marketing Science 38(2), 2019).

Note what that research does not say. It doesn't say last click undercredits the upper funnel by some fixed percentage — an "X% undercredit" figure is a thing you'll see quoted and a thing that does not exist in any verifiable form. The direction isn't even consistent: last click can overestimate ad effects when ads are targeted at people who would have bought anyway, and underestimate when conversions fall outside the attribution window. The error isn't a constant you can correct for. That's worse news than a fixed bias would be, and it's the actual finding.

Choosing Between Attribution, MMM, and Incrementality

Since no model measures causation, the real question is which instrument to use for which decision.

	Attribution (DDA / last click)	Marketing mix modelling	Incrementality testing
What it does	Divides credit among tracked touchpoints	Models aggregate spend against outcomes	Randomised holdout — measures causal lift
Data needed	User-level tracked paths	Aggregate, geo-level; no user-level data	Enough scale to power a test
Answers	"Which touchpoints preceded conversions?"	"How does spend relate to outcomes overall?"	"Did this actually cause incremental sales?"
Blind spot	Confuses correlation with cause	Low granularity; needs long history	Costly; not for routine decisions
Open-source option	—	Meridian (Google), Robyn (Meta)	Platform lift tools

Both major platforms now ship open-source MMM. Google's Meridian uses Bayesian causal inference with a NUTS MCMC sampler, under Apache-2.0 (GitHub). Meta's Robyn uses ridge regression with evolutionary hyperparameter optimisation, under MIT — though note its last release was December 2024, and its Python port is self-described as an LLM-translated beta that may contain bugs (GitHub). The maintenance signal there is worth weighing before you build on it.

The practical sequence: use attribution for day-to-day directional steering, because it's the only thing available at that cadence. Use incrementality tests before big budget shifts, because they're the only thing that measures cause. Reach for MMM when you have years of history and need to reason about channels that resist tracking entirely.

What This Looks Like in Southeast Asia and India

Honestly? Exactly the same — and that's itself a finding worth stating.

We looked for attribution research specific to Singapore, India, or Indonesia: model performance studies, regional platform differences, any benchmark on which model works better where. There is none. No study, no platform documentation, no regional dataset. If you see a claim about attribution behaving differently in SEA, ask where it came from.

The one thing that is genuinely regional isn't about models at all — it's consent. Google Consent Mode v2 applies to the European Economic Area, the UK and Switzerland, and not to Singapore, India, or Indonesia (Google). Global agency content routinely implies otherwise. Our companion guide to First-Party Data Rules in Southeast Asia & India covers the actual regional obligations market by market.

Statistics We Rejected Writing This

This topic is unusually polluted. Roughly speaking, if a number about attribution sounds decisive, it's invented. These all failed verification:

1. "Organizations implementing MTA see a 19% improvement in marketing ROI" — attributed to Forrester. No such report exists on forrester.com.
2. "64% of attribution implementations fail to reflect reality" — attributed to Gartner. Not on gartner.com.
3. "78% of marketing leaders say attribution data doesn't match revenue reports" — attributed to Forrester, 2025. Not locatable.
4. "McKinsey: 15–20% marketing ROI improvement from MTA." The number is real; the claim isn't. McKinsey's figure refers to integrated analytics freeing up marketing spend generally — not multi-touch attribution. A real number attached to the wrong claim is harder to catch than an invented one.
5. "Last-touch misses up to 90% of the credit owed to upper-funnel." No traceable primary source, and it would be a vendor claim even if there were.

Frequently Asked Questions

What attribution models are available in GA4?

Three: data-driven, paid and organic last click, and Google paid channels last click. First click, linear, time decay, and position-based were removed in November 2023 per Google's own documentation. Google Ads retains only last click and data-driven, and auto-upgraded accounts still using deprecated models.

Does GA4's data-driven attribution use the Shapley value?

Google's current GA4 documentation doesn't say so — it describes a "counterfactual approach" using machine learning across converting and non-converting paths. The Shapley claim traces to a legacy Universal Analytics document about Multi-Channel Funnels, a product that stopped processing data in 2023. Google has not disclosed GA4's DDA algorithm.

Do I need 200 conversions to use data-driven attribution?

No. Google states all conversion actions are eligible for data-driven attribution regardless of conversion or interaction volume. The "at least 200 conversions and 2,000 ad interactions within 30 days" figure is a Google recommendation about result reliability, not an eligibility requirement.

Which attribution model is most accurate?

None of them measure causation, so "accurate" is the wrong frame. Across 2,226 Meta experiments, last-click attribution had an R² of 0.19 against experimentally-measured incrementality. Use attribution for directional day-to-day steering and incrementality tests for decisions that actually matter.

Should I use MMM or multi-touch attribution?

They answer different questions. MMM models aggregate spend against outcomes without user-level data and suits channels that resist tracking; attribution divides credit among tracked touchpoints at daily cadence. Both major platforms ship open-source MMM — Google's Meridian and Meta's Robyn — though Robyn's last release was December 2024.

Does attribution work differently in Southeast Asia or India?

There's no research showing it does — we searched for regional attribution studies and found none. The genuine regional difference is consent, not modelling: Google Consent Mode v2 covers the EEA, UK and Switzerland only, and does not apply to Singapore, India, or Indonesia.

Conclusion

The attribution question has a smaller answer than the internet suggests. You have three models in GA4, two in Google Ads, and one of them — data-driven — is the default for good reason. Nobody outside Google knows exactly how it computes credit, and the Shapley explanation you've read is describing a product that no longer exists. What no model does is tell you whether your ads caused anything, and the research measuring that gap is unambiguous: pick a model for steering, run an experiment before you bet.

How this post was compiled. Every factual claim about model availability, data-driven attribution mechanics, and eligibility thresholds was verified directly against Google's own documentation on 2026-07-12, with each cited page confirmed live. The incrementality findings come from peer-reviewed and preprint research by Gordon, Moakler, Zettelmeyer and colleagues, cited with sample sizes. Five widely-circulated attribution statistics failed verification — including two attributed to Forrester and one to Gartner that do not appear on those firms' sites — and are named in the article rather than quietly omitted. Where Google documents a mechanism but publishes no comparative outcome data, we've said so rather than inferring a result. No attribution research specific to Singapore, India, or Indonesia exists as far as we could establish. Coact is a performance marketing agency working with ecommerce and app businesses across Singapore, India, and Indonesia.

Continue Learning

• How to Set Up Conversion Tracking That Actually Works — the measurement layer underneath any attribution model
• First-Party Data Rules in Southeast Asia & India — what Consent Mode v2 actually requires in your markets
• Why Is My CAC Increasing? What the 2026 Data Shows
• Talk to our team about measurement that survives contact with an experiment.