What Is Causal Inference and Why Every Marketer Needs It

Causal inference is the statistics of what would have happened otherwise: when Ads Manager reports 200 purchases from a campaign, it asks how many of those you'd have had with the campaign switched off. The difference is what the campaign caused. Marketers need it because attribution reports, from last click in GA4 to Meta's own numbers, share out credit for sales without asking that question. A sales spike in the week a campaign launched can be the campaign, or payday, or the weather, and only a comparison with what would have happened otherwise tells them apart [3].

It's the same trap as believing ice cream causes shark attacks because both rise in summer: a third factor, the heat, drives both. Statisticians call that third factor a confounder, and marketing is full of them: seasonality, payday, a competitor running out of stock, a mention in the press. When one of them lifts sales in the same week as your campaign, the report hands the campaign the credit, and nothing on the screen tells you.

The High Cost of Correlation

Marketing attribution assigns credit for a sale to the touchpoints a customer passed on the way to it.

For years, the industry has relied on marketing attribution models for this. First-touch, last-touch, linear, time-decay: each promises to tell you which channels deserve credit for a sale. GA4 removed first click, linear, time decay and position-based models in November 2023, but plenty of other tools still offer them. They're systems for sharing out credit, not for finding cause. They look at a customer's journey and assign value to touchpoints, but never answer the question that matters: would the sale have happened anyway?

Where each common model goes wrong:

• First-Touch Attribution: gives 100% of the credit to the first touchpoint. It rewards the channels that create awareness and ignores everything that happens after. It's like crediting a marathon to the friend who first mentioned the race, not the months of training or the final sprint.

• Last-Touch Attribution: the most common model, and often the most misleading. It gives 100% of the credit to the final touchpoint, so branded search ads, which people click when they're already about to buy, collect the credit for sales that social campaigns or influencers started weeks earlier.

• Linear Attribution: gives equal credit to every touchpoint. It looks fairer, but it's still a guess: a glance at a banner ad gets the same weight as a long product video.

• Time-Decay Attribution: gives more credit to touchpoints closer to the sale. Recent touches may well matter more, but the rate of decay is a setting someone chose, not something anyone measured.

That's how one channel ends up cannibalizing another. A Google Ads campaign can get 100% of the credit for a sale because it had the last click, when the customer decided to buy after a TikTok video three weeks earlier. The ad collected a sale that was already going to happen. The report rewards the wrong channel, and the budget scales a campaign that adds nothing. It isn't a small error: across a whole account it can put a real share of the budget into channels that aren't winning new customers, which is how correlation-based decisions waste budget.

For a Dutch beauty or fashion brand, where every euro of ad spend gets questioned, that uncertainty is expensive. Causal inference is how you shrink it.

The Causal Revolution: From Guesswork to Evidence

See also: Best Causal Inference Tools for Marketing Attribution

Causal inference is the scientific process of working out cause and effect from data. Correlational analysis shows that two variables move together; causal inference uses experiments and statistical methods to estimate whether a change in one causes a change in the other, and by how much. For marketers, that's the move from guessing what works to measuring it.

The discipline grew out of statistics and economics, and researchers like Judea Pearl shaped its modern form [2]. At its core is one question: what would have happened if we had done nothing? That's the counterfactual, and estimating it is the critical step in any causal analysis [1]. If you ran a campaign and got 100 sales, you need to know how many you'd have had without it. The difference between the two is your incremental sales, and everything else is noise.

Three ideas do most of the work:

• Counterfactuals: the heart of causal thinking, comparing the world as it is with the world as it would have been. There's more in the guide to counterfactual analysis for ad spend.

• Treatment and Control: to measure the counterfactual, you need two groups: a treatment group that sees your marketing and a control group that doesn't. If the two were comparable to begin with, the difference in how they behave is the effect of your campaign.

• Directed Acyclic Graphs (DAGs): diagrams of which variables cause which. Drawing one before you analyze anything shows which confounders you need to control for, so you don't mistake correlation for cause. There's a practical guide to directed acyclic graphs in marketing.

