Causal Inference in Advertising
Causal inference in advertising is the discipline of identifying which specific change caused a metric to move — as opposed to correlation, which only shows that two things moved together. It operates at two levels: channel-level (incrementality, MMM) and operational (which account change moved which metric).
Updated July 2026.
Causal inference in advertising means identifying which specific change caused a metric to move, rather than noting that two things moved together. Correlation says "CPA rose while the new creative launched." Causal inference says whether the creative did it — or whether it was a bid change, a learning-phase reset, an auction shift, or a platform update.
Two levels of causality
| Level | Question it answers | Methods |
|---|---|---|
| Channel-level | "How much revenue does this channel actually cause?" | Incrementality testing (holdouts, geo-tests), marketing mix modeling |
| Operational | "Which change in the account moved this metric?" | Change records + baseline modeling + attribution of shifts to specific events |
Channel-level causality is having a moment: 60% of marketers now trust incrementality testing most, versus roughly 40% for MMM and 37% for in-platform attribution (Haus survey via eMarketer, January 2026, N=500; see the incrementality era). Operational causality is the newer discipline — and the one that answers the day-to-day question "why did my numbers move?"
Why correlation fails in ad accounts
Ad accounts change constantly — team edits, automated rules, platform automation, auction dynamics, and platform rule changes (Meta's March 3, 2026 attribution redefinition moved measured CPA for every advertiser with zero real-world change). Many things move together every week; without a method for isolating cause, teams optimize noise (see statistical significance in paid media).
Related
The Ad Spend applies operational causal inference: every change is recorded, detection runs against the account's own baseline roughly every three hours, and metric moves get the likely cause attached — see Signal and how this differs from asking an AI assistant.