A customer sees an ad on Instagram and ignores it. Days later they search for the brand on Google, click a search ad, but don't buy. A week after that they get an email, return to the site and finally purchase. Simple question, hard answer: which of these touches deserves the credit for the sale? The answer you choose decides where your next marketing euro goes — which is why attribution is one of the most consequential, and most poorly resolved, decisions in digital marketing.
For years, the default answer was simple and wrong: give all the credit to the last click. It is convenient, it is what most tools show by default, and it systematically overvalues the channels that appear at the end of the journey at the expense of those that create demand at the start. Choosing the wrong attribution model is not a technical detail; it is redistributing budget based on a distorted snapshot of reality.
This article compares the main approaches to attribution — from last-click to data-driven models, by way of Marketing Mix Modeling and incrementality testing — with the pros, the cons and the 2026 context, in which measurement has become harder for privacy reasons. The goal is not to crown a single winner, but to see that each method answers a different question.
What attribution is and why it is so hard
Attribution is the process of distributing the credit for a conversion across the various marketing touches that preceded it. It looks like accounting, but it is really an attempt to answer a causal question: how much of this outcome is due to each action we took? And causal questions are notoriously treacherous.
The difficulty has three roots. First, real journeys are long and multichannel, often crossing devices and weeks. Second, correlation is not causation: a channel can always appear before the purchase without having caused it — people who were already going to buy also search for the brand. Third, much of marketing's effect is indirect and delayed: an awareness campaign today can generate sales months from now that no click model can trace back to their origin.
Last-click: the default model that misleads
The last-click model assigns 100% of the credit to the final touch before conversion. Its only virtue is simplicity: it is easy to calculate, easy to explain and demands no sophisticated data. It was for a long time the de facto standard of web analytics tools.
The problem is that it rewards whoever arrives last and ignores whoever prepared the ground. Bottom-of-funnel channels — brand search, retargeting, email — receive disproportionate credit, while the channels that generate demand at the top — display, video, social, awareness — look ineffective and see their budget cut. The typical result is a spiral: you cut the top of the funnel, demand dries up months later, and no one understands why, because the model never saw that link. Last-click is not useless, but using it as the only lens is one of the most common ways to make bad budget decisions with apparently objective data.
Rule-based models: first-click, linear, time-decay
Between last-click and the more advanced methods there was a generation of rule-based models that tried to share the credit more fairly. First-click gives everything to the first touch; linear splits the credit equally across all points; time-decay gives more weight to touches closer to conversion; position-based reinforces the first and the last. They all share the same underlying limitation: the split is arbitrary, decided by a fixed rule and not by the data.
It is worth noting a relevant practical change: in 2023, Google discontinued the rule-based models (first-click, linear, time-decay and position-based) in GA4 and Google Ads, leaving essentially data-driven attribution and last-click. Many teams that still think in terms of these models are, in practice, working with tools that have already abandoned them.
Data-driven attribution: letting the data assign the credit
Data-driven attribution (DDA) replaces the fixed rule with an algorithm that learns, from the observed journeys, how much each channel actually contributes. Techniques such as Shapley values or Markov chains compare the paths that convert with those that do not and estimate the marginal weight of each touch. In theory, it is the fairest answer within what is observable.
It has, however, three important limits. It needs enough data volume to be reliable — in small accounts, it is noise. It is specific to each platform and often a black box: Google attributes its way, Meta its own, and each tends to pull the credit toward itself, which produces double counting when you add up reports from different channels. And, crucially, it remains limited to what it can observe at the user level — precisely what privacy is making scarcer.
Marketing Mix Modeling: the top-down view
Marketing Mix Modeling (MMM) approaches the problem from the opposite side. Instead of following individual users, it uses aggregate data — spend by channel, sales, price, seasonality, promotions, external factors such as weather or the economy — and, through statistical models, estimates how much each channel contributed to total sales. It needs no cookies and no identifying of people, which makes it naturally resistant to privacy restrictions.
Its strengths are exactly the weaknesses of click models: it captures the effect of offline channels, awareness campaigns and media where there is no click to measure; it incorporates delayed and long-term impact; and it gives a whole-picture view that is useful for allocating budget across large blocks of investment. In return, it is a low-granularity approach — it answers "how much to invest in video versus search" well, and "which specific ad to optimise today" badly — and it demands a long history and methodological care to separate correlation from causation.
