A marketing mix model fits the weeks. Can you believe its split of orders between channels?
Updated October 2026
A model can fit your weekly orders well and still give an impossible split between channels. Read the verdicts before any channel share.
Updated October 2026
The answer first: one brand's weekly orders
Not automatically. On one brand's weekly orders, the model followed the latest weeks closely, and those weeks were left out of the fit. The settings were chosen using all the weeks, so this is a softer test than a true forecast. Yet it credited the four channels in the model with more than all the orders, which leaves the baseline below zero. That split cannot be believed, and the tool says so itself: it labels the result degenerate.
The data are one anonymised brand's weekly web orders, 182 weeks in all. The model had four channels: Google paid search and three Meta channels. To check the fit we left the last 37 weeks out of the coefficient fit and compared them. The split below comes from the model fitted on all the weeks.
Our position: a model that fits well can still give an impossible split. Read the verdicts before you read any channel share.
| Check | What we looked at | This brand's result |
|---|---|---|
| Does it fit weeks left out of the fit? | Average error on the last 37 weeks | 7.7 percent of weekly orders |
| Does the split add up? | Orders credited to the four channels in the model | 115.0 percent of all orders; baseline minus 15.0 percent |
| Are the channels moving together? | Overlap between channels (VIF) | Google paid search 5.02, Meta Instagram 5.00 |
Population: one brand, 182 weeks, four channels in the model. The baseline is the orders the model attributes to no channel: a baseline, a trend and a yearly shape together.
Data: Conjura Multi-Region MMM Dataset (Anderson 2024), figshare. CC BY 4.0. Changes: one series chosen by rule, daily rows summed to Monday-to-Sunday weeks, four channels.
By MCP Analytics Team. Real data: one brand's weekly web orders, credited above.
Also on this topic: the idea, no data · the analyst version, on simulated data.
What a marketing mix model is
A marketing mix model takes your weekly sales and your weekly spend in each channel. It splits the sales into a baseline, the orders you would have had anyway, and a part for each channel. It is built from past weeks. What it reports are links between spend and orders. They are not proof that a channel caused an order.
A marketing mix model vs a test with a held-back group
A test with a held-back group splits your customers at random. One group sees the ads and one does not, so the gap between them estimates what the ads caused. A marketing mix model has no held-back group. It reads the weeks you already have, and the channels often moved together in those weeks. That is why its split between channels can be shaky even when its fit is good. The model is quick and uses data you already keep. The test is the one to run when you need to know what one channel adds.
What this means for your decisions
- A good fit is not a believable split. The model was close on the weeks left out. It still adds up to more than all the orders. Both are true at once.
- Check that the split adds up first. Here the model's baseline comes out below zero, and orders cannot be negative. The tool's own label for this is "degenerate".
- Do not read the split between channels as the effect of each. The channels moved together. Google paid search and Meta Instagram overlap the most, so they are hard to tell apart. The tool's label for this overlap is "fragile".
- Treat the intervals as optimistic. The weeks' errors are linked to each other, so the tool says its intervals are optimistic. They also leave out the uncertainty in the fitted carryover and ceiling, so none of them is the whole uncertainty.
- These figures describe one brand. They show associations in one brand's weekly data. They say nothing about cause. They say nothing about any other brand. They say nothing about channels not in the file, such as branded search, email or organic. This is not an incrementality test, and it ranks no channel.
How to run a check on your own file
You need a date column and an orders column, weekly, plus a spend column for each channel.
Our steps, for a first pass:
- Leave the last weeks out of the fit, then compare the model with them.
- Add up what the model credits to the channels and compare it with all the orders. If it is over all of them, stop there.
- Look at the overlap between channels. If two overlap a lot, do not read the split between them.
- If the split does not add up, quote neither the split nor the total. Keep the fit on the weeks left out, and the verdict. To learn what one channel adds, run a test with a held-back group.
The free marketing mix tool runs these checks on your file. Run the free marketing mix tool on your own weekly data.
What you need and what you get back
What your file needs:
- One row per week.
- A date column and an orders column.
- One spend column for each channel.
- Enough weeks to leave some out of the fit and still have a long run to fit.
What you get back:
- How the model did on the weeks left out of the fit.
- What the model credits to each channel and to the baseline, set against all your orders.
- The overlap between channels, with a plain verdict.
- A range for each result, with a note on what the range leaves out.
- The method named and the code that ran it.
Questions readers ask
What is a holdout?
Weeks left out of the fit, to see how the model does on those weeks. Here it is the last stretch of weeks.
If the model fits so well, why distrust the split?
Fit says the total follows orders. It does not say how the total divides between channels. Here the divided total is more than all the orders.
What does "the baseline is below zero" mean?
The model credits the four channels in the model with more orders than there were. What is left for the baseline is below zero. That is not possible.
Why can't the split between channels be read?
The channels moved together. Google paid search and Meta Instagram overlap so much that their shares are hard to tell apart.
Does this show which channel works best?
No. It ranks no channel and makes no claim about cause. It is not an incrementality test.
Can I use the model to plan next quarter?
Not from this split: it adds up to more than all the orders. Use the fit on the weeks left out and the verdicts; for what one channel adds, test it with a held-back group.
How we checked
Every number on this page is copied from the tool's own output for this file. Our test station's gate passed on this file. The data are anonymised in the source, so the brand is not named. We left the latest weeks out of the fit and compared them. The settings were chosen using every week, so that test is softer than a true forecast.
Related articles
Your turn: run it on your own file
The same question, on your own weeks. Check the fit on weeks left out of the fit. Then check that the split adds up. Run the free marketing mix tool on your own weekly file.
How we ran it
Three things to know about this reading.
Weeks left out of the fit
We left the last 37 weeks out of the fit and checked the model on them. On those weeks the average error is 7.7 percent of weekly orders. The settings were chosen using every week, so this test is softer than a true forecast. The split and the other figures come from the model fitted on all 182 weeks.
Does the split add up?
We added up what the model credits to the four channels in the model and compared it with all the orders. It comes to 115.0 percent, so the baseline is minus 15.0 percent. The tool labels this "degenerate".
Are the channels moving together?
The channels moved together. Google paid search (5.02) and Meta Instagram (5.00) overlap so much that they are hard to tell apart. The tool labels the split "fragile".