The theory can get complex, but the arithmetic often isn't. One of the most basic formulas in causal inference is the uplift calculation:

Uplift = (Conversion_Rate_Treatment - Conversion_Rate_Control) / Conversion_Rate_Control

It gives the percentage increase in conversions your marketing caused, as long as the treatment and control groups were comparable to begin with. That's the shift to make: away from "attributed revenue" as the headline number and toward incremental sales. To put a euro figure on the difference, see what better attribution is worth.

Three more methods are common when you can't run a clean experiment:

• Instrumental Variables: find a variable, the instrument, that shifts the treatment but affects the outcome only through the treatment. The variation it creates works like a natural experiment and isolates the treatment's causal effect.

• Regression Discontinuity: for treatments assigned by a cutoff. If customers who spend over a set amount get a discount, regression discontinuity compares customers just above and just below the cutoff to estimate what the discount caused.

• Difference-in-Differences: compares the change over time in a treatment group with the change in a control group. It's the method behind a regional holdout: the region where the campaign changed, against the regions where it didn't, each against its own before-period.

Applied to your marketing channels, the same idea is called causal attribution. You don't need a data team to start; the first test can run on your own store this month.

Where to start, without buying anything:

1. Put the report in proportion. Add up the purchases Meta, Google and TikTok each claim for last month, and the total often comes out higher than your actual orders. Every platform grades its own homework, so the gap is credit counted more than once, not extra sales.
2. Start from your own data. List your channels by spend and by what last click gives them in GA4. The channel with the biggest budget and the least certain credit is the one to test first.
3. Run the holdout. Pause that channel in one region for two weeks, keep everything else the same, and compare that region's revenue with the regions where it kept running, each against its own two weeks before. That's difference-in-differences on your own sales, and a holdout on a weak channel is the usual first test.
4. Watch MER while you test. Total revenue over total ad spend doesn't depend on any attribution model, so no report can inflate it.

Once one holdout has run, the same logic answers the questions behind the numbers. Did the discount bring in new customers, or subsidize ones who were coming anyway? Did the influencer post reach new buyers, or your existing followers? Each is a counterfactual question, and each can be tested the same way.

Frequently Asked Questions

What is the difference between causal inference and A/B testing?

An A/B test is one form of causal inference: a randomized experiment, ideal for isolated changes like a button color or a subject line. Causal inference is the wider discipline, and it also covers the cases where you can't randomize, such as a multi-channel campaign, with methods like difference-in-differences or regional holdouts.

How does causal inference work without cookies?

By working on totals instead of individuals. A regional holdout compares total sales by region, and models of incrementality work on daily or weekly totals, so neither needs to follow a person across sites. That's why the approach holds up as third-party cookies disappear.

Is causal inference only for large companies?

No. A regional holdout works for any store that sells in more than one region, and it needs a spreadsheet, not a data team. Smaller stores may need to run it longer, because a few orders a day take longer to stand out from normal noise. Cutting spend that causal analysis shows wasn't adding sales is worth as much to a small store as to a big one.

What are the first steps to implementing causal inference?

Question one number, then test it. Pick the channel with the most budget at stake, pause it in one region for two weeks, and compare that region's revenue with the regions where it kept running, each against its own two weeks before. Software can scale the approach later; the first test needs none.

Can causal inference predict future sales?

Not directly. It's explanatory: it estimates what caused past sales. That makes forecasts better, because a forecast built on what actually drives sales holds up when you change the budget, but the forecasting itself is a separate step.

References

[1] Varian, H. (2016). Causal Inference in Economics and Marketing. Proceedings of the National Academy of Sciences, 113(27), 7310-7315. Available at: https://www.pnas.org/doi/10.1073/pnas.1510479113

[2] Pearl, J., & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.

[3] Frost, J. (2023). Correlation vs. Causation: Understand the Difference. Statistics By Jim.