Attribution that survives a privacy-first web
Last-click reporting is now measuring a shrinking slice of reality. Modelled attribution plus incrementality testing is what replaces it.
Key takeaways
- Agree the attribution model before spend starts, not when the results are argued about.
- Incrementality tests answer the budget question that any attribution model only estimates.
- Server-side conversion capture is now table stakes for paid channels.
How should marketing attribution work now?
Combine three layers: server-side conversion capture for accuracy, a modelled attribution view for day-to-day allocation, and periodic incrementality tests to validate the model. No single number should decide a budget on its own.
What actually changed
Consent rates, cross-device journeys and platform reporting windows have each removed a slice of observable data. What remains is real, but it is a sample, and the sample is biased towards the channels that happen to be observable.
Reporting on that sample as though it were the whole picture reliably shifts budget towards branded search and away from everything that made anyone search for the brand.
What each layer is for
| Layer | Question it answers | Cadence |
|---|---|---|
| Server-side capture | Did the conversion happen at all | Continuous |
| Modelled attribution | How should credit be split today | Weekly |
| Incrementality tests | Did this spend cause the outcome | Quarterly |
Run the holdout before the argument
Incrementality testing is unpopular because it means deliberately not spending in a region or audience for a few weeks. It is also the only method that settles the question.
On one account, a channel that last-click credited with 34% of revenue tested at 11% incremental. Reallocating that gap funded an entire second channel without increasing total spend.
Agree the model in writing first
The attribution conversation is easy before spend and impossible after results. We write the model into the engagement: which conversions count, what lookback applies, how modelled and observed conversions are labelled in the report.
It removes the quarterly ritual where two dashboards disagree and everyone picks the one that supports their channel.
Report modelled and observed conversions separately
The fastest way to lose trust in an attribution rebuild is to hand stakeholders a single number that is quietly part measurement and part estimate. The first time someone reconciles it against a platform dashboard, the whole model becomes suspect.
We label the columns. Observed conversions are the ones captured server-side with a full path. Modelled conversions are estimated, with the method and the confidence stated. Incrementality results are reported separately again, with the test window and the holdout size.
It makes the report longer and it makes it survivable. When a channel owner disputes a number, the conversation is about which layer they are disputing, and that is a question with an answer. A single blended figure only ever produces an argument about whose dashboard is right.
FAQFAQ
Frequently asked questions
About the author
Aisha runs acquisition across search, paid and lifecycle. She writes about the work that connects a channel report to pipeline — crawl and indexation, topical authority, incrementality testing and attribution that survives a privacy-first web.
- Technical SEO
- Topical authority
- Performance media
- Marketing attribution