MMM is no longer the exclusive preserve of large advertisers with expensive consultancies. In February 2025, Google open-sourced Meridian, the successor to LightweightMMM, with a Bayesian approach and geo-level modelling; Meta maintains Robyn, which automates parameter optimisation. These open-source tools have lowered the barrier to entry, even if they still demand quality data and statistical knowledge to deliver reliable results.
Incrementality: the only question that matters
Behind all attribution lies a question that no correlational model fully answers: would this sale have happened anyway, without the ad? This is called incrementality, and the most robust way to measure it is not to model the past but to run experiments. A geo test (geo-lift) switches spend off in some regions and keeps it on in others, then compares the results; a holdout keeps a control group that is not exposed to the campaign.
These tests are the gold standard because they create a real counterfactual — an observed "what if we had done nothing", not an estimated one. They are harder to set up and sometimes mean giving up some reach to hold a control group, but they answer the causal question that attribution only approximates. A mature practice combines attribution models for the day-to-day with periodic incrementality tests to calibrate what the models say.
Privacy in 2026: less signal, not fewer cookies
There is a persistent narrative that "cookies are going away". It pays to be precise. In 2024, Google reversed its intention to remove third-party cookies from Chrome and, in October 2025, announced the winding down of much of the Privacy Sandbox technologies, keeping only a reduced set of features. In other words: third-party cookies still exist in Chrome.
That does not mean the measurement problem has gone away — quite the opposite. User-level signal keeps degrading through other routes: Safari and Firefox have blocked third-party cookies for years, iOS limited cross-app tracking with App Tracking Transparency, and the consent required by GDPR reduces the fraction of observable users. The practical effect is that attribution based on individual tracking sees less and less of the real journey. It is that context — and not the end of cookies as such — that explains the return of MMM and incrementality testing, methods that do not depend on identifying people.
How to choose: a layered approach
The question "which is the best attribution model" is badly posed, because each method answers a different question and operates at a different scale. The most mature organisations do not choose one; they combine them in layers:
- MMM for the high-level strategic decision: how to split the budget across major channels over quarters.
- Data-driven attribution for day-to-day tactical optimisation within digital channels, aware of its limits and of double counting.
- Incrementality tests as the referee: when MMM and attribution disagree, a well-designed experiment says who is right.
The rule of thumb is to use each tool for what it is good at and to distrust any number that comes from a single source. When three independent methods point in the same direction, you can act with confidence; when they diverge, that divergence is itself valuable information about where your data is weak.
Mini case: reconciling two stories
Consider an e-commerce company that ran its campaigns almost entirely on last-click. Through that lens, brand search and retargeting were the heroes, and spend on video and social looked like waste — so it was gradually cut. Over two quarters, "direct response" sales held up, but total sales kept slipping, with no apparent explanation in the reports.
The team built a simple MMM with two years of history. The model suggested that the top-of-funnel channels, apparently ineffective under last-click, contributed to a meaningful share of demand that only materialised weeks later in other channels. Rather than decide on a model alone, they ran a geo test: they reactivated video in half the regions and kept it off in the other half.
After six weeks, the regions with video showed total sales visibly higher than the control group, confirming what the MMM indicated and last-click could never see. The company rebalanced budget toward the top of the funnel gradually, guided by experiments rather than intuition. The point is not that video is always good — it is that no single lens showed the full picture.
In practice
Attribution is not a problem you solve by picking the right model from a list; it is a discipline of triangulation. Last-click remains useful as a quick operational signal, as long as you know it overvalues the end of the journey. Data-driven attribution sharpens optimisation within digital channels. MMM gives back the whole-picture view and resists privacy restrictions. And incrementality tests are the only method that directly answers the causal question everyone cares about.
In 2026, with individual signal ever scarcer but cookies still present in Chrome, the recommendation is not to bet everything on one technology, but to build a measurement system that cross-checks approaches and treats each number with the healthy scepticism of someone who knows how it was produced. Anyone allocating budget on the basis of a single source will keep rewarding the channels that appear last — and keep wondering, quarter after quarter, why demand fails to show up.
