Causal Inference for Marketing Teams: A Non-Technical Guide

Causal inference is the discipline of answering one question honestly: what would have happened without the thing you did. Every attribution report answers a different question, which channel touched the sale, and the two are routinely confused in marketing meetings. This guide explains the counterfactual, why observational numbers get it wrong, what an honest design looks like, and the four things a causal number has to state, with no equations.

The question, and the one it is usually confused with

Meta says a campaign produced 400 purchases. That sentence means: 400 people who bought had seen or clicked the campaign inside Meta's window. It does not mean 400 fewer people would have bought without it. The second statement is the counterfactual, and no platform report produces one, because the platform can only see the world in which the ad ran.

The Price of Being Found describes the reconciliation problem as a category error. The ad platform reports conversions it can associate with its own inventory inside its own window. Analytics reports sessions it could resolve to a source. The store reports orders, the only fact about money. None of the three answers "what would have happened otherwise".

Why a bigger model does not fix it

The trap has a name: selection. Platforms show ads to the people most likely to buy. Compare buyers who saw the ad with people who did not and you are comparing people the platform chose against people it skipped. The naive comparison fits the data beautifully and measures the platform's targeting, not the ad's effect.

The evidence is experimental. Gordon, Zettelmeyer, Bhargava and Chapsky, in Marketing Science in 2019, ran standard observational attribution methods against 15 randomised experiments at Facebook. In half the studies the estimated lift was off by a factor of three across all methods, and six of the fourteen checkout studies could not detect a significant lift at all. Their worked example: true lift 72.8%, naive comparison 316%. The book's gloss is the useful one: the better a model fits observed journeys, the more faithfully it may be reproducing the targeting rather than the effect. Data-driven attribution vs causal attribution covers the GA4 version of the same trap.

The two honest designs

There are exactly two ways to observe the counterfactual, and everything else is a dressed-up version of one of them.

The individual holdout. Randomly withhold the ad from a slice of the audience and compare purchase rates. The gold standard, and what the Facebook experiments used. Its power depends on conversion rate and sample size, and the numbers are unforgiving: fifty thousand users a month at a 1% conversion rate can detect a 25% lift and nothing smaller.

The geographic holdout. Switch the channel off in some regions, model what would have happened from the pre-period, compare. This is what most brands can actually run. Its power depends on how much regional revenue wobbles week to week and how long the test runs.

Both designs have a minimum detectable effect: the smallest true lift they could distinguish from nothing. Below it, "no significant difference" means "we learned nothing", not "the channel does nothing". Almost everyone reads it the second way. Incrementality testing for ecommerce is the step-by-step version.

What sits between experiments

Experiments answer one channel at a time and take weeks. Between them, brands run on models, and the honest models say so. Causality Engine uses causal inference on the GA4 history you already have to estimate what each channel caused.

The four properties of a number you can defend

The book gives a defensible number four properties, and they are the whole of causal inference translated for a budget meeting:

1. It comes from a source that does not sell you media.
2. It states its coverage: revenue of X, of which the systems can attribute Y.
3. It states its design: experimental, quasi-experimental, or observational, and if observational, that it is a description rather than a causal claim.
4. It states its interval and its floor: "12%, interval 4 to 20, from a design whose minimum detectable effect was 8.3%".

Most arguments in marketing meetings, the book observes, are two people comparing an experimental number to an observational one without either noticing. Property three ends those.

What to do this week

• If you own the budget: ask, for every number in the plan, which of the three systems it came from and whether it has property three. Most will not.
• If you have to defend the number: rewrite your best channel's line in the four-property form. If you cannot fill the interval or the floor, that is the finding.

As of 9 September 2026. The experimental findings and the four properties are quoted as The Price of Being Found (Edition 2.10) presents them in Chapters 9, 15 and 19, with the book's scope notes.