What attribution stopped measuring in 2024, and what to do about it.
The web of cookie deprecations, privacy regulations, and platform-level signal restrictions that landed between 2022 and 2024 produced, by the end of the period, a measurement environment substantially different from the one most marketing teams were trained on. Where we stand now, and what the teams who have adapted are doing.
The story of how digital marketing attribution stopped working as advertised has been told often enough that I am not going to repeat the timeline in detail. The short version: between roughly 2022 and 2024, a combination of browser-level cookie deprecations, regulatory restrictions on cross-site tracking, platform-level changes to advertising identifiers on both major mobile operating systems, and the deprecation of several major third-party measurement APIs, produced a measurement environment in which the standard attribution models — multi-touch, last-click, view-through — were no longer measuring what they had been measuring previously.
The marketing industry, broadly speaking, responded to this in three phases. Phase one, in 2022 and 2023, was denial: most teams continued to report on the same metrics, using the same models, while the underlying data quality quietly degraded. Phase two, in 2024 and early 2025, was panic: a substantial fraction of teams discovered, often abruptly, that the numbers they had been reporting were no longer reliable, and that several of the channels they had been crediting with their growth had been substantially overstated. Phase three, from late 2025 onwards, has been adjustment: the teams who survived phase two have, in many cases, started to rebuild their measurement practice from a different starting position.
I want to spend this essay on phase three. The teams that have adapted well have, in our reading, converged on a set of practices that are noticeably different from the pre-2024 norms. The practices are not, individually, novel. Several of them have been recommended by sophisticated marketing analysts for years. What is new is that they are now being adopted by mainstream marketing teams who, until two years ago, would have considered them too imprecise to bother with.
The first practice: incrementality testing, not attribution modelling
The standard pre-2024 measurement approach was attribution modelling: the use of statistical models to assign credit for a conversion to the various marketing touches that the user had encountered before converting. The models varied in sophistication — from simple last-click models to elaborate multi-touch attribution systems — but they shared a common underlying assumption: that the user-level data on which the models were built was reasonably complete and reasonably accurate.
That assumption no longer holds. The user-level data on which attribution models depend is, in 2026, substantially incomplete in most cases. The models, applied to incomplete data, produce results that are confidently wrong.
The teams that have adapted have, in most cases, abandoned attribution modelling as the primary measurement framework and replaced it with incrementality testing. The basic idea of incrementality testing is to compare the conversion rate of a group that was exposed to a particular marketing intervention with the conversion rate of a group that was not, and to attribute the difference between the two to the intervention. This is, in measurement terms, considerably more robust than attribution modelling. It does not require complete user-level tracking. It only requires that the two groups be comparable in expectation.
Incrementality testing is harder than attribution modelling in some specific ways — it requires the discipline to hold out a control group, which many teams find culturally difficult, and it produces fewer dimensions of analysis per test. It is also, on every measure of statistical robustness that matters, a substantially better approach. The teams that have made the switch are getting better information from their tests than they were from their attribution models.
The second practice: marketing mix modelling, at small scale
Marketing mix modelling — the use of econometric techniques to estimate the contribution of different marketing channels to overall business outcomes — was, until recently, a technique used almost exclusively by very large advertisers. The technique requires substantial historical data, some statistical expertise, and a willingness to work with point estimates rather than precise attributions. It was, in the pre-2024 environment, considered too coarse for most digital marketing teams.
This has changed. The marketing mix model, with its more aggregate inputs and its more aggregate outputs, has turned out to be remarkably robust to the loss of user-level data. The model does not depend on tracking individual users through a conversion funnel. It depends on understanding, at the aggregate level, how variations in marketing spend over time correlate with variations in business outcomes over time. This kind of analysis is now broadly accessible — the open-source tooling has matured considerably, the cost of running the relevant statistical work has fallen, and a number of smaller agencies now offer this kind of analysis to teams that would not have considered it three years ago.
"The death of user-level attribution turned out to be the rebirth of aggregate-level measurement. We are, in many ways, going back to the measurement practices of the pre-digital era — and, in many ways, getting better answers as a result."
The third practice: the post-purchase survey
The third practice is, in some ways, the most surprising. Several of the teams I have spoken to have, in the past eighteen months, added a question to their post-purchase or post-signup flow that asks the user, in plain language, how they heard about the product. The question has, traditionally, been considered too imprecise to be useful: users do not, in many cases, accurately remember how they first encountered a brand, and the data collected by self-reporting is widely understood to be unreliable.
The teams using this approach in 2026 are, in our observation, finding it more useful than they had expected. The reason is that the self-reported data — while individually unreliable — is, in aggregate, often the only source of information available about the channels that the algorithmic attribution systems no longer see. A user who first encountered the brand through a podcast appearance, an unprompted recommendation from a colleague, or a long-form article they read three months earlier, will sometimes report exactly that — and the report, however imprecise, is information about a channel that the rest of the measurement stack is completely blind to.
The teams using post-purchase surveys well are, in most cases, combining the self-reported data with the incrementality and mix-modelling work I described above. The self-reported data does not, on its own, support precise decisions. It does, however, often surface channels that the rest of the analysis would miss entirely, and it provides a useful sanity check on the conclusions that the more rigorous methods produce.
The fourth practice: acceptance of irreducible uncertainty
The fourth and most important practice is, in a sense, not a practice at all. It is a cultural shift in how the marketing function relates to measurement. The pre-2024 norm was that the marketing team was expected to know, with reasonable precision, the return on each marketing pound spent. The norm was unrealistic even before the measurement environment changed; it is, in 2026, simply not achievable on the data that is now available.
The teams that have adapted have, in most cases, made peace with this. They report ranges rather than point estimates. They explicitly acknowledge the uncertainty in their numbers when presenting to the rest of the company. They argue, where necessary, for the strategic value of activities whose attribution they cannot prove. The cultural shift is harder than any of the technical changes. It requires the rest of the company — particularly the finance function and the executive team — to accept that the marketing dashboard is not what it used to be, and that better decisions will result from honest uncertainty than from precise-looking estimates that the underlying data does not support.
What I would say
I would say that the post-2024 measurement environment, while harder to work in than the pre-2024 environment, is in several important ways better. The pre-2024 environment encouraged a kind of false precision that, in retrospect, led many companies to make worse decisions than they would have made with less data. The post-2024 environment forces a kind of intellectual honesty about what we know and what we do not that, in my view, the industry was overdue for.
The teams that have adapted to the new environment are, in our observation, making better strategic decisions than they were three years ago — even though the dashboards they show their executives are, by every measure of precision, considerably less impressive. That is, in the end, the headline of the story: the dashboards got worse, the decisions got better, and the relationship between the two was always more complicated than the dashboards admitted.