A Media Mix Model Fits the Weeks. Does Its Split of Orders Add Up?
Updated October 2026
A mix model can fit the weeks and pass a holdout, yet credit its channels with more than all the orders. Add the shares up before you read any channel.
Updated October 2026
By MCP Analytics Team. Data: an anonymised Food & Drink brand from the Conjura Multi-Region MMM Dataset, credited under Sources. For a business reading of a different brand's weekly orders, see the business version.
Also on this topic: the idea, no data · the business version, on one brand's weekly orders.
The answer first: add the split up before you read any channel
A marketing mix model can reproduce your weekly orders and still hand out more than all of them. On one anonymised Food & Drink brand (territory labelled US), the model fits the weeks with an R-squared of 0.829 and misses the 24 weeks left out of the fit by 9.6 percent of weekly orders. Yet its split credits the four channels in the model with 310.8 percent of all orders and leaves the baseline at minus 210.8 percent. That is more than the whole, so the split is not read, whatever the fit says.
Our position: a model that fits the weeks and passes a holdout can still hand out more than all the orders. Add the shares up before you read any channel.
The check costs one addition.
What a marketing mix model is
A marketing mix model relates weekly orders to weekly spend in each channel. It adds a baseline for the orders you would have had anyway, and allowances for carryover (spend today still working next week) and diminishing returns (each extra unit of spend adds less). It is fitted to past weeks, so it reports associations in that data, not proof of cause.
The data and the question
The series is one anonymised brand in the Conjura Multi-Region MMM Dataset (Anderson 2024, figshare 25314841, CC BY 4.0). The brand's vertical is Food & Drink, its territory is labelled US by the dataset, and it reports in USD. We summed the daily rows to Monday-to-Sunday weeks. That gives 116 weeks, from 2021-10-18 to the week starting 2024-01-01, with none dropped.
- Outcome. Weekly orders: web purchases by all customers, 243,048 over the 116 weeks. It is a count, so it has no currency.
- Channels in the model. Four, chosen by a rule fixed before the fit: Meta Facebook, Google paid search, Google Performance Max and TikTok. Spend is in USD.
- Left out of the model. Five other channels (Google video, Meta Instagram, Google Shopping, Google display, Meta other) hold 11.3 percent of this series' spend. Email, organic, direct and affiliate traffic are not in the model either. We say nothing about their effect.
We chose the series from its days, territory and currency before any outcome was summed, and we ran the fit once. The result is poorer than we expected. We keep it and quote it as the result.
The question: when a media mix model fits the weeks and passes a holdout, how much of its split can you read?
The fit and the holdout
We ran our free media mix tool on the 116 weeks, in one run, with no tuning after the fit. R-squared, the share of the ups and downs in weekly orders that the model fits, is 0.829 over all 116 weeks (adjusted 0.818).
To check the fit, the tool compared the last 24 weeks, which were left out of the fit. On those weeks its typical miss is 9.6 percent of weekly orders (mean absolute percentage error), or 505.5 orders a week. The shape settings, carryover and saturation, were chosen using all 116 weeks, so this is a softer test than weeks held back from every choice.
On the usual measures the fit looks good. The sum check is what changes the reading.
The sum check
Add the share of orders credited to every channel and to the baseline. A split that hands out more than 100 percent of orders has not produced a split.
Here the four channels in the model are credited with 310.8 percent of all orders. The baseline, the orders the model gives to no channel (baseline, trend and a yearly shape together), comes out at minus 210.8 percent. The two sum to 100, as they must, but a baseline below zero means the model says the brand would have had negative orders without its advertising. The tool's label for this is "degenerate", and its own warning is that no model can honestly claim advertising produced more than the whole outcome, so the split should not be used.
Because the split is not read, this page quotes no figure for any single channel and ranks none. The tool does print per-channel numbers. We do not use them.
Why the split can fail while the fit passes
The tool gives two further signals.
Overlap. The tool measures how much of one channel's spend pattern the others already explain (the variance inflation factor). About 5 and over means that channel's split is shaky, and 10 and over means it cannot be separated. The highest value here is 49.85, for one of the four channels. In plain words, that channel's spend moves with the baseline and with the others so closely that the model cannot say which one the orders go with. The tool's label is "unreliable". The other three channels are well under 5.
Autocorrelation. The Durbin-Watson statistic is 1.46. A value near 2 means the model's misses are unrelated from week to week. A value below 2 means they run in streaks, and the tool says its intervals are therefore optimistic.
A good total fit does not need a good split. Our reading: the model can trade orders between the baseline and a channel whose spend rises and falls with demand, and the weekly total stays right either way. The overlap number is the warning that this trade is available.
How this compares
A media mix model is one of six ways to ask what advertising did. All six read the same weeks. They differ in what they assume, so none of them ranks above the others, and the right choice depends on the decision.
| Method | What it tells you about the same weeks | What it assumes | When it is the right tool |
|---|---|---|---|
| Last-touch | Credits each order to the final tracked touch before it. | The final touch deserves the credit; earlier touches add nothing. | Reading the last step of a path you can track end to end. |
| First-touch | Credits each order to the first tracked touch. | The first touch deserves the credit; later touches add nothing. | Asking where people first met the brand, among tracked paths. |
| Linear and time-decay | Spreads credit across tracked touches, evenly or weighted toward recent ones. | The spreading rule, fixed in advance, matches how touches work together. | A negotiated compromise between first and last when you need one rule. |
| Platform data-driven attribution | A platform's own model of credit across the touches it can see. | The platform sees the whole path, and its model is right about it. | Reading a platform's view of its own touches, knowing it sees only its own. |
| Media mix modeling | Relates weekly orders to weekly spend, in aggregate, with no user-level tracking. | Enough independent variation in spend across channels (check the overlap); carryover and saturation shapes that are right; no omitted demand driver that moves with spend. | Budget questions across channels where user-level paths are missing, and weeks of history exist. |
| Incrementality test (holdout, geo) | Compares a group that saw the advertising with a comparable group that did not. | Random or matched assignment; no interference between groups; comparable exposure; a representative window. | Asking what one channel added. Of the six, it is the one built to compare against what would have happened without the advertising; it finds cause only if its assumptions hold. |
Three points follow from the table.
