Attribution After the Cookie's Death

Attribution was never truly solved — but it was manageable for a long time. As long as third-party cookies worked universally and users browsed primarily on a single device, last-click and first-click at least provided orientation: crude, distorted, but reproducible. That comfort is gone.

Google Chrome has progressively restricted third-party cookie support and positioned the Privacy Sandbox as its successor. Safari and Firefox have blocked them by default for years. iOS 14.5 radically changed the mobile measurement landscape with App Tracking Transparency. What remains are measurement pathways with structural gaps — and attribution models that behave as though those gaps don't exist.

The real problem isn't the missing cookie. The problem is that first-click and last-click rest on an assumption that was already wrong before cookies died: that a single touchpoint is causally responsible for a conversion. In a typical B2B buying process in the DACH mid-market, the time between first touchpoint and deal close is often 30 to 90 days, involving seven to twelve channel contacts and at least two to three key decision-makers. Last-click attributes all of that to the final Google Ad. First-click attributes it to the first organic result. Both models lie in their own way.

Before discussing attribution models, the data foundation must be sound. The most common mistake I see in audits: companies attempting to apply sophisticated attribution logic to corrupted or incomplete raw data. The result is precise nonsense.

First-party data is the foundation. By this I mean data that originates directly from your own systems: CRM records, login events, transaction data from your own backend, newsletter subscriptions with opt-in. This data belongs to the company, is subject to its own data governance practices, and is not dependent on browser restrictions. It is also more durable: a CRM entry does not disappear because a user clears their cache.

Server-side tagging complements this foundation at the measurement layer. Instead of capturing event data exclusively in the browser — where ad blockers, ITP, and browser policies intervene — data collection is shifted to a controlled server. Google Tag Manager Server-Side, Meta's Conversions API, and similar implementations send events directly from the server to advertising platforms. In practice, server-side setups regularly show 15 to 35 percent more measured conversion events than purely browser-based implementations — not because more happens, but because less is lost.

The third component is a clean identity graph: linking anonymous sessions to identified users via first-party identifiers such as hashed emails or customer IDs. Without this graph, any attribution remains a collection of isolated session fragments.

Data-driven attribution (DDA) is not a buzzword. It is a statistical method — originally based on Shapley values from game theory — that calculates the marginal contribution of each touchpoint based on actual conversion paths. Not because a model was defined that way, but because the data suggests it.

Google Ads offers DDA for accounts with sufficient data volume. The model analyzes which touchpoint combinations lead to conversions and which do not, and weights accordingly. This is substantially more robust than rule-based models — with one important caveat: DDA within a platform is only platform-agnostic within that platform. Google DDA does not see Meta touchpoints, organic search queries, or email clicks.

Probabilistic attribution at a cross-platform level solves this problem conceptually — but is technically demanding. The logic: when a user remains anonymous, statistical pattern recognition is used to calculate probabilities for channel attribution and conversion contribution. Tools like Northbeam, Triple Whale (for e-commerce), or Rockerbox take this approach. For B2B companies with long cycles and small sample sizes, however, these models quickly reach their limits: the statistical populations are too small for robust probabilistic conclusions.

My assessment for DACH mid-market companies: DDA within Google Ads and Meta is a sensible first step that can be implemented immediately and requires no additional tool investment. Cross-platform probabilistic models are relevant for companies with more than 500 monthly conversions and stable first-party data pipelines. Below that threshold, incremental measurement is the more reliable instrument.

Incremental measurement is conceptually simple and methodologically robust: it does not ask 'which channel measured the conversion', but rather 'how many additional conversions occur when this channel is active'. The difference is fundamental. Attribution describes correlation. Incrementality measures causal effect.

Geo-lift tests are the most practical form for mid-market companies. The logic: two geographically comparable regions are defined — one test region, one control region. In the test region, a channel or campaign is activated or increased. In the control region, everything remains constant. The difference in conversion rates between the two regions is the channel's incremental signal.

Meta offers its own geo-lift test framework. Google has made CausalImpact available as an open-source library based on Bayesian time-series analysis. For companies that lack the technical resources or data volume for their own geo-lift tests, Marketing Mix Models (MMM) offer an aggregated alternative: they model, based on historical spend and revenue data, which channel weights produced the observed results.

Marketing Mix Models previously had a reputation for being expensive, slow, and consultant-dependent. That has changed. Meridian (Google, open source, published 2024) and Robyn (Meta, open source) make MMM accessible to technical teams. The models are not trivial to set up, but they are no longer an exclusive management consulting deliverable.

For B2B companies with long sales cycles, I recommend a combined approach: geo-lift tests for the two or three largest media channels, quarterly MMM runs for strategic budget allocation, and DDA within platforms for tactical optimization. No single instrument delivers the complete truth — but three complementary instruments approximate it.

Theory is useful. Execution is decisive. The following framework is not academic — it is based on what works in real projects with limited resources.

Step one is data hygiene. Before any attribution model makes sense, tracking must be clean. That means: server-side tagging for the top three conversion events, Enhanced Conversions / CAPI activated, a defined conversion hierarchy (macro-conversions with revenue value, micro-conversions as lead quality signals), and an identity graph concept that at minimum links email hashes from CRM and newsletter systems.

Step two is platform-level DDA. Once data volume permits — Google recommends 300 conversions per month as a minimum threshold — DDA should be set as the default attribution model in Google Ads and Meta. It is not a silver bullet, but it is better than last-click and requires no additional investment.

Step three is the first geo-lift experiment. Not for all channels simultaneously, but for the channel with the highest budget uncertainty — typically brand awareness campaigns or upper-funnel video. A clean geo-lift experiment takes four to eight weeks and delivers a causal answer to a question that rule-based attribution can never address.

Step four, for companies with sufficient data history, is the first MMM run. At minimum 24 months of weekly channel spend and revenue data are needed. Meridian or Robyn can be implemented by a technical analyst. The result is a foundation for strategic budget allocation that does not depend on platform-owned attribution reports — which are structurally platform-partisan.

What this framework does not deliver: real-time attribution at the individual user level. In the post-cookie world, this is no longer possible for most companies — and not necessary. Budget decisions do not require single-path visibility; they require aggregated causal signals. That is where this framework leads.

Attribution is not a measurement problem — it is a decision problem. The question is not whether attribution can be perfectly solved. The question is whether decisions about budget allocation, channel mix, and campaign optimization are based on the best available signal.

Companies still relying exclusively on last-click reports from Google Analytics to drive budget decisions in 2026 are making choices based on a model that structurally favors performance channels and systematically undervalues awareness and mid-funnel activity. This leads to an optimization paradox: channels that appear most measurable receive more budget — not because they are causally more effective, but because they are more visible in the last-click model.

The transition to a sustainable framework is not a one-time technical implementation. It is an organizational decision about how measurement and decision-making are structured within the marketing team. That means: accepting uncertainty, moving away from the illusion of day-level causal statements, and building measurement routines that accommodate longer cycles.

That is uncomfortable. It is also necessary. The alternative — holding on to measurement models that ignore the reality of the post-cookie world — is not a neutral decision. It is a decision for systematically misdirected budget.

The following sources form the methodological basis of this article. They are freely accessible and directly relevant for technically-minded marketing decision-makers.