What changed · Measurement
Attribution was always estimation
Marketing spent a decade treating a modelled number as a measured one, then treated its decline as a crisis rather than a correction.
20 August 2026 · 6 minute read
Third-party cookies never measured attribution accurately. They produced a number that looked precise, covered a partial and biased slice of behaviour, and systematically credited whichever channel appeared last. The industry built budgets, incentives and careers on that number and is now describing its decline as a loss of measurement. It is better understood as the removal of a false precision the industry should never have trusted. The practical consequence is that measurement should return to methods that are honest about their uncertainty: holdouts, geographic tests and incrementality, all of which were available throughout and were skipped because they were slower.
What the number actually was
A cookie-based attribution figure was never a measurement of causation. It recorded a sequence of touchpoints it happened to be able to see, applied a rule about which one deserved credit, and returned a value to two decimal places.
Each of those steps loses information. It could not see mobile apps, blocked browsers or logged-in environments. It could not see the conversation that made someone consider you. It could not distinguish the ad that created demand from the one that harvested it. Then it credited the last one anyway.
Why the false precision mattered
Because budgets follow reported returns. Channels that report well get funded, and channels that report badly get cut, regardless of what either actually did. That systematically favoured harvesting over creation, because harvesting is easy to attribute and creation is not.
A decade of that produced a generation of brands very good at converting existing demand and very bad at creating any. It is a large part of why acquisition costs rose: a lot of brands were bidding against each other for the same finite pool of already-interested people.
The tell
If a channel reports a return that seems too good, it is usually taking credit for demand created elsewhere. Retail media and branded search are the two most common examples, and both are excellent channels that flatter themselves in reporting.
What honest measurement looks like
| Method | What it tells you | What it costs |
|---|---|---|
| Geographic holdout | Real incremental effect | Time, and giving up some spend |
| Time-based holdout | Whether a channel does anything | Nerve |
| Matched-market test | Comparative effect | Planning |
| Modelled attribution | A directional signal | Cheap, and easy to over-trust |
| Last-click | Which channel was nearest the finish | Nothing, and it is worth about that |
The first three were available for the whole cookie era and were mostly not used, because they take weeks and involve deliberately not spending money somewhere. The industry preferred a number it could have on Monday.
The part that should be liberating
Once you accept that measurement is estimation with error bars, several long-running arguments dissolve. You stop needing attribution to justify brand work. You stop cutting channels because they report poorly. You start asking what would have happened otherwise, which is the only question that was ever interesting.
It also lowers the cost of being wrong, because a range invites a test rather than a reorganisation.
We did not lose the ability to measure marketing. We lost a number that was confident, convenient and wrong, and the discomfort is mostly about the convenience.
If you are an agency
Clients ask for attribution certainty and the honest answer is a range with a method attached. Agencies that promised precision during the cookie era are now in a difficult position, and the ones that were candid about modelling are finding the conversation easy. Propose a holdout on the next campaign; it is the single most credible thing you can offer, and almost nobody does.
How we do this ourselves
We report what each connected source returns and join it to your own customer segments, which is more useful than a blended number. We do not claim a proprietary attribution model, because most of them are assumptions wearing a confident interface.
How we report performance