A DTC skincare brand spending $6.4 million annually across seven marketing channels had a problem that plagues most growth-stage companies: they had no idea what was working. Their attribution platform said Google was driving 60 percent of conversions. Their Facebook dashboard said Meta was driving 45 percent. The numbers did not add up. Every channel was taking credit for the same sales. Meanwhile, their TV spend — $1.8 million per year — showed zero conversions in their digital attribution system because there was nothing to click.
We built a Bayesian marketing mix model that estimated the true incremental contribution of each channel using three years of weekly spend, revenue, and external data. The results were illuminating: TV was their second most efficient channel (not zero as digital attribution suggested), paid social was over-invested by 40 percent (most of the attributed conversions would have happened anyway through organic), and email — which received almost no budget — had the highest ROI of any channel. We reallocated $2 million in budget based on the model's recommendations. Revenue grew 18 percent in the following quarter while total marketing spend increased by only 3 percent.
Marketing mix modeling is one of the most valuable and most defensible services an AI agency can offer. It addresses a universal pain point — "where should we spend our marketing budget?" — with a rigorous, privacy-compliant methodology that does not depend on cookies, pixels, or any user-level tracking. Here is how to deliver it.
Why Marketing Mix Modeling Is a Tier-One Agency Service
The marketing measurement landscape has shifted dramatically, and MMM is the biggest beneficiary.
What is driving demand:
- Cookie deprecation and privacy regulations: Third-party cookies are effectively dead. GDPR, CCPA, and iOS privacy changes have gutted traditional digital attribution. MMM does not need user-level tracking data.
- Cross-channel measurement gap: Most companies run marketing across 5-15 channels. No single attribution platform can measure all of them in a unified framework. MMM can.
- Budget pressure: CMOs are under intense pressure to prove marketing ROI. MMM provides the most rigorous framework for quantifying the return on every marketing dollar.
- Offline channel measurement: TV, radio, out-of-home, events, and sponsorships are invisible to digital attribution. MMM captures their impact.
- Strategic budget allocation: MMM answers the most important question in marketing: "If I have one more dollar to spend, where should I put it?"
What clients will pay: MMM engagements range from $80,000 for a focused model covering 3-5 channels to $300,000+ for comprehensive models covering all channels, geographies, and product lines. Ongoing measurement retainers run $10,000-30,000 per month.
Client types: Companies spending $2 million or more annually on marketing across multiple channels. DTC brands, consumer goods companies, retail chains, financial services, travel and hospitality, and automotive are the most common verticals.
Understanding Marketing Mix Modeling
What MMM Actually Does
Marketing mix modeling uses statistical techniques to estimate the relationship between marketing activities (inputs) and business outcomes (outputs) over time.
The core model: Sales = Baseline + Channel1Effect + Channel2Effect + ... + External_Factors + Error
Where:
- Baseline represents sales that would happen with zero marketing (brand equity, repeat customers, organic demand)
- Channel effects capture the incremental impact of each marketing channel
- External factors account for seasonality, economic conditions, competitive activity, weather, and other non-marketing drivers
- Error captures unexplained variation
Key Concepts
Adstock (carryover): Marketing has a delayed and decaying effect. A TV ad shown today still has impact next week, but less. The adstock function models this carry-over effect with a decay rate that varies by channel.
Saturation (diminishing returns): The first $100,000 spent on a channel drives more incremental revenue than the last $100,000. Saturation curves model this diminishing returns relationship, which is critical for budget optimization.
Decomposition: Breaking total revenue into contributions from each channel, baseline, and external factors. This tells the client what percentage of revenue each channel is responsible for.
ROAS (Return on Ad Spend): Revenue generated per dollar of marketing spend, for each channel. The primary metric clients care about.
Marginal ROAS: The incremental revenue from the next dollar spent on a channel. This is what drives budget optimization — spend more on channels with high marginal ROAS and less on channels where you are already in the flat part of the saturation curve.
MMM vs Attribution
Attribution and MMM answer different questions and should be complementary, not competing:
Attribution answers: "Which touchpoints did this specific customer interact with before converting?" It is useful for tactical, within-channel optimization.
