What is causal attribution?
Causal attribution gives each channel credit only for sales that would not have happened without it. It is usually estimated with a holdout test, a geo experiment or a mix model. Last click and data-driven only split up sales that already happened.
By Joris van Huët, Founder & CEOUpdated 7 min read
Causal attribution credits each channel only with the sales it caused: the sales that would not have happened without it. It is usually estimated by comparing people or regions that met a channel with similar ones that did not. Last click and other rules just split up sales that already happened.
What one store's data shows
The usual answer calls causal attribution a cleverer attribution model: the same click paths, smarter maths. One store's export shows why no model on those paths can get there alone.
One store's anonymised GA4 export, 1 January 2024 to 21 August 2026. It holds shares of revenue only: no ad spend, no order counts.
| What the export shows | Share of revenue | Source cell |
|---|---|---|
| Distinct path sequences in the file (a count of path shapes, not a share) | 3,670 | Journeys sheet, path counts of all four rows added |
| Direct, in last click, first click and touched views | 57.7% | Channels sheet, Direct row |
| Journeys with 1 touch (0.5 days to buy) | 79.5% | Journeys sheet, 1 touch row |
Start with what the file holds. On the Journeys sheet, the four rows add up to 3,670 distinct path sequences (14 + 296 + 1,568 + 1,792). They are path shapes, not orders. And each one ended in a purchase. GA4 only draws a path when a key event happens: without a key event, there is no path to show.
So the export is a file of winners. It shows what buyers touched before they bought. It holds nothing on people who saw the same ads and never bought, or on what buyers would have done without a channel. Causal attribution lives in exactly that missing half.
Now the biggest row. On the Channels sheet, Direct holds 57.7% of revenue in every view. A causal reading would split that money three ways. Some buyers would have come anyway. Some were sent by ads or emails that left no trail. Some came from a friend's tip or a press mention. The file has no column for any of the three.
Single-touch journeys raise the same question. On the Journeys sheet, they hold 79.5% of revenue, at 0.5 days to buy. Attribution must give each of those sales to its one recorded touch. A causal reading may give part to an ad seen a week earlier, and part to nobody at all.
That last option is the heart of it. Google's Meridian guide defines the baseline as the expected outcome in the counterfactual scenario, with paid and organic media set to zero. Sales in the baseline belong to no channel. Attribution has no such bucket: every sale must land on some touch.
Why isn't a smarter model the same thing?
Because a model on click paths can only rearrange credit for sales that happened. To learn what a channel caused, you need a comparison: similar buyers, weeks or regions, with and without it. There are three ways to get one.
- A randomized test. The platform hides your ads from a random holdout and compares the two groups. Google calls such incrementality experiments the way to measure the causal impact of ads.
- A geo experiment. You switch a channel off in some regions and keep it on in others. Google's researchers describe regions randomly assigned to a control or treatment condition, and call the method conceptually simple.
- A model of spend and sales over time. A marketing mix model looks for sales that move with spend. Google's Meridian guide calls that causal inference from observational data, resting on an assumption no statistical test can confirm.
GA4's data-driven model sits in between. Google trains it on randomized trials of Google ad exposures, yet it still shares out only the touches GA4 logged. Its output is credit, not a measured effect.
How far off can a model be without a test?
Further than most dashboards admit. Researchers at Kellogg and Facebook took 15 large ad experiments at Facebook and tried to match them with observational methods. The methods often missed, even with rich demographic and behavioural data.
A later paper scaled that up to 663 experiments at Facebook. For lower-funnel outcomes, the median experimental lift was 5%. The better of two observational methods put it at 24%. The authors concluded they could not reliably estimate a campaign's causal effect from the observational data alone.
So treat any causal number built from history as an estimate to check. It is good for ranking channels and choosing what to test. A test is what settles a big budget move.
What can causal attribution not tell you?
- What each order owed to each ad. A test or a model gives an average effect over a period, never a verdict on one sale.
- What happens at another time or budget. Google describes lift tests as measuring effect at a certain point in time. Change the season, the offer or the spend, and the answer can move.
