Unified Marketing Measurement Framework

Context

Unified Marketing Measurement was delivered through standalone client-requested projects rather than as a single fixed implementation. Each engagement addressed a specific marketing-effectiveness question, while a governed analytics foundation of shared KPI definitions, standardised channel taxonomy and validated commercial outcomes made the work easier to configure, compare and govern.

The framework combined multi-touch attribution, marketing mix modelling and incrementality where appropriate, with the exact design adapted to the customer's data, channel mix, market structure and commercial objective.

Measurement Challenge

Channel dashboards and last-touch attribution could not answer the questions leadership was actually asking. Each channel reported its own attributed result, journeys spanned many touchpoints before a commercial outcome, and part of what moved was explained by seasonality and market conditions rather than by marketing at all. Different questions needed different methods, answered on one consistent set of definitions:

• Which touchpoints influenced progression through the customer or advertiser journey?
• Which channels contributed at an aggregate level over time?
• How much of observed performance was explained by external market conditions?
• Which observed outcomes were genuinely incremental?
• How should the evidence influence marketing and investment decisions?

My Role

I designed and delivered standalone marketing-measurement projects for customers requiring deeper effectiveness analysis. Working from an existing governed data foundation, I tailored attribution logic, MMM approach, controls, causal design and outputs to each business context, while remaining hands-on in model configuration, validation and executive interpretation.

That covered measurement strategy and business-question definition, model selection, multi-touch attribution design, Robyn and Meridian modelling, control-variable design, model validation, KPI governance, scenario interpretation and executive recommendations.

Governed Data Foundation

The models were not starting from raw, inconsistent reporting every time. A reusable foundation was already in place: governed KPI definitions, a standardised channel taxonomy, consistent revenue and conversion outcomes, validated date logic, reusable transformation layers, documented business rules, common dimensions across channels, and QC and reconciliation checks.

That gave each engagement faster setup, less ambiguity in model inputs, better comparability across channels, fewer reconciliation issues, easier model validation and more confidence in executive interpretation.

The modelling approach varied by engagement, but the governed measurement foundation remained consistent. That reduced time spent reconciling definitions and increased time available for modelling, validation and interpretation.

Measurement Architecture

Media and marketing inputs, digital engagement, customer and advertiser touchpoints, commercial outcomes and external market factors are brought onto the governed measurement layer. Three lenses then read from it: multi-touch attribution for journey-level contribution, marketing mix modelling for aggregate channel contribution, and incrementality for causal validation. Their combined evidence feeds executive marketing-effectiveness and investment decisions.

The measurement architecture stayed consistent, while attribution logic, MMM configuration, controls, experimental design and outputs were adapted to the specific customer context.

Example Engagement: Luxury E-commerce Retailer

A luxury e-commerce retailer wanted to understand which marketing channels were genuinely contributing to revenue, where diminishing returns were emerging and how much confidence to place in journey-level attribution when making future investment decisions. The questions the engagement had to answer:

• Which channels influenced customer journeys?
• Which channels contributed at an aggregate level over time?
• Where were channels approaching saturation?
• Which channels still had marginal headroom?
• How much performance was explained by seasonality, promotions or external factors?
• Were attributed outcomes actually incremental?
• How should the evidence affect future media allocation?

Multi-Touch Attribution

I built and managed multi-touch attribution for a B2B advertiser journey and applied the same approach to customer journeys in later engagements. Journeys are constructed from a touchpoint taxonomy and standardised channel definitions, with identity resolution where the data allowed, so that upstream engagement is comparable across sources and connected to downstream commercial outcomes.

First-touch, last-touch and multi-touch views sit side by side, with assisted touchpoints and channel interaction made visible rather than collapsed into the final click. Lookback-window sensitivity is tested explicitly, because attribution settings depend on the buying cycle, channel mix, conversion latency, journey length and data availability. No two engagements used the same window or logic by default.

Attribution Limitations

User-level attribution is limited by consent, cookie loss, platform restrictions, walled gardens and incomplete identity resolution. It shows where contribution appears within the journeys that can be observed, not the whole picture. That is why marketing mix modelling and incrementality complement deterministic attribution rather than compete with it.

Choosing the MMM Approach

Marketing mix modelling provided the aggregate strategic view, separating baseline performance from marketing contribution while accounting for seasonality, promotions, macroeconomic context and other external effects. The modelling approach was chosen per engagement rather than fixed, based on:

• The business question and the depth of scenario planning required
• Time-series depth, number of channels and geographic granularity available
• Whether experimental or incrementality evidence existed to inform the model
• Prior knowledge worth encoding, and the need for explicit uncertainty estimates
• The control variables available to isolate marketing effects

Meta Robyn

Robyn was the fit where channel-level time-series data was rich, many media variables needed exploring, and fast model iteration mattered: typically strongly digital, direct-response contexts where response curves, saturation and carryover were the central questions and budget-allocation scenarios were required.

