MMM + Incrementality: Your 2026 Attribution Stack

80% of cookie tracking is gone. Multi-Touch Attribution (MTA) is no longer reliable. Platform dashboards contradict each other. In 2026, marketers measure contribution using two complementary methods: Marketing Mix Modeling (MMM) and incrementality testing. The problem: few know when to use which. This article shows where to place Robyn (Meta's open-source MMM library), Meta Lift API, and geo-based holdout tests in the same setup.

Last-touch attribution is dead—but what replaces it?

Google Analytics 4 claims "data-driven attribution." Meta says "modeled conversions." TikTok reports its own numbers. All three differ. In 2025, an e-commerce brand spending $100 sees 8 conversions in GA4, 12 in Meta, and 6 in TikTok. Which channel actually works? Last-touch attribution can't answer because users pass through multiple touchpoints and each platform self-credits.

Marketing Mix Modeling solves this differently: it treats channels as independent variables, sales or revenue as the dependent variable, and uses regression to calculate each channel's marginal contribution. Incrementality tests are more direct: expose one group to a channel, withhold it from another, measure the difference. Both break the last-touch illusion, but their use cases don't overlap.

The difference is scale: MMM is macro (long-term, all channels), incrementality is micro (short-term, specific campaign or channel). By 2026, combining both is standard practice.

MMM: Weekly regression setup with Robyn

Meta's Robyn library is the Facebook Marketing Science team's open-source MMM framework. It runs in R, uses Bayesian ridge regression, and auto-fits adstock (lagged effect) and saturation (diminishing returns) curves. At weekly granularity, it outputs each channel's (TV, display, paid social, SEO, email) percentage contribution to sales.

Robyn's 4 setup components:

1. Data gathering: Minimum 1.5 years of weekly data. One row per week. Columns: spend per channel, impressions or clicks; independent variables (price, inventory, seasonality); dependent variable (revenue, orders, conversions). Missing data breaks the model.
2. Hyperparameter tuning: Robyn searches adstock decay (α) and saturation shape (γ) for each channel, running 2000+ model combinations and suggesting the best 5–10 from the Pareto frontier. This step takes 10–30 minutes on 64 cores.
3. Model selection: Pick the model with lowest NRMSE (Normalized Root Mean Squared Error) plus highest decomp.rssd (decomposition stability). Output: each channel's percentage contribution to total sales, estimated ROI, optimal spend allocation.
4. Budget allocation: Robyn's "budget allocator" function redistributes total budget to equalize each channel's marginal ROI. Use this output for next quarter's plan.

When to use Robyn:

• Cross-channel budget decisions (e.g., Q3 planning)
• New channel addition/removal simulation
• Long-term trend analysis (6+ months)

When NOT to use Robyn:

• In-campaign optimization (periods under 2 weeks)
• Platform-specific creative decisions (MMM can't detect creative differences)
• Real-time bidding feedback (weekly lag)

Roibase integrates Robyn in its Digital Marketing service: we connect GA4, server-side GTM, Meta CAPI, and BigQuery, build a weekly ETL pipeline, and visualize model output in Data Studio.

Incrementality tests: Meta Lift and geo-based holdout

MMM answers "how much?" Incrementality answers "does it really work?" Two different questions. If you spend $100K on Meta and get 120 conversions, is that "good"? MMM says "Meta captures 15% of budget, drives 12% of sales." But how many of those conversions would have happened anyway (organic)? That's where incrementality testing comes in.

Meta Conversion Lift

Meta Lift API measures the true impact of Facebook and Instagram ads. How? It withholds the campaign from a small holdout group, shows it to others, and measures the difference 7–14 days later. The difference = incremental conversions.

Setup:

• Open a Lift study before campaign launch (Ads Manager > Measure & Report > Conversion Lift)
• Holdout rate is 5–10% (too small = noise, too large = impression loss)
• Test runs at least 7 days (shorter = low statistical power)
• Output: incremental conversions, incremental CPA, confidence interval

Example result interpretation: Control group: 1,000 people, 40 conversions Test group: 9,000 people, 450 conversions Incremental conversions = (450/9,000 − 40/1,000) × 9,000 = 90 Lift = 90 / (450 − 90) = 25%

So of the 450 conversions the campaign saw, only 90 were truly from the ad. The rest would have happened anyway. Incremental CPA = spend / 90. This number is 30–60% higher than MTA—because it's real.