- Anything about a channel that never changed. If you spend the same every week, your history holds no contrast to learn from. A pause or a push gives the estimate something to measure.
- Whether your platform reports agree with it. Meta says its lift results are not meant to be compared with campaign results in Ads Manager. The two count different things over different windows.
What to do this week
- Find a pause you already ran. In GA4, open Reports > Acquisition > Traffic acquisition. Pick the weeks a channel was off, click Compare, then Apply. Pass: total revenue fell by about what the channel's row lost, or more, so its credit looks mostly real. Fail: the channel's row emptied but total revenue barely moved, so its credited sales probably came anyway; test it before you spend more.
- Ask Meta to run the randomized version. Meta's Conversion Lift guide wants a past-year campaign with USD 5,000 or more spent and 500 conversions. Tests are created in Meta's Experiments tool. Pass: a campaign clears both bars, so book the test there. Fail: none does, so plan your own regional holdout.
- Check whether Google can split your test by region. In Google Ads, open Campaigns, then Settings, and expand Locations on your biggest campaign. Pass: it targets one country, which Google requires for its geo-split Conversion Lift, so ask your Google representative. Fail: it spans several countries, so test one country at a time.
Check the homework. Your GA4 Attribution paths export already holds the evidence. Causality Engine reads that one file and shows what each channel caused next to what last-click gave it, in 1 to 2 minutes, for €99 once (excluding VAT), refundable within 30 days. Check the homework
Sources, 1 October 2026: Key events attribution paths report (Google); Assess the baseline (Google); About Conversion Lift (Google); Measuring Ad Effectiveness Using Geo Experiments (Google Research); About MMM as a causal inference methodology (Google); Get started with attribution (Google); A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook (Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science, via RePEc); Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement (Gordon, Moakler and Zettelmeyer, arXiv); Differences between Conversion Lift test results and other reporting tools (Meta); About Conversion Lift (Meta); Traffic acquisition report (Google); Change and compare date ranges in reports (Google); Set up Conversion Lift based on geography (Google); Exclude ads from geographic locations (Google)
Related answers
Frequently asked questions
Is causal attribution the same as incrementality?
Nearly. Incrementality is the extra sales a channel caused. Causal attribution is the practice of crediting each channel with that extra, whether a test measured it or a model estimated it. In both, sales that would have happened anyway belong to no channel.Can I do causal attribution without running experiments?
Yes, with care. A marketing mix model estimates effects from your history, and Google's Meridian guide calls that causal inference from observational data. It rests on assumptions no data can fully check, so confirm big budget moves with a holdout when you can.Why can't attribution paths show what a channel caused?
Because GA4 only builds a path when a key event happened, so the report shows buyers alone. Causal attribution needs a comparison: people, weeks or regions without the channel. A file of buyers cannot show what would have happened without your ads.
Go deeper: Incrementality testing, explained.
Sixty-second versions of these ideas: Causality Engine on YouTube Shorts.
Keep reading
Terms in this article
- Attribution ModelAn Attribution Model defines how credit for conversions is assigned to marketing touchpoints. It dictates how marketing channels receive credit for sales.
- Causal AttributionCausal Attribution uses causal inference to determine which marketing touchpoints genuinely cause conversions, not just correlate with them.
- Causal InferenceCausal Inference determines the independent, actual effect of a phenomenon within a system, identifying true cause-and-effect relationships.
- CounterfactualCounterfactual is a hypothetical outcome that would have occurred if a subject had received a different treatment.
- IncrementalityIncrementality measures the true causal impact of a marketing campaign. It quantifies the additional conversions or revenue directly from that activity.
- Incrementality TestingIncrementality Testing measures the additional impact of a marketing campaign. It compares exposed and control groups to determine causal effect.
- Marketing AttributionMarketing attribution assigns credit to marketing touchpoints that contribute to a conversion or sale. Causal inference enhances attribution models by identifying true cause-effect relationships.
- Marketing MixThe marketing mix is the set of actions a company uses to promote its brand or product. It traditionally includes product, price, place, and promotion.