1. Why the classical model broke

The classical last-click and multi-touch attribution models assumed that a coherent identifier — the third-party cookie — followed a user across the web. That assumption is now false almost everywhere. Safari and Firefox block third-party cookies entirely. Chrome is rolling out its Privacy Sandbox. iOS App Tracking Transparency capped identifier availability. Adblockers strip pixels. Consent frameworks legitimately suppress tracking for users who opt out.

Even where identifiers still work, cross-device journeys mean the same person shows up as three different users on desktop, phone, and tablet. Traditional MTA models weren’t built for this; they systematically overcredit the last touchpoint before conversion and undercredit upper-funnel channels.

2. The modern three-layer stack

The stack that works in 2026 has three layers, each with a clear job.

• Layer 1 — Platform-native modeled attribution
  GA4 Data-Driven Attribution, Google Ads Enhanced Conversions, Meta modeled conversions, LinkedIn revenue attribution. Use these for tactical channel-level optimization. Fast, granular, but privacy-limited.
• Layer 2 — Marketing Mix Modeling (MMM)
  Aggregate-level econometric model that estimates channel-level ROI from spend and outcome time-series. Use for strategic budget allocation and cross-channel decisions. Slower, less granular, but privacy-safe.
• Layer 3 — Incrementality testing
  Controlled experiments — geo holdouts, matched-market tests, platform-native lift studies. Use quarterly to validate the model outputs and correct systematic biases.

3. Layer 1 in practice — platform-native modeled attribution

Each major ad platform now runs its own modeled attribution system. Google Ads uses Data-Driven Attribution, Meta uses modeled conversions from its Advantage+ machinery, LinkedIn uses revenue attribution off matched audiences, TikTok uses its own attribution model.

These models are usable for tactical channel optimization — deciding which ad set to scale, which creative to kill, which audience to retire. They are not usable for cross-channel decisions because each platform overclaims for itself. That overclaim isn’t malice; each platform can only see its own touchpoints and assumes it’s responsible for outcomes it observed.

4. Layer 2 in practice — Marketing Mix Modeling

MMM is having a renaissance. Modern MMM tools like Meridian (Google), Robyn (Meta), and specialist vendors run in weeks, not months, and estimate channel-level ROI from spend, outcome, and controls data — without any user-level identifiers.

Practically, MMM needs 18–24 months of weekly time-series data across spend, outcomes, and control variables (seasonality, promotions, competitor activity, macro factors). The output is a set of channel-level response curves that answer the strategic question: if I move $100K from channel A to channel B, what happens to revenue?

MMM cannot answer tactical questions (should I scale this ad set?), but it is the only method that fairly credits upper-funnel channels — brand, YouTube, connected TV, PR, out-of-home — that platform MTA systematically underweights.

5. Layer 3 in practice — incrementality testing

Incrementality tests are the truth serum for the entire measurement stack. Well-designed tests measure the causal lift a channel contributes — not the correlation an attribution model implies.

The most common tests we run:

• Geo holdouts
  Turn a channel off in half the geographies for 4–8 weeks; measure the difference in outcomes. Simple, robust, works for most channels.
• Matched-market tests
  Same idea, but on statistically-matched geo pairs. Higher statistical power on smaller budgets.
• Platform-native lift studies
  Meta, Google, and TikTok all offer randomized-holdout lift studies. Fast, valid within-platform, but limited by platform.
• Media-mix experiments
  A/B test between two spend allocations across multiple channels. Best for validating MMM outputs.

6. Sequencing the build — where to start

The build order matters. Skip a layer and the ones above stall.

• Month 1–2
  Fix the data foundation. Server-side tracking, Enhanced Conversions, CAPI on every major channel, offline conversion imports.
• Month 3–6
  Deploy platform-native modeled attribution correctly. GA4 DDA, Ads DDA, Meta modeled conversions. Standardize channel taxonomy.
• Month 6–9
  Start MMM. Assemble the historical spend and outcome data, add controls, ship the first model. Validate against known intuition.
• Month 9–12
  Begin quarterly incrementality tests. Use them to correct MMM and platform-attribution biases. Formalize the operating cadence.

7. Tooling — what we use

A partial list of the tooling we deploy or integrate with:

• MMM
  Google Meridian (open source, well-documented) or specialist vendors like Recast or Cassandra AI.
• Attribution warehouse
  BigQuery streaming from sGTM, dbt or Dataform models, event-level plus daily-aggregate tables.
• Experimentation
  GeoLift (open source) for geo holdouts; platform-native tools for within-platform lift studies.
• BI
  Looker or Metabase over the warehouse copy. Never point BI at platform APIs directly.

8. The operating cadence

What we run on a weekly, monthly, and quarterly cadence:

• Weekly
  Review platform-native modeled attribution for tactical channel decisions.
• Monthly
  Update MMM with the latest week. Re-estimate channel response curves. Flag material shifts.
• Quarterly
  One incrementality test, one MMM validation review, one budget-allocation revision based on refreshed MMM output.

9. What still doesn’t work

A few honest caveats. Cross-device attribution is still fundamentally impaired without third-party cookies. Small-brand MMM (under $1M annual spend) can be unstable due to insufficient time-series signal. Incrementality tests on brand and thought-leadership investments remain hard to design well. And platform-native modeled attribution is only as good as the underlying signal feed — a poorly-implemented sGTM deployment produces poor modeled attribution.

The stack is a lot better than what came before. It isn’t magic.

10. Where measurement is going

Two shifts are worth watching. First, on-device measurement — models that run locally on the user’s device without sending identifiers back to the platform — will materially improve modeled attribution over the next 24 months. Second, cross-channel MMM standards will formalize; multiple platforms are converging on shared open-source frameworks. The brands that get their measurement foundation right in 2026 will be positioned to take advantage of both.