What a National Marketing Model Hides (And How Segmentation Reveals the Truth)

The Invisible Channel

A national retailer with over 300 stores ran a marketing mix model across two years of data. The model measured every channel: search, radio, out-of-home, programmatic, TV. It produced clear ROI numbers for most of them.

But one major channel returned no usable signal at all. Paid social, one of their biggest digital investments, was statistically undetectable.

Not underperforming. Not low ROI. Completely invisible. The model couldn’t separate its effect from everything else happening at the same time.

The marketing team knew paid social was driving something. They could see it in campaign dashboards. But the econometric model, which was supposed to be the source of truth for budget decisions, had nothing to say about it.

Why It Disappeared: Collinearity

The culprit was collinearity, a statistical problem that occurs when two or more variables move together so closely that a model can’t tell them apart.

In this case, paid social spend was highly correlated with seasonal demand peaks and with spend in other channels. When everything ramps up together (more social ads, more search spend, more foot traffic, more sales), the model can’t attribute the incremental lift to any single channel. It sees a combined wave and shrugs.

This isn’t a flaw in the model. It’s a limitation of aggregated data. When you blend everything into one national view, patterns that move in lockstep become indistinguishable.

The Fix: Split the Model

The breakthrough came from a simple structural change: instead of running one national model, the team built two, one for metro stores and one for regional stores.

This worked because the correlation patterns were different in each segment. Metro and regional stores had different spend levels, different competitive dynamics, and different customer behaviours. Splitting the data broke the lockstep pattern that was hiding paid social’s true contribution.

The result: paid social went from undetectable to one of the highest-performing channels in the mix. Metro stores showed an ROI of approximately 2-3x. Regional stores showed approximately 4-5x. Both were strong, clear signals that the national model had completely missed.

What Else the Split Revealed

Paid social wasn’t the only hidden finding. Segmenting the model surfaced several insights that the national view had obscured:

Search ROI was inflated by seasonality. The national model showed very high search returns, but the segmented view made it clear that search spend was concentrated during peak demand periods. The channel was partly riding a seasonal wave, not purely generating incremental revenue. The recommendation: test search during off-peak periods to get a cleaner read.

Radio underperformed everywhere. This one was visible in the national model too, but the segmented view confirmed it wasn’t a geographic anomaly. Radio returned below 1x ROI in both metro and regional markets, consistently, across two years.

Channel misclassification appeared. When the data was split by region, the team spotted that a metro-only channel (broadcast video on demand) was showing up in regional results. The root cause was a data mapping issue where state-level spend was being split across both segments by default. A national model would have buried this error in the aggregate.

TV saturation differed by segment. TV showed higher ROI in metro stores where spend was lower, and lower ROI in regional stores where spend was higher. This suggested the regional TV budget was past the saturation point, while metro still had headroom. A single national number would have averaged these two realities into one misleading figure.

When to Segment Your Model

Not every brand needs to split their marketing mix model. But you should consider it when:

You operate across geographically distinct markets. Metro versus regional is the obvious split for retail, but it could also be country-level for international brands, or urban versus suburban for service businesses.

You suspect a channel is being masked. If your campaign dashboards show strong performance for a channel but your MMM can’t detect it, collinearity is a likely cause. Segmentation can break the correlation.

Your spend patterns are uniform across channels. If you tend to increase and decrease budgets for all channels at the same time (common with seasonal businesses), a single model will struggle to isolate individual channel effects.

You’re seeing ROI numbers that seem too high or too good. Inflated ROI often means a channel is capturing credit for broader demand effects. Segmentation helps separate the signal from the noise.

How to Do It Right

Choose meaningful segments

The split needs to create groups where customer behaviour or media exposure genuinely differs. Geography is the most common axis, but you could also segment by store format, customer cohort, or product category. The key is that the segments should break the correlation patterns in your data.

Maintain sample size

Each segment needs enough data to produce reliable estimates. Splitting a small dataset into too many groups will give you noisy, unreliable results. Two or three well-defined segments is usually better than ten thin ones.

Compare the segments

Once you have separate models, the comparison is where the insight lives. Which channels perform differently across segments? Where are the saturation points different? These gaps tell you where your budget allocation should differ by market.

Use it for budget optimisation

Segmented models enable segmented budget strategies. Instead of one national media plan, you can allocate differently for metro and regional (or whatever your segments are), directing spend toward the channels that work hardest in each context.

The Bottom Line

A national marketing mix model gives you one version of the truth. It’s a useful version, but it’s also a blended one that can hide as much as it reveals.

If you’re making budget decisions based on a single aggregate model, you might be underinvesting in your best channels and overinvesting in your worst, simply because the data couldn’t tell them apart.

Segmentation is the fix. It’s not more complex modelling for the sake of complexity. It’s a structural change that lets the data speak more clearly. For the brand in this case, it turned an invisible channel into a proven performer and revealed geographic differences that reshaped their entire media strategy.

The question isn’t whether your model is wrong. It’s whether it’s hiding something.

Seeda builds marketing mix models that can be segmented by geography, customer type, or any dimension that matters to your business. Get in touch to find out what your national model might be missing.