How a CPG Brand Used MMM and Geo Lift Tests to Reallocate $4M in Media Spend and Lift Marketing ROI 31%

The Brand, the Budget Problem, and What Was at Stake

The company at the center of this case study is a mid-size CPG brand operating across the personal care category — selling through a mix of national grocery chains, mass-market retailers, and a growing direct-to-consumer e-commerce channel. Annual revenue sat around $180 million, and the total media investment being managed by their marketing team was approximately $22 million per year across paid social, linear TV, streaming audio, display, paid search, and retail media networks.

On the surface, performance looked acceptable. Marketing leadership could point to year-over-year revenue growth of roughly 4%, and their agency partners were reporting strong platform-native metrics: video completion rates above 70%, return on ad spend figures that looked compelling in Facebook Ads Manager, and solid click-through rates on display. The problem was that none of those numbers answered the one question that the CFO had started asking with increasing urgency: how much of our revenue growth would have happened anyway, without the media spend?

That question had real financial weight. The company's board was preparing for a Series C raise, and sophisticated investors wanted to see defensible evidence of marketing efficiency — not platform-reported ROAS that every savvy buyer knows is inflated by view-through attribution and last-click credit. Meanwhile, internal pressure was building to cut the overall marketing budget by 15% heading into the following fiscal year. If the team couldn't prove what the spend was actually doing, the cuts would be arbitrary rather than surgical.

The marketing director estimated that if the team failed to credibly defend the budget — or worse, if they cut the wrong channels — they could lose between 8% and 12% of incremental revenue with no way to recover it before the board presentation. Something had to change.

"We had great-looking dashboards and terrible answers. Every channel looked like it was working if you squinted at it the right way — but we couldn't tell the CFO what would actually happen to sales if we turned any of it off."

— Marketing Director, personal care CPG brand (paraphrased from internal debrief notes)

The stakes were clear: prove incrementality or face across-the-board cuts. The team had roughly two fiscal quarters to build and validate a measurement framework that could withstand scrutiny from both the CFO and outside investors.

Strategy: Why MMM and Geo Lift Tests, Not Just One or the Other

The measurement team considered three options: run a pure marketing mix model, run a series of geo holdout experiments, or combine both approaches in sequence. They chose the third path — and the reasoning behind that choice is one of the most instructive parts of this case study.

A standalone MMM would give them a retrospective view of how each channel had contributed to sales over the past two to three years, controlling for price, distribution, seasonality, and competitive activity. It would help them understand the relative efficiency of each channel and build a scenario-planning tool for reallocation. But MMM has a known limitation: the model's output is only as good as the data fed into it, and for channels that are relatively new or where spend levels haven't varied much historically — in this brand's case, retail media and streaming audio — the model would have wide confidence intervals and limited statistical power. Relying on MMM alone would leave the most uncertain channels unvalidated.

Geo holdout tests, on the other hand, would give them clean, prospective causal evidence for specific channels — but designing, running, and reading out a geo test takes 8 to 12 weeks per channel, and the brand didn't have time or budget to run sequential tests across all seven channels in their mix. Geo tests are also blind to the interaction effects between channels, which the MMM would capture.

The solution was to use the MMM to prioritize. Rather than testing everything, the model's output would identify the two or three channels with the most ambiguous incremental contribution — the ones where the confidence interval was widest or where the model's implied ROI was suspiciously high compared to industry benchmarks. Those channels would then be subjected to geo lift tests to get ground-truth causal evidence. Channels where the MMM showed high confidence in a strong or negligible contribution would be acted on directly without a geo test.

This approach — using MMM vs incrementality testing as complementary tools rather than alternatives — let the team move faster and spend their experimental budget where the uncertainty was actually highest. It's a framework more brands are adopting as measurement sophistication grows: use modeling to orient, and experimentation to validate at the margin.

Critically, the team decided not to rely on multi-touch attribution data from their ad server as a primary input. MTA had been the dominant measurement currency internally for the previous three years, and the team recognized it was distorting channel valuations — particularly overvaluing paid search and display retargeting, which habitually appear at the bottom of the funnel but mostly intercept consumers who would have converted anyway.