Its ridge-regression-based approach, adstock and carryover, saturation and response curves, channel contribution and media-efficiency outputs, and built-in scenario analysis and budget optimisation suited those engagements. Hyperparameters and coefficients are engagement-specific and are not published.

Google Meridian

Meridian was the fit where geographic variation mattered, Bayesian uncertainty was valuable and prior information could improve estimation: engagements where geo-level modelling was available, where ROI and marginal ROI uncertainty needed to be explicit, and where experimental evidence could calibrate the model when it existed.

Its Bayesian formulation, priors and credible intervals, geo-level variation, response curves, ROI and marginal ROI estimates, and scenario analysis with explicit uncertainty suited those engagements. Experiment calibration was used only where an experiment had actually been run.

Using Both Approaches

Where both approaches were appropriate, Robyn and Meridian were also run as complementary sensitivity checks. Differences in contribution, saturation, response curves or ROI estimates were investigated rather than averaged. Agreement increased confidence; disagreement triggered deeper review of assumptions, controls, priors and data quality.

MMM Outputs

Outputs reviewed with stakeholders, with exact values kept confidential:

• Channel contribution, baseline demand and seasonality
• External and control effects
• Adstock and carryover, saturation and response curves
• Average ROI, marginal ROI and diminishing returns
• Scenario analysis and budget-allocation scenarios
• Uncertainty and credible ranges where the approach provided them

Incrementality & Causal Measurement

Attribution shows where contribution appears in the journey. Mix modelling estimates aggregate contribution and response. Incrementality asks the question both leave open: would the outcome have happened anyway?

Designs were chosen to fit the engagement and the data available: holdouts, geo experiments, conversion-lift tests, pre and post comparisons with appropriate comparison groups, and other causal measurement techniques. A specific design is only claimed for an engagement where it was actually run.

Measurement Triangulation

Multi-touch attribution answers where journey-level contribution appeared. Marketing mix modelling answers which channels and external factors contributed at an aggregate level. Incrementality answers whether the intervention genuinely created additional outcomes.

The methods do not always agree, and that is the point. Paid social could appear strong in journey-level attribution while MMM showed diminishing marginal returns. Incrementality evidence was then used to determine whether additional exposure was creating genuinely incremental demand. The decision came from the combined evidence rather than any single model.

Disagreement between models was treated as a diagnostic signal, not something to hide.

Decision Output

The investment conversation moved from "which channel has the highest reported ROAS?" to "which channels are creating incremental demand, which still have marginal headroom, and how confident are we in that conclusion?" That reframing, more than any single model output, is what changed how media decisions were discussed.

Results

Exact values are confidential to each client and are not published. The results that mattered were decisions, and they took the same shape across engagements:

• Identified channels approaching saturation and channels with stronger marginal-return headroom.
• Separated baseline and seasonal demand from media contribution.
• Showed where last-touch reporting overstated or understated a channel's contribution.
• Highlighted the difference between average ROI and marginal ROI when discussing the next increment of spend.
• Identified where MMM and MTA told different stories, and resolved them with incrementality evidence.
• Shifted the investment discussion from reported ROAS toward incremental value and marginal opportunity.

Validation

Validation was proportionate to the engagement; not every check applied every time:

• Holdout or validation periods where the series allowed
• Residual diagnostics, model stability and sensitivity to controls and assumptions
• Plausibility of contribution and response curves against commercial knowledge
• Confidence or credible intervals where the approach provided them
• Comparison against experiments where they existed
• Reconciliation against governed source metrics

Business Value

• Less reliance on last-touch attribution and a more defensible interpretation of marketing contribution.
• Visibility into saturation and marginal headroom, and a clearer separation of baseline and external effects from marketing effects.
• Attributed impact compared with incremental impact before it informs investment.
• Stronger evidence for media-investment discussions.
• Reusable governance across standalone client projects, and faster execution because KPI and channel definitions were already governed.

Methods & Measurement Design

• Attribution: multi-touch attribution, first-touch, last-touch and multi-touch perspectives, journey construction, attribution-window design
• Marketing mix modelling: Meta Robyn, Google Meridian, adstock and carryover, saturation and response curves, ROI and marginal ROI, Bayesian priors where applicable, external and control variables
• Causal measurement: incrementality, holdouts, geo experiments and conversion-lift designs where run
• Decision support: scenario planning, measurement triangulation, KPI governance, model validation

Key Lessons

• A common measurement framework does not mean a common model. The design must adapt to the customer, data and business question.
• Governed data and shared definitions reduce analytical friction before modelling begins.
• Robyn and Meridian answer similar MMM questions through different modelling approaches; model choice should follow the evidence and use case.
• MTA, MMM and incrementality are strongest when treated as complementary evidence rather than competing answers.
• Disagreement between methods is often useful because it exposes assumptions that need to be investigated.