Why Advertisers Need More Than Attribution

Attribution answers a narrower question than most people think it does. It asks: of the conversions that happened, which touchpoint gets the credit?

The question your budget decision actually rests on is different: would this conversion have happened anyway? Attribution cannot answer that, and no amount of model sophistication changes it, because the answer requires observing a world where the ad did not run.

Why this gap costs real money

Retargeting is the clearest case. It reliably reports excellent ROAS, because it is shown to people already close to buying, and it gets credited for purchases that were largely already coming. Attribution is not wrong here. It faithfully reports that a retargeting ad was the last touch. It just cannot distinguish credit from causation.

The consequence is a well-known and expensive loop: scale retargeting, in-platform ROAS improves, total revenue does not move, and the account concludes it needs more retargeting.

The check takes a minute. Compare blended ROAS or MER, total revenue over total spend with no attribution involved, against your in-platform number over the same period. If platform ROAS improves while blended stays flat, you have moved credit around rather than generating sales.

Every attribution system disagrees, and they are all internally right

Meta reports one number, Google Analytics another, your CRM a third. This is not usually a bug.

They differ on the window: how long after a click or view a conversion still counts. They differ on the model: first, last, or distributed. They differ on whether view-throughs count at all. And they differ on whether modeled conversions are included, which matters more since signal loss made modelling a larger share of what platforms report.

Given different rules, different totals are the correct outcome. The useful response is to stop reconciling and choose: pick the system you make decisions with, understand what it is counting, and use the others as sanity checks rather than as contradictions to resolve.

What actually answers the causal question

Geo holdouts. Suppress advertising in a set of comparable regions, run normally elsewhere, compare. Geo holdout testing is the most practical incrementality method for most advertisers, because it needs no platform cooperation and measures what happened rather than arguing about credit.

Conversion lift studies. Meta’s own randomised holdout. Cleaner in design and dependent on sufficient volume: see conversion lift studies.

Marketing mix modelling. Top-down, channel-level, no user tracking required. Useful at larger spend and unhelpful for deciding what to do with one ad set: MMM.

All three cost something: suppressed revenue, time, or analytical capacity. That cost is the point. Causal answers are expensive because they require deliberately not doing something, and that is exactly why so few advertisers have one.

Attribution still earns its place

None of this makes it optional. Attribution is what lets Meta optimize at all: the delivery system needs conversion events tied to ads to learn from, which is why fixing your event pipeline is usually worth more than any model refinement. Accurate attribution also catches genuine breakage, and a wrong window will misattribute real performance in ways that are worth correcting.

The failure is treating attribution as a complete answer to whether spend is working. It is one layer of one part of the account, doing the job of measurement, not the job of deciding.

The practical position

Use attribution to run the day: which ads to test, which to refresh, where the delivery is going. Use a periodic incrementality read, even a rough geo holdout once or twice a year, to check whether the channel-level story attribution tells you is roughly true.

And keep the two questions separate in your head. “Which ad gets the credit” and “what would have happened without it” are different questions, and confusing them is how confident, well-reported, comprehensively wrong budget decisions get made.