Break the intuition first: the dashboard doesn't know which ad worked. Platform-reported conversions answer 'was our ad somewhere on the path?' — and paths are crowded, so every platform finds itself present and claims the sale. Summed platform claims routinely exceed actual revenue several times over. They can't all be right, and structurally none of them is, because presence isn't influence. The question a budget actually needs answered is counterfactual: how many of these sales would have happened without the spend? That question has a name — incrementality — and no attribution model, however sophisticated, answers it, because models allocate observed conversions. They never observe the world where the ad didn't run.
Understand the models as opinions with known biases. Last-click credits the final touch: it systematically flatters harvest channels (branded search, retargeting, cashback affiliates) that intercept buyers who've already decided, and starves the creation channels (video, content, upper-funnel social) that made them decide — optimising to it slowly reallocates budget from farming to gleaning. First-touch inverts that bias. Linear, time-decay, and position-based split the credit by arbitrary formula — different opinions, same epistemic status. Data-driven attribution (Shapley values, Markov removal effects computed over observed paths) is genuinely better bookkeeping: it at least asks how conversion probability changes when a touchpoint appears. But it stays bookkeeping of the tracked world — offline influence, dark social, and TV don't emit clicks — and it still can't see the counterfactual. Changing attribution models reallocates credit, and therefore budget, without a single fact about causation changing. That sentence should be laminated above every dashboard.
Incrementality is the counterfactual discipline, and it's older than the internet: hold something back, compare, believe the difference. The instruments, in rising order of ambition: conversion-lift studies (platform-run randomised holdouts — users eligible for your ad but shown none — measuring the true delta); geo-experiments (matched cities or states dark versus exposed — the workhorse, because it needs no user tracking at all and survives every privacy change); and the bravest, cheapest instrument in marketing, which is simply turning it off — regionally, briefly, on purpose. The classics earned their fame: eBay's economists found paid clicks on brand terms were overwhelmingly harvest (the buyers arrived anyway via the organic link one centimetre below), and P&G's public digital cuts — around two hundred million dollars — reportedly changed sales not at all. Neither result proves your brand search or your digital is waste. Both prove the only way to know is to test the counterfactual rather than admire the click.
The working stack triangulates three instruments at three altitudes. Marketing-mix modelling — econometrics regressing sales on spend, seasonality, pricing, and competition — is back from the 1960s because it needs no user-level tracking: it sees TV, print, and trade alongside digital, estimates diminishing-returns curves per channel, and answers the quarterly allocation question. Its price is assumptions, data hunger, and lag (and the majors now open-source their tooling, which says a lot about where measurement is heading). Experiments are the truth layer: run a handful of geo-tests a year on your biggest, most contested lines, and use the readouts to calibrate the MMM — an MMM anchored to experimental lift is twice the instrument. Attribution survives at the tactical altitude, where its bias is roughly constant and its speed is the point: comparing creatives, audiences, and bids within one channel, daily. One instrument for strategy, one for truth, one for steering — the failure mode is using any of them at another's altitude.