Add up the revenue each advertising platform reports and the total routinely exceeds what the business actually took. That is not fraud and it is not a bug — each platform applies its own attribution model over its own window to its own view of the customer journey, and a sale touched by three channels is claimed by all three.
Understanding which model produced a number is the difference between reading a report and being managed by one.
What is an attribution window?
The period after an ad interaction during which a conversion is credited to it, and platforms differ on both the length and what counts as an interaction.
| Setting | Typical default | Effect |
|---|---|---|
| Click-through window | 7 to 30 days | Longer credits more sales |
| View-through window | 1 to 7 days | Credits sales with no click at all |
| Model | Last click, or data-driven | Decides who gets the credit |
View-through attribution is where the double counting becomes largest. A platform claiming a sale because the ad was displayed — not clicked — will credit conversions that another platform also claims for a click, and neither is applying an unreasonable rule.
Changing a window changes the reported ROAS without changing a single thing about the campaign, which is worth remembering when a number moves after a settings change.
How do the models differ?
By where they place the credit along the journey. Last-click gives everything to the final interaction, which systematically undervalues anything that introduces a customer. First-click does the opposite. Linear splits evenly, and data-driven models assign weights from observed patterns.
None of them is correct, because attribution is not a measurable property of a purchase — it is an accounting convention applied to a decision that happened in someone’s head. The useful question is whether the convention is consistent enough to compare periods.
What can you actually trust?
MER, and incrementality tests. Total revenue divided by total marketing spend cannot be double counted, because it has no model inside it — the ROAS article covers why it became the figure operators fall back on.
A holdout test is the stronger tool and the harder sell: turn a channel off in a matched region or audience and measure what happens to total revenue. It answers the question attribution only approximates — what would have happened anyway.
The answer is frequently uncomfortable. Branded search and retargeting both attribute well and test poorly, because a large share of the people they claim were going to arrive regardless.
Why did tracking get worse?
Because the identifiers the models relied on stopped being available. Third-party cookie restrictions, mobile tracking prompts and privacy features that strip parameters all reduce the share of journeys a platform can observe end to end.
Platforms responded by modelling the gap rather than reporting the shortfall, which means a growing share of reported conversions are statistical estimates rather than observed events. That is a defensible engineering choice and it makes the numbers less literal than they look.
It also broke the arithmetic quietly rather than visibly. A report that says 4.2× is not flagged as partly modelled, and the share that is modelled is not disclosed.
What is a server-side conversion feed?
A way of reporting conversions from your own systems rather than from the browser, which restores some of what the tracking restrictions removed. The platform receives the event from your server, matched to the ad interaction by whatever identifiers are available.
It improves the platform’s view and it does not resolve the credit question — three platforms receiving the same server-side event will still each claim it. What it does fix is the shortfall where a conversion was simply never observed, which is the difference between under-reporting and over-crediting.
What should a report actually contain?
Three numbers rather than one, so the disagreement between them is visible.
- Platform-reported ROAS per channel, for steering bids.
- MER for the whole business, for judging whether marketing is working.
- New-customer share, so retargeting cannot flatter the blend.
When the first rises and the second does not, the channels are competing for credit rather than creating demand. That divergence is the most useful signal in the whole set and it is invisible if only one of the three is tracked.
Questions people ask
Should I use a shorter window? A shorter window reports less and is more conservative, which makes it a reasonable default when the alternative is over-crediting. Consistency matters more than the specific length.
Is last-click really that bad? It is simple, stable and systematically biased toward the bottom of the funnel. As a decision rule for scaling top-of-funnel spend it is actively misleading.
Does UTM tagging fix this? It fixes your view — which channel the visit came from — rather than the credit question. The medium field article covers what it does and does not answer.
How big is the over-count? Varies enormously by mix. Comparing the sum of platform-reported revenue against the accounting figure gives your own number, and it is worth doing once a quarter.
Every platform is answering the question it can see; MER is the one that reconciles. The MER calculator gives the blended figure, the ROAS and CPA calculators cover the channel view, and the conversion rate calculator is the part of the chain no attribution model can distort.