Attribution

A customer sees a Meta ad on Monday, searches the brand on Wednesday, clicks a Google ad, and buys. Meta knows about the impression and not the search. Google knows the search and not the impression. Each reports one conversion, because from where each is standing, one conversion is exactly what it caused.

Add them together and you have two conversions and one order. Do that across four channels and the reported total can approach double what the business actually took, which is the single most common way an agency report loses credibility.

This is not a defect anybody can fix. It follows from each platform measuring its own surface with its own data, and there is no shared identifier across them that privacy rules would permit. The mistake is not the over-claim, which is structural. The mistake is presenting the sum as though it were a total.

A conversion window is how long after an ad interaction a purchase still gets credited to it. Widen it and attributed conversions rise; narrow it and they fall. Nothing about the advertising changed, so any comparison across a window change is measuring the measurement.

View-through attribution is the sharper version. It credits a conversion to somebody who saw an ad and did not click it. There are defensible arguments for counting some of that, since advertising does work without clicks, and there is no defensible way to distinguish an impression that influenced a purchase from one that was scrolled past two days earlier.

View-through is the largest single contributor to the gap between platform-reported revenue and the client order total, and it is usually switched on by default. Knowing whether the figures you are reporting include it is a basic check that a surprising number of accounts fail.

Safari has restricted third-party cookies through Intelligent Tracking Prevention since 2017, and has progressively shortened the life of first-party cookies set by script. Firefox followed. Chrome has taken a longer and more contested route, but the direction across the industry has been one way.

Apple App Tracking Transparency in iOS 14.5, in 2021, made cross-app tracking opt-in, and most people declined. That removed a large share of the signal Meta had been using to attribute mobile conversions, which is the origin of the measurement disruption advertisers noticed that year.

In Europe, consent requirements under GDPR and ePrivacy mean a share of traffic is never measured at all, because the visitor declined and the tags never fired. That share varies by country and by how the consent banner is designed.

Platforms have filled the gaps with modelling: inferring conversions they cannot observe from patterns in the ones they can. The modelling is often reasonable. It is also invisible in the interface, so a meaningful and growing part of any modern attribution report is estimated rather than counted, and the report does not say which part.

It is a reasonable tool for comparing things measured the same way. Campaign A against campaign B inside one platform, this month against last month with the model unchanged, one audience against another. Those comparisons hold because the distortion applies equally on both sides.

It is a poor tool for stating what the advertising earned, because the number includes conversions the advertising did not cause and excludes ones it did. For that question the instrument is an incrementality test, which measures against a holdout rather than inferring from paths.

The practical division most agencies land on is to use platform attribution for day-to-day optimisation, where relative accuracy is enough, and to reconcile against the client own sales data for anything commercial. Both belong in the report, side by side, with the gap explained rather than hidden.

The conversation goes badly when the client discovers the over-claim on their own, because the discovery frames it as something that was being concealed. It goes fine when it appeared in the first report as a normal feature of how this works.

Three things make it routine. Show platform-attributed conversions next to the client own order count from the beginning. Label the model and the window on the report. Never sum across platforms, and say why not when somebody asks for a single total.

It is also worth saying plainly that the platforms are not being dishonest. They are each answering a narrower question than the one the client is asking, and the arithmetic of adding those answers together is what produces the nonsense.