- The four attribution rules at the top read tracked paths. They cannot see orders that no tracked touch led to, and they do not test what would have happened without the advertising.
- A media mix model reads association in aggregate. It never shows cause. Its identification rests on the three conditions in its row, and this page's brand failed the first: the overlap number is 49.85.
- An incrementality test compares against what would have happened without the advertising, and finds cause only if its assumptions hold. It answers a narrower question, one channel in one window, and it needs the groups to be set up before the spend.
This page makes no claim about which method is more accurate in general, and gives no statistic about how often any method is wrong.
What this does not tell you
- Nothing about other brands or markets. One anonymised brand, one territory as the dataset labels it, one window, four channels. These results describe this series.
- No cause. These are associations in one fitted model. The tool uses no experiment or lift test, and this is not an incrementality test.
- No ranking. No channel is called best, largest or most efficient. With a split that adds to 310.8 percent, any such statement would be unsupported.
- A softer holdout. The shape settings were chosen on all 116 weeks, so the 9.6 percent figure is not a forecast error.
- Channels left out. Five channels with 11.3 percent of this series' spend, and email, organic, direct and affiliate traffic, are not in the model.
How to check a media mix model on your own file
- Leave the last weeks out of the fit and compare the model with them. Say that the shape settings were chosen on all the weeks.
- Add up the shares the model gives the channels and the baseline. If the channels take more than 100 percent, or the baseline is below zero, do not read the split.
- Check the overlap between channels. About 5 and over, treat that channel's split as shaky; 10 and over, it cannot be separated.
- Check the Durbin-Watson value. Below 2, treat the intervals as optimistic.
- If a channel decision depends on the split, run an incrementality test for that channel.
What you need and what you get back
What your file needs: one date column, one numeric outcome and at least one spend column, with no gaps.
What you get back: the split of orders into baseline and channels with a verdict on whether it adds up, a return per unit of spend with a 95 percent interval for each channel, the overlap between channels with a verdict, a check on weeks left out of the fit, and the method named with the code that ran it. The tool says its intervals are narrower than the truth when the misses run in streaks.
How we checked
Every figure on this page is copied from the answer key for this series. The answer key is the free tool's recorded output on these 116 weeks, kept in our registry as stations/registries/concepts/mmm-academic/results_7569a6.txt. The rule for what the page may say was written before the fit and the fit was run once.
Sources
Data: Conjura Multi-Region MMM Dataset (Anderson 2024), figshare. CC BY 4.0.
Changes: one series (territory labelled US, anonymised Food & Drink brand) chosen by structure, four channels chosen by rule, daily rows summed to Monday-to-Sunday weeks.
Related articles
- What a media mix model separates: the idea, no data
- Can you believe the channel split? One brand's weekly orders
Your turn: run it on your own file
You need one date column, one numeric outcome and at least one spend column. The free tool infers the cadence from the dates and refuses a series that is too short or too broken. Run the free media mix tool on your own weekly data, and add up the split before you read any channel's share.
Questions people ask
These questions are inferred from the key and the brief, not taken from observed search queries.
Can a media mix model fit well and still give a wrong split?
Yes. On one anonymised Food & Drink brand the R-squared is 0.829 and the typical miss on the last 24 weeks, left out of the fit, is 9.6 percent of weekly orders. The split still credits the four channels with 310.8 percent of all orders.
How do I check a split in one step?
Add the shares of every channel and the baseline. If the channels take more than 100 percent of orders, or the baseline is below zero, the split has not been produced and no channel's number should be read.
What does an overlap of 49.85 mean?
One channel's spend pattern is almost fully explained by the others and the baseline, so the model cannot separate its effect. About 5 and over means shaky, and 10 and over means it cannot be separated.
Is a media mix model the same as an incrementality test?
No. A media mix model reads association in aggregate weekly data. An incrementality test sets up a comparison group and is built to find cause. They answer different questions.
Does this say anything about other brands?
No. It describes one anonymised brand, one territory as the dataset labels it, one window and four channels.
References
- Anderson (2024). Conjura Multi-Region MMM Dataset. figshare, 25314841. CC BY 4.0.
How we ran it
Three things to know about this reading.
Weeks left out of the fit
We left the last 24 weeks out of the fit and checked the model on them. The typical miss is 9.6 percent of weekly orders. The shape settings were chosen on all 116 weeks, so this is a softer test than weeks held back from every choice.
Does the split add up?
The four channels in the model are credited with 310.8 percent of all orders and the baseline with minus 210.8 percent. The tool labels this "degenerate".
Why can it fail?
One channel overlaps the others and the baseline so closely that the model cannot separate it (highest overlap 49.85). The Durbin-Watson value is 1.46, so the misses run in streaks.