MMM answers: "How much total revenue did each marketing channel drive, including effects that cannot be tracked at the user level?" It is useful for strategic budget allocation.
Where attribution fails and MMM succeeds:
- Measuring TV, radio, out-of-home, and event marketing
- Capturing view-through effects that are not clickable
- Accounting for cross-device behavior
- Working without user-level tracking in a privacy-restricted world
- Measuring the interaction effects between channels
Technical Architecture for MMM
Data Requirements
Marketing data (by channel, by week):
- Spend
- Impressions, GRPs, or reach
- For digital channels: clicks, click-through rate
- For TV: spot counts, daypart distribution, program type
- For promotion: discount depth, promotional type, in-store display
Sales/revenue data (by week, ideally by geography):
- Total revenue or units sold
- Broken down by product line or category if possible
- Online vs offline if applicable
External data:
- Seasonality indicators (week of year, holiday flags)
- Economic indicators (unemployment rate, consumer confidence, GDP growth)
- Weather data (temperature, precipitation by geography)
- Competitive activity (competitor spend if available, or proxy measures)
- Distribution changes (new store openings, distribution gains/losses)
- Pricing data (own pricing and competitor pricing)
Data granularity: Weekly data is the standard for MMM. Daily data can work for digital-only businesses. Monthly data is too coarse for most models. Geographic breakdowns (by DMA, state, or region) dramatically improve model accuracy by providing cross-sectional variation in addition to time-series variation.
Modeling Approach
Bayesian MMM is the current state of the art. It offers several advantages over traditional frequentist regression:
- Prior information: Incorporate business knowledge (e.g., TV ROAS should be between $1 and $5) to regularize estimates
- Uncertainty quantification: Every estimate comes with a credible interval, not just a point estimate
- Flexible model structure: Easily incorporate non-linear effects, hierarchical structures, and complex interaction terms
- Robust with limited data: Priors help stabilize estimates when data is limited (which it always is in MMM)
Model implementation:
- Data preparation: Aggregate to weekly granularity, align all data sources, handle missing values
- Adstock transformation: Apply geometric or Weibull adstock functions to each channel's spend or impression data
- Saturation transformation: Apply Hill or logistic saturation functions to model diminishing returns
- Model specification: Define the Bayesian model structure including priors for all parameters
- Model fitting: Use MCMC sampling to estimate posterior distributions for all parameters
- Convergence diagnostics: Verify that the MCMC chains have converged and the model is well-specified
- Model validation: Hold out recent data for out-of-sample validation, cross-validate across geographies
- Decomposition: Calculate the contribution of each channel to total revenue
- Budget optimization: Use the fitted model to find the budget allocation that maximizes total revenue
Budget Optimization
The optimization layer is what makes MMM actionable:
Inputs:
- Fitted model with saturation curves for each channel
- Current budget allocation
- Total budget constraint (same budget, 10 percent more, 10 percent less)
- Channel-specific constraints (minimum or maximum spend per channel)
- Business rules (must maintain presence in specific channels)
Outputs:
- Recommended budget allocation by channel
- Expected revenue under current vs recommended allocation
- Marginal ROAS by channel at current and recommended spend levels
- Scenario analysis (what if budget increases by $500K? What if we cut TV entirely?)
Sprint-Based Delivery Framework
Sprint 1: Data Collection and Preparation (Weeks 1-4)
Activities:
- Kick-off meeting to align on scope, channels, and success criteria
- Data request documentation with specific requirements for each data source
- Data collection from client and external sources
- Data quality audit and gap identification
- Data transformation and alignment (common time granularity, currency normalization)
- Exploratory data analysis to understand relationships and identify anomalies
Key risk: Data collection is always slower than expected. Clients struggle to provide clean, complete data. Marketing platforms have different reporting conventions. Budget 60-70 percent of this sprint for data wrangling.