Implementation: Timeline, Tools, and How the Tests Were Run

The project ran across two quarters — roughly 24 weeks from kickoff to final readout — and was structured in three overlapping phases.

Phase 1 — MMM Build (Weeks 1–10): The team partnered with an external measurement consultancy to build a Bayesian marketing mix model using three years of weekly data. Inputs included national TV GRPs, weekly digital spend by channel, retail scanner data from their top four retail partners, distribution-weighted ACV, price and promotional variables, and macro indicators including fuel prices and consumer sentiment indices. The model was built in a Bayesian framework specifically because Bayesian MMM surfaces credible intervals around each channel's contribution estimate — making uncertainty explicit rather than hiding it in a point estimate. For a team trying to prioritize which channels to test, that uncertainty quantification was essential. Teams exploring this approach for the first time will find a thorough primer in this guide to marketing mix modeling for digital teams.

Phase 2 — Geo Holdout Tests (Weeks 8–20): Two geo experiments ran in parallel. The first tested paid social (Meta) by selecting 40 designated market areas (DMAs) matched on baseline sales velocity, demographics, and retail distribution. Twenty DMAs were assigned to a holdout group where Meta spend was reduced to zero for eight weeks; the other 20 maintained normal spend levels. The second test examined retail media by similarly splitting 36 markets — holding out retail media activations in 18 of them while keeping standard RMN investment in the control group. Both tests used a difference-in-differences framework with synthetic control matching to isolate the treatment effect from organic sales trends.

Phase 3 — Integration and Reallocation (Weeks 18–24): Geo test results were fed back into the MMM as calibration constraints — a technique sometimes called "MMM calibration with experiments" that anchors the model's channel coefficients to real causal estimates rather than allowing them to float freely on historical correlation. The recalibrated model was then used to run budget optimization scenarios, identifying the allocation that maximized projected incremental revenue within the same total budget envelope. Final reallocation decisions were reviewed by the CMO and CFO before implementation.

The primary tools used were: a custom Bayesian MMM built in Python (using PyMC), Google's Matched Market Test tool for DMA selection and assignment, BigQuery for data warehousing, and Looker for executive dashboards. Total investment in the measurement program — including consultancy fees, internal analyst time, and tooling — was approximately $340,000, or about 1.5% of total media spend.

Results: The Numbers Before, During, and After

The findings were uncomfortable in exactly the way the team had suspected — and in one case, they were surprising in a direction nobody had anticipated.

The MMM's initial output flagged paid search (branded keywords) and programmatic display as generating minimal incremental revenue. The model estimated that branded paid search was capturing consumers who would have navigated directly to retail product pages anyway, with an incremental ROAS of approximately 0.6x — meaning the brand was spending $1 to generate $0.60 in revenue it wouldn't have gotten otherwise. Display retargeting showed a similar pattern, with the model attributing most of its apparent conversions to users who were already deep in the purchase funnel before the ad appeared.

The geo holdout results on retail media were the real surprise. The brand's retail media investment — $3.9 million per year across two major RMNs — had been reported by the networks themselves at a ROAS of 4.2x. The geo test told a different story: in holdout markets where retail media was switched off, sales declined by only 4% relative to control markets, implying a true incremental ROAS closer to 0.9x. The gap between the platform-reported number and the causal number was staggering, and it identified retail media as the single largest source of misallocated spend in the portfolio.

"The retail media test was the moment everything clicked. A 4.2x ROAS in the dashboard versus 0.9x in reality — that $3M gap wasn't performance, it was attribution credit for sales that were going to happen regardless."

— Senior Data Scientist, measurement consultancy (paraphrased from project debrief)

The paid social geo test, by contrast, vindicated the channel. Holdout markets showed a 9.3% decline in sales versus control, suggesting that Meta was generating meaningful incremental volume — though at a lower efficiency than the platform's own reporting suggested. The team kept paid social spend roughly flat but shifted creative strategy toward upper-funnel video, which the MMM flagged as the higher-ROI format within the channel.

The 31% lift in marketing ROI was measured using the recalibrated MMM applied to post-reallocation spend data — not self-reported platform metrics. The brand also ran a follow-up geo validation in Q4 to confirm the model's projections, which aligned within 6% of the experimental results, giving the CFO sufficient confidence to present the numbers to the board.

