A performance dashboard measures correlation. Your budget pays for cause. Here is the gap, and the one method that reveals it.

A rooster crows every morning, and every morning the sun rises. The correlation is perfect. The rooster causes nothing.

Performance dashboards report the rooster. They show that when a campaign spends, conversions appear beside it, and they call that the campaign's return. What they cannot show is whether the campaign caused those conversions or merely crowed next to them. The two readings split apart whenever a hidden factor drives both spend and sales, and in advertising that factor is almost always pre-existing demand. Branded search and retargeting look best precisely because they target people already about to buy, so their reported return is inflated by sales you would have won for free.

Judea Pearl's adjustment formula names that gap. On a retargeting line, a reported 6x return can hide a real one closer to 2x (illustrative). This piece walks through why the dashboard measures correlation rather than the causal effect of your budget, why sharper targeting makes the problem worse, and the single method that recovers the truth.

The dashboard measures correlation, not cause

This is the foundational problem of causal inference, stated by Judea Pearl: correlation does not identify cause whenever a third factor influences both. In a Google Ads account that third factor is usually demand. A person searching your brand name was already going to find you. A shopper you retarget had already added the item to a cart. The ad runs, the sale lands, and the platform records the sale next to the ad. The reported return looks excellent because the campaign is standing beside conversions it did not create.

Performance dashboards report the rooster, not the sunrise.

The confounder in almost every account is demand you did not create

Call the hidden factor intent: the pre-existing probability that a person buys regardless of your ad. Intent drives the outcome, and intent also drives where your spend lands, because both you and the platform aim money at people who already look ready to buy. When the same hidden factor moves both spend and sales, the relationship between them is inflated. Pearl's adjustment formula describes the fix in one move: estimate the effect as if intent were spread across the whole population the way it naturally occurs, not concentrated where the algorithm chose to spend.

A retargeting line shows an $8 cost per conversion and a 6x return. A geo holdout, where a random set of regions is withheld from the campaign, reveals that most of those conversions happened in the withheld regions too. Only 40 percent were incremental. The honest cost per conversion is closer to $20, and the honest return closer to 2x (illustrative). Nothing on the dashboard changed.

This is not a thought experiment. In 2015 eBay ran one of the largest controlled tests of its kind, reported by economists Steven Tadelis and colleagues, and found that paid search ads on the company's own brand terms produced almost no measurable new sales. The customers were coming anyway. A company spending heavily on the campaign with the cleanest-looking return was, on the causal reading, buying traffic it already owned.

The dashboard was never measuring the right thing.

Sharper targeting widens the gap

Here is the part that inverts intuition. The better a platform targets, the worse this bias gets. Smart Bidding earns its name by finding the users most likely to convert, which is the same population that would have converted without the ad. As the model improves, spend and pre-existing intent lock together more tightly, so the observed relationship drifts further from the causal one. The algorithm is manufacturing the exact correlation between confounder and treatment that the adjustment formula exists to remove.

The setting makes this worse. Major platforms control roughly 65 percent of global digital ad spend, and each one sets the auction price and grades the result. The system that decides what you pay also writes the report card, and that report card drifts toward the platform's incentive, not yours. So the campaigns your team is proudest of, the ones the platform rewards with more budget, are often the least incremental.

Optimization does not break the correlation trap. It tightens it.

Pearl's formula says subtract the confounder, randomization is how you do it

Pearl's adjustment formula is a recipe: to recover the causal effect, average the outcome across the natural spread of the confounder rather than the spread the algorithm produced. The catch is that you must be able to observe the confounder to subtract it. Intent is not in your data. You see device, geography, time of day, audience labels, none of which is the thing that actually decided the purchase.

When the decisive confounder is unobservable, no amount of statistical correction on observational data recovers the truth. The one operation that does is randomization. A randomly chosen holdout, withheld from exposure, breaks the link between intent and spend by force: the held-out group carries the same natural intent as the treated group, so the difference in their conversion rates is the incremental effect, with the confounders cancelled. This is the Analyze stage of the M.A.T.H. framework doing its real job, separating what your spend caused from what merely happened beside it.

Randomization is not free. A holdout sacrifices revenue from the suppressed group, and it needs enough volume to reach significance, which rules it out for small accounts. The working discipline is to run lift studies occasionally and use them to calibrate the cheap observational reads you look at daily, so a reported return becomes a number you discount by a known factor rather than trust at face value.

Where correlation is good enough

The strongest argument against treating the dashboard as fiction is that correlation is often a stable, useful proxy. If the bias is large but roughly constant, a reported 6x that is really 2x still ranks campaigns correctly, and for pacing and ranking decisions a biased-but-consistent signal can be enough. Incrementality testing also carries its own errors. Holdouts leak through other channels, a suppressed audience can still be reached by organic or email, and a noisy lift study can read a real effect as zero and kill a campaign that was working. Causal rigor is a cost, not a virtue in itself.

The case for it is specific. In advertising the bias is not constant, because the algorithm keeps changing how tightly spend tracks intent. Yesterday's discount factor is wrong today, which is why the correction has to be re-measured, not assumed once and trusted.

When this does not apply

The correlation trap binds when a hidden, decisive confounder drives both spend and outcome. Several cases fall outside it: cold prospecting into audiences with little pre-existing intent; new products and launches with no established demand; small accounts that cannot power a holdout; and cases with observable, stable confounders where adjustment on data works without an experiment (rare in paid media, where the decisive factor is intent).

What this changes in practice

The most dangerous line item in an account is rarely the one with poor numbers. It is the one with excellent numbers and low incrementality, because its reported return shields it from scrutiny while it absorbs budget that would compound elsewhere. The forensic move is to ask, for each campaign, how far its reported return would fall under a holdout, then aim the next experiment where that gap is widest. Found early, the budget recovered on a single mid-market account runs into six figures a year (illustrative). The skill is not better optimization. It is refusing to optimize against a number that measures correlation when the decision requires cause.

A performance dashboard measures how spend and conversions move together, not what your spend caused, and the two split apart whenever pre-existing demand drives both. Pearl's adjustment formula names the correction, but the decisive confounder in advertising is unobservable, so only randomization recovers the truth. The campaigns that look best are often the least incremental, and sharper targeting widens the gap. Measure cause with the occasional holdout, then discount the daily dashboard by what the holdout taught you.

Cite as: Ivitskiy, I. (2026). The M.A.T.H. Framework: A First-Principles Approach to Quant Marketing in High-Uncertainty Environments. SSRN. https://dx.doi.org/10.2139/ssrn.6258238 (open-access mirror: Zenodo, https://doi.org/10.5281/zenodo.18552246)

Originally published on LinkedIn. Igor Ivitskiy writes on Profit Forensics, the science of hidden profit in advertising data, at thedoctorads.com.