Attribution After Cookie Loss: What Still Works
Attribution has quietly stopped meaning what most reports imply it means. Tracking prevention, app-level restrictions, shorter cookie windows and cross-device journeys have all degraded the signal. The reports still produce confident-looking numbers, which is the dangerous part.
What broke, specifically
The assumption underneath last-click attribution is that you can observe a continuous journey and identify the final touch. Today you frequently observe fragments: a session that ends, a return visit that looks like a new user, a conversion with no visible prior interaction.
The predictable consequence is that channels which happen to sit near the end of the journey — branded search, direct, email — absorb credit for demand that something else created, while upper-funnel activity looks worse than it is. Budget decisions made on those reports systematically underfund the things that actually generate demand.
What still works
First-party data, handled properly
Logged-in sessions, email identification at checkout, and order history are durable in a way that third-party signal is not. A customer who can be recognised across visits restores much of what was lost — which makes account creation and email capture measurement infrastructure, not just marketing tactics.
Server-side event collection
Sending conversion events from your server rather than the browser avoids a whole category of client-side loss. It does not solve identity, and it is not a way around consent — it is a way to stop losing events you legitimately have.
Incrementality testing
Turning a channel off in a region, or holding out a cohort, answers the question attribution was always trying to answer: what would have happened anyway. It is slower and less granular than a dashboard, and it is the only approach that survives signal loss intact.
Modelled attribution, understood as modelled
Platform models fill gaps with estimates. That is reasonable and often the best available answer — provided everyone reading the number knows it is partly inferred. Problems start when a modelled figure is treated as a count.
How to report it honestly
Stop reconciling platform numbers to each other. Each platform counts its own contribution generously and they will never agree; time spent forcing agreement produces no insight.
Pick one internal source of truth for revenue — your order data — and use channel reporting for direction rather than precision. Then judge spend on blended efficiency alongside periodic incrementality checks.
State the uncertainty in the report itself. A decision made knowing the range is a better decision than one made on a false point estimate, and it protects the analysis when the number inevitably shifts.