Sprint 2: Model Development (Weeks 5-8)
Activities:
- Feature engineering (adstock transformations, saturation functions, interaction terms)
- Prior specification based on industry benchmarks and business knowledge
- Model fitting and convergence diagnostics
- Model validation (holdout testing, cross-validation)
- Sensitivity analysis (how do results change with different priors or model specifications?)
- Channel decomposition and ROAS estimation
Sprint 3: Optimization and Insights (Weeks 9-10)
Activities:
- Budget optimization under various constraints and scenarios
- Develop strategic recommendations based on model results
- Create visualization of channel contributions, saturation curves, and optimization results
- Prepare client presentation with actionable recommendations
- Internal review of results for coherence and business validity
Sprint 4: Delivery and Handoff (Weeks 11-12)
Activities:
- Client presentation of results and recommendations
- Interactive workshop to explore scenarios and what-if questions
- Deliver model documentation including methodology, data sources, assumptions, and limitations
- Set up ongoing model refresh process
- Plan for validation through controlled experiments (geo-tests)
Common Delivery Challenges
Data Quality and Completeness
MMM is only as good as the data feeding it. Common issues:
- Missing spend data: A channel's spend was not tracked consistently for part of the analysis period
- Aggregation mismatches: Digital spend is daily, TV spend is weekly, promotion data is monthly
- Definitional inconsistencies: What counts as "paid social" varies across teams and time periods
- External factor data gaps: Competitor spend data is rarely available; you need proxy measures
Our approach: Create a comprehensive data specification document before the project starts. Review it with the client in detail. Build in a "data remediation" buffer of 1-2 weeks.
Client Pushback on Results
MMM results sometimes contradict what the client believes about their marketing. Maybe the model says their beloved TV campaign is not working. Maybe it says their pet channel is over-invested.
Handle this proactively:
- Set expectations during kick-off that the model may reveal uncomfortable truths
- Present results with uncertainty ranges, not just point estimates
- Validate surprising results with additional analysis (does the finding hold across sub-periods? Across geographies?)
- Propose controlled experiments (geo-tests) to validate controversial findings
- Acknowledge model limitations honestly — MMM is a statistical estimate, not ground truth
Multicollinearity
Marketing channels are often correlated — companies increase spend across all channels during peak seasons and reduce across all channels during slow periods. This correlation makes it difficult to isolate individual channel effects.
Mitigations:
- Use Bayesian priors to regularize estimates when data is not informative enough to distinguish channels
- Incorporate geographic variation (channels vary by market) to break time-series correlation
- Recommend controlled experiments to validate model-based estimates
- Be transparent about which channels have well-identified effects and which are uncertain
Short Data Histories
Ideally you want 2-3 years of weekly data for a robust MMM. Many clients only have 12-18 months, especially for newer channels.
Strategies:
- Use informative Bayesian priors based on industry benchmarks to compensate for limited data
- Incorporate geographic variation to increase effective sample size
- Be upfront about wider uncertainty intervals with shorter data histories
- Plan for model refinement as more data accumulates
Pricing MMM Engagements
Standard MMM project:
- 3-5 channels, single geography: $80,000-120,000
- 5-10 channels, single geography: $120,000-180,000
- 10+ channels, multiple geographies: $180,000-300,000
- Enterprise (multiple brands, products, and geographies): $300,000-500,000
Ongoing MMM retainer:
- Monthly model refresh and reporting: $10,000-20,000 per month
- Quarterly model rebuild with new data: $15,000-25,000 per quarter
- Ad hoc scenario analysis and optimization: $3,000-8,000 per request
Value justification: A company spending $10 million on marketing that reallocates 20 percent of budget based on MMM insights and achieves a 15 percent improvement in marketing ROI is generating $300,000 in incremental revenue per year. A $150,000 MMM engagement pays for itself in 6 months.
Your Next Step
Identify a company spending at least $2 million annually across multiple marketing channels, especially one with significant offline spend that is invisible to digital attribution. Offer a free diagnostic meeting where you review their current measurement approach and identify the blind spots. When you show them that 30-50 percent of their marketing spend is unmeasured by their current attribution system, the case for MMM sells itself. Start with a focused model covering their top 5 channels and expand from there.