It's worth noting that the broader adoption of this kind of rigorous measurement is growing: Supermetrics' 2026 marketing data report found that 71% of enterprise marketers surveyed were using marketing mix modeling — a signal that MMM has moved from a tool used only by the largest CPG conglomerates to a standard practice for mid-market teams willing to invest in the methodology.

Key Learnings: What Worked, What Failed, and What Surprised Everyone

What worked: Using the MMM to prioritize which channels to test experimentally was the single highest-leverage decision the team made. Without that triage step, they would have spent six months running sequential geo tests with no guarantee of reaching actionable conclusions before the board deadline. The Bayesian MMM's uncertainty quantification directly enabled that prioritization — point-estimate models would not have surfaced the same signal.

Calibrating the MMM with geo test results was also a meaningful improvement over the uncalibrated model. The recalibrated version's channel coefficients changed significantly for retail media and display — and those changes made the budget optimization recommendations substantially different from what the original model would have produced.

What failed: The team underestimated the internal communication challenge. When geo test results showed that nearly $2.8 million of the retail media investment was generating minimal incremental return, the RMN partners pushed back hard — providing alternative measurement approaches that, predictably, showed better results for their own platforms. Managing those relationships while holding firm on the experimental methodology required more CMO-level involvement than the team had planned for. Build stakeholder alignment early, and document your methodology before you share results.

What surprised everyone: Linear TV performed significantly better than the uncalibrated MMM initially suggested. Before the geo tests were run and used to recalibrate the model, TV's contribution was being undersold because the model was over-attributing sales to digital channels that ran simultaneously. Once retail media's true incremental contribution was re-anchored at the experimental estimate, TV's implied coefficient rose — which justified the decision to shift $2.2 million from retail media and display into incremental TV weight. That counterintuitive finding — that a "traditional" channel was being undervalued relative to supposedly data-rich digital channels — is one that many practitioners report when they move from MTA or siloed platform measurement to integrated causal frameworks.

How to Replicate This: An Actionable Checklist

This framework is replicable for any CPG brand with at least $8–10 million in annual media spend and 18 or more months of consistent weekly data. Smaller budgets and shorter data histories will produce less statistically reliable MMM outputs, though a geo-first approach may still be viable.

• Audit your existing measurement stack first. Identify which channels have platform-reported metrics that diverge significantly from your intuitive sense of their business impact. Those gaps are where misallocation tends to hide.
• Assemble at least two years of weekly data covering spend by channel, baseline sales (ideally scanner data or direct e-commerce transactions), price, distribution ACV, promotional events, and relevant external variables. Data quality matters more than model complexity.
• Build a Bayesian MMM, not a classical regression. The credible intervals around each channel's contribution are what make uncertainty-driven prioritization possible. If your current vendor or tool only produces point estimates, push for uncertainty ranges.
• Use the MMM to rank channels by uncertainty, not just by implied ROI. High uncertainty + high spend is your highest-priority candidate for a geo holdout test. High confidence + low ROI is a candidate for immediate budget reduction without a test.
• Design geo tests with statistical rigor. Match test and control markets on at least 12 weeks of pre-period sales data, demographic composition, and retail distribution. Use synthetic control or difference-in-differences analysis, not simple before/after comparisons.
• Calibrate your MMM with experimental results before running budget optimization. Uncalibrated optimization is just reorganizing your prior beliefs — calibration anchors the model to causal reality.
• Build stakeholder alignment before you share findings. If a channel partner is going to receive a result that implies their platform is not generating incremental value, prepare your CMO and procurement team for the conversation before the readout meeting.
• Plan a post-reallocation validation. Run a second, lighter-weight geo test or monitor the MMM in the two quarters after reallocation to confirm the model's predictions are tracking reality. This closes the loop and builds internal credibility for future measurement investments.
• Budget 1–2% of media spend for measurement. Industry practitioners widely observe that teams spending less than 1% of media on measurement tend to underinvest in data quality and analytical depth, which undermines the credibility of results.

Frequently Asked Questions