Recently a client asked us whether they should keep running TV campaigns in the Netherlands, one of the countries they operate in. It's a question we get a lot: which channel drives additional revenue, customers or conversions?
In this case TV was the biggest line on the media budget, and the marketing manager wanted to know if that was still defensible. Because in a click-based attribution model, TV obviously doesn't show up.
1. Why the average marketing dashboard doesn't answer this question
Most dashboards show how channels perform, based on platform-reported or Google Analytics 4 attribution. That gives you insights, but it's far from covering all your channels.
Channels that are easy to measure structurally get too much credit this way: Search captures demand that was often sparked somewhere else, and TV, radio and out-of-home don't generate a click, so they don't get a row in your attribution model, no matter how hard they work. On top of that, every platform reports in its own favour, so the sum of all claimed conversions almost always ends up higher than your actual revenue.
Allocate your budget based on those numbers, and you systematically shift money to what's easy to measure. Not to what works.
A Media Mix Model (MMM) works differently. Instead of tracking individual customer journeys, an MMM looks at the whole picture: how does your KPI move with what you spend, corrected for everything outside your control, like seasonality and search demand. That means offline channels simply count.
2. Three types of questions, three instruments
An MMM doesn't replace attribution, and it isn't an end point in itself. We place it under what we call triangulation: three instruments that each answer a different question and reinforce each other:
Instrument | Answers | Time horizon | In this example |
|---|---|---|---|
Media Mix Model (MMM) | Which channels contribute to your KPI, and where is each channel's saturation point? | Months to years, based on historical data | Over a year and a half of daily data across TV, YouTube, Google Ads, DV360 and the smaller channels |
Incrementality testing (geo-experiment) | What really happens to your results when you switch one channel off or scale it up? | Weeks, one experiment at a time | Recommended to validate the Google Ads finding from the model |
Attribution | Which touchpoints are in the customer journey of the conversions you can track? | Daily, close to real time | The dashboard where the TV question came from |
The three reinforce each other. We feed geo-experiment results back into the MMM as priors, so the model guesses less and knows more on its next run. And you use the MMM results to calibrate your attribution numbers.
3. The case: should TV stay on air?
For this client we built an MMM on over a year and a half of daily data: all media spend, reach and frequency for TV and YouTube among others, plus control variables like search volume and seasonal patterns.
4. What the MMM shows
Three things stood out.
- TV was 67% of the budget, but only ran on 37% of the days. Short, heavy flights with long silences in between. The model estimated the average frequency at around 1.7, while the estimated optimum was around 2.7. The same euros, spread differently across the flight, deliver more: less reach, but reach that sticks better.
- Google Ads was the largest digital channel and also the least efficient. Almost €400 per incremental lead, well past the optimal point on the response curve. Every extra euro there barely added anything. This is exactly the risk of a channel that looks good in an attribution model and gets scaled up because of it, without any certainty about its incremental value.
- Google DV360 DemandGen came in at around €20 per incremental conversion. Twenty times more efficient than Google Ads, and more efficient than TV as well. The channel was also nowhere near its saturation point, and the same went for YouTube. Together these two channels accounted for only a fraction of the budget, but they already showed clear potential.
A channel that looks good in your click-based attribution model often gets scaled up. Whether it actually works incrementally, you don't know.
Channel | Budget share | Finding |
|---|---|---|
TV | 67% | Ran on only 37% of the days. Frequency around 1.7 where the optimum was around 2.7. Same budget, spread differently across the flight, delivers more. |
Google Ads | Largest digital channel | Almost €400 per incremental lead, well past the optimum of the response curve. Extra budget barely adds anything here. |
DV360 DemandGen | Fraction of the budget | Around €20 per incremental conversion, twenty times more efficient than Google Ads. Nowhere near its saturation point. |
YouTube | Fraction of the budget | Not yet saturated, room to scale up. |
5. The answer: yes to TV, but the gains were elsewhere
When we had the model reallocate exactly the same total budget, the result was around 1,600 extra incremental conversions. Almost 9% more results, with no extra investment.
The shift behind it wasn't spectacular: budget away from Google Ads, more towards DV360, YouTube, Meta and Bing. TV stayed at the same amount, but with a higher frequency per flight. So no channel disappeared from the mix, and there was no need for extra budget either. The budget was simply allocated differently, based on the model.
6. How accurate is a model like this (and what it doesn't solve)
In this case we could explain about 90% of the data, and the predictions were off by less than 10%. That's a good score, but an MMM is always part of the truth, never the whole truth.
One challenge, for example, was that TV was only present in the mix on 37% of the days. More data points there would probably help estimate TV even better. The budget of the four smallest channels combined was also less than a tenth of the total. Little variation in spend means little information for the model.
7. The unintended conclusion: budgeting too neatly doesn't make you smarter
For this client this was perhaps the most valuable insight, and it is not mentioned often enough.
A channel you put the same amount of budget into every month gives any model nothing to work with. No variation means no learning.
An extremely skewed split is also mostly noise for a statistical model: if one channel gets 67% of your budget and four others get 2% each, those small ones get statistically drowned out.
If you want to learn something from your marketing spend, deliberately build in enough variation and aim for a healthy balance between channels. Let budgets fluctuate on purpose, switch a channel off for a short period, run TV on more days with a lower daily budget. It might feel inefficient, but it's an investment in the quality of your data, and so in the quality of your next decision.
8. From MMM to action
The biggest risk of an MMM project is that it ends the way most research ends: a deck that disappears into a folder after the presentation, with numbers that no longer hold up six months later.
That's why we can connect the MMM to a budget optimizer. Want to know what happens if you scale up Q1 by 15%, or what's left of your results if the TV budget is cut in half? You run that scenario yourself.