When to use Meta Lift:

• A/B testing new campaigns or creatives
• Platform decisions (Meta vs. Google vs. TikTok—which is more incremental?)
• Measuring retargeting's true contribution (common issue: retargeting always looks cheap but 80% would convert anyway)

Drawback:

• Only works on Meta (Google has Display & Video 360, but limited)
• Creating a holdout group causes impression loss (revenue dips short-term)
• Minimum 1-week test period—not suitable for daily decisions

Geo-based experiments (geographic holdout)

For non-Meta channels (Google, TikTok, TV), run geo tests: activate campaigns in some cities, leave others untouched, measure the sales difference. Academically, this is the cleanest incremental measurement because no user-level manipulation.

Example setup:

• Select 30 cities (similar population, economic profile)
• Launch Google Ads campaign in 15, keep 15 as control (randomize)
• Wait 4 weeks
• Compare city-level conversions in Google Analytics 4

Analysis:

• Treated cities: average 120 conversions/city
• Control cities: average 95 conversions/city
• Incremental lift: (120 − 95) / 95 = 26.3%

Extrapolate this 26.3% lift nationally. If Google Ads spend is $200K, calculate incremental revenue and incremental ROAS.

When to use geo tests:

• Measuring each channel's net contribution in a multi-channel setup
• Assessing non-digital channels (TV, OOH, podcasts)
• You don't trust platform dashboards

Drawback:

• Too few cities = low statistical power (minimum 20 cities)
• Geographic heterogeneity skews results (Istanbul ≠ Şanlıurfa)
• Slow (4–8 weeks)

Decision tree: which method, when?

Organize all three methods in one setup like this:

Practical workflow:

1. Weekly: Monitor platform dashboards (MTA-like, but don't trust them)
2. Monthly: Test big campaigns with Meta Lift
3. Quarterly: Model all channels with Robyn, redistribute budget based on long-term contribution
4. Twice yearly: Validate each channel's true lift with geo-based tests

This 3-layer setup lets you drive both short-term tactics (which creative works) and long-term strategy (how much budget per channel) with data.

Common misconceptions and tradeoffs

Misconception 1: "If I do MMM, incrementality testing is unnecessary" Wrong. MMM shows correlation, assumes causation. Incrementality testing measures causation. They complement each other. Example: MMM says "Instagram contributes 15%," but Lift test reveals 40% of that would happen anyway. True contribution: 9%.

Misconception 2: "Run incrementality tests on every campaign" No. Holdout groups mean lost impressions. Test only on major decisions (new channel, new creative direction, retargeting strategy shift). Small optimizations use A/B tests instead.

Misconception 3: "Set up Robyn once, then it runs itself" No. Retrain the model every quarter. New channels, price changes, or seasonality shifts require model updates. Robyn setup requires ongoing maintenance.

Tradeoff 1: Speed vs. precision MMM requires 1.5 years of data, results lag by 1 week. Geo tests take 4–8 weeks. Want fast decisions? Rely on platform dashboards but accept 30–50% error margins.

Tradeoff 2: Granularity vs. sample size City-level geo tests keep sample size tight and confidence intervals wide. District-level increases heterogeneity. Weekly MMM can't answer daily questions. Every method has resolution limits.

How to build your attribution stack in 2026

The technical setup consists of:

1. Server-side GTM + first-party cookies: Send clean signals to GA4 and Meta CAPI (consent-based data enrichment, not iOS ATT bypass)
2. BigQuery data warehouse: Consolidate all platform data in one place (GA4, Meta Ads API, Google Ads API, TikTok Ads API, CRM)
3. dbt transformations: Create weekly aggregate tables (one row = one week, each column = one channel spend + one outcome)
4. Robyn pipeline: Run R script in Cloud Run weekly, write model output to BigQuery
5. Looker Studio dashboard: MMM output + platform MTA numbers + incrementality test results side-by-side
6. Slack alerts: Alert on data anomalies if model NRMSE exceeds 10%

Building this stack takes 4–6 weeks. After that, weekly maintenance is 2–3 hours. ROI: budget allocation becomes 15–25% more efficient (Robyn reports 18% improvement in benchmarks).

What to do now

If you're still making decisions based on last-touch attribution, you won't compete in 2026. First step: stream platform data to BigQuery, build 1.5 years of weekly tables. Second step: set up Robyn, train the first model. Third step: open a Meta Lift study on your next major campaign. Fourth step: in 6 months, run a geo-based test to validate cross-channel lift. These 4 steps move your attribution stack from MTA illusion to incrementality foundation.