How a model splits sales between channels
8 minute read
How much of last month's sales came from each channel, and how much would have happened anyway? A marketing mix model answers that from weekly totals, with no tracking of people.
What a model reads
A marketing mix model reads two lists, week by week. The first is what you spent in each channel: Meta, Google Ads, email and so on. The second is what you sold. That's all it needs. It never follows a visitor, never needs a cookie, and doesn't care whether anyone accepted a banner.
Take 26 weeks of both. The model looks for the pattern that explains the sales line. Weeks when Meta spend rose and sales rose with it. Weeks when an email went out and sales jumped for a few days. Weeks when you spent nothing extra and sales still climbed. It tries many combinations and keeps the ones that fit the sales you actually made.
Forge Mark's model is Bayesian. That means it doesn't give one answer. It gives the range of answers that fit your data, and shows you how wide the range is. A narrow range means the data agree. A wide range means they don't, and the honest reading is that you can't be sure yet.
Because it works on totals, it can see what moved with what. It can't see which person saw which ad. That's the trade, and the last section of this guide comes back to it.
Baseline and season
Before any channel gets credit, the model sets aside what you'd sell anyway. Two things sit under every week's sales.
- The baseline is steady demand from people who already know you: repeat customers, word of mouth, people who type your name into Google.
- The season is the calendar: the Christmas rush, the January lull, a bank holiday weekend, your own sale.
Without this step, a model gives Meta the credit for November because Meta spend rose in November. Spend rose because Black Friday was coming, and sales would have risen anyway. Taking out baseline and season first is what stops a model flattering whichever channel happened to scale up alongside demand.
In the sample, the pale layers at the bottom of the second scene are the baseline and the season. Everything above them is what the channels added. In the sample the baseline is the biggest layer. That isn't a failure. It means the brand does some of the selling on its own.
Channels
Each channel gets its own spend line and its own effect. Paid channels are the obvious ones. Email counts too, because a send is a burst of activity you control. Anything you can put a weekly number on can go in.
For each channel the model estimates how many pounds of sales an extra pound of spend brings in, and how sure it is. Two channels that always rise and fall together are hard to separate, because the model can't tell which one did the work. When that happens the range for each one widens. Forge Mark shows the wider range rather than picking a winner.
It also helps to vary spend. A channel you've kept flat for a year gives the model nothing to learn from. A channel you've turned up and down, even by accident, gives it a lot.
Carry-over
An ad doesn't stop working when the spend stops. Someone sees a video on Monday, thinks about it, and buys on Thursday. Marketers call this carry-over, or adstock. A model has to account for it, or it credits Thursday's sale to whatever you spent on Thursday.
68% of the effect lands in a later week
Week 5, no ads£2,500 of sales
Use the control in the picture. With a short carry-over, about 30% of one week's effect shows up in the next week. With a long one, about 70% does. In both settings the sample ads bring in the same £2.50 of sales for every £1 spent. What changes is when.
With long carry-over, week 5 has no ads at all and still brings in £2,500 of sales from the weeks before. A model with no carry-over would look at week 5, see no spend, and file all £2,500 under baseline. The baseline would look bigger than it is and the ads would look weaker.
Channels carry over differently. Search ads often land fast, because the shopper is already looking. Video and social can take longer. A good model lets each channel have its own rate.
Saturation
The tenth pound you spend in a channel earns less than the first. The best searches get bought first. The most interested audience sees the ad first. More spend reaches people who are less likely to buy. This is saturation, or diminishing returns.
The model draws a curve for each channel, with spend across and extra sales up. The curve rises steeply and then flattens. Where you sit on the curve matters more than the channel's average return. A channel can have a great average and nothing left to give at the margin.
In the third scene of the sample, spending 85% more in one channel buys only 35% more sales, because that channel's curve is already flattening. The other channel is still on the steep part. That gap is what a budget move uses.
Moving a budget
With baseline, carry-over and saturation estimated, the model can answer a what-if. Move £1,300 a week from Meta to Google Ads: what happens to sales?
The answer should be a range, not a figure. The last scene of the picture shows one: most likely +£4,100 a month, somewhere between +£2,300 and +£6,900. Those are sample numbers. In real use the range gets wider when you have fewer weeks of data and narrower when you have more.
A range tells you how to act. If the whole range sits above zero, the move is low risk. If it runs from -£1,000 to +£9,000, the model is telling you the data can't settle it yet. Make a smaller move, or run a holdout test to check. Forge Mark's budget planner shows a range and never a single promise.
What it can't tell you
- Which person saw which ad. It works from totals. Attribution follows visits, with consent, and a model doesn't.
- Cause, on its own. A model finds patterns that fit. If two things always moved together, it can't split them. A test is how you check cause.
- A channel you've never used. No spend history means no estimate.
- The future, for certain. It says what's likely if next month looks like the weeks it learned from. A new competitor, a stock-out or a new website can break that.
- Small effects. A channel that is 2% of spend and moves sales by 1% will be lost in the noise.
None of these makes the model wrong. They make it one lens among several. Forge Mark's "Measured by" bar puts the platform's own count, the pixel, a blend, the survey, the model and tests side by side, so you pick the lens and nothing is forced on you.
How much data it needs
There's no magic number, but three things always help.
- Weeks. A full year lets the model see each season once. The sample here uses 26 weeks to keep the picture readable. A real model does better with more.
- Spend that moves. A channel with a trickle of spend, or the same spend every week, can't be separated from noise.
- Enough sales. A store with a handful of orders a week has a lot of noise in its sales line, and the model's range will say so.
When there isn't enough, the right output is a wide range, not a confident guess. Forge Mark's diagnostics are there to show you how much the model has to go on.
Frequently asked questions
Does a marketing mix model need cookies or consent?
No. It reads weekly totals of spend and sales, not visitors. Forge Mark records visits for its other screens only with the visitor's consent.
How many weeks of data do I need?
More is better, and a full year lets the model see each season once. With fewer weeks any model gives a wider range, and a wide range is a fair answer.
Is it better than attribution?
They answer different questions. Attribution follows visits and credits each one. A model reads totals and estimates what each channel added. They can disagree. You pick the lens.
How is it different from an incrementality test?
A model estimates from patterns in past weeks. A test switches ads off in some regions and measures the gap. A test is stronger evidence for one channel, and a model covers every channel at once.
See a model on your own numbers.
Connect your store and ad accounts in a few minutes, read only, and see your sales split by channel with a range on every figure.