Media measurement
One number is a guess. A range is an answer.
MMM Studio measures what your media actually did, then reports the range the data supports — not the midpoint that reads best in a deck. When a method can’t answer, it says which one would.
The film is an illustration, and the figures in it are examples. Every figure on the rest of this page is from a real fit.
Read the film instead
- 1. Your data
Bring one spreadsheet: what you spent each week, and what you sold. So what did $50M of advertising actually do?
- 2. The check
First it checks whether your data can answer that. Good to model, with two caveats.
- 3. The fit
Then it tries 4,000 plausible answers per channel, in four separate runs. The runs agree. Where the answers pile up is your range.
- 4. One number
Most reports boil that down to one number per channel. One number is a guess.
- 5. The range
A range is an answer. Break-even is $1 back for every $1 spent. Only Paid Search clears it from end to end.
- 6. The decision
So you get the one move your data supports. And a straight answer on what it can't call.
- 7. The refusal
And when your data can't answer at all? Here, two Search channels always ran together. Nothing can split the credit. So the numbers come off the screen. Removed, not greyed out. And it tells you what would fix it.
- 8. The point
One number is a guess. A range is an answer.
Two real runs
Same product. Same screen. Two honest answers.
On the left, a fit you can act on. On the right, one you can't — and the screen says so instead of dressing it up.
Ordered from the channel your data pins down best to the one it pins down least. Each bar is the range the data supports, with a 90% chance the return is inside it; the notch is the middle estimate. Watch the bars widen down the list.
What it will say
No channel here paid for itself across its whole range. That is a finding, and it is the one the chart above supports.
What it won’t say
Which of the rest did best. Their ranges overlap and cross break-even. Any ranking would be ours, not your data’s.
The return column has been removed, not greyed out.
Why you are seeing this
These are the checks the fit failed, named rather than scored. This is the screen that decides whether a measurement tool is trustworthy, and it is the one nobody demos.
Not one method
The right method for the question you're asking.
A mix model answers some questions well and others not at all. Rather than stretch one method over everything, we gate honestly and tell you which method your question needs.
What each channel added, what it returned and how long its effect lasted, each with its range. Runs on two years or more of weekly spend and sales data.
Each channel's return fitted year by year, with its range, instead of one number for the whole window. It needs most of a year of spend in each year. When your data falls short of that, it says so and reports the average.
Past data can't prove this on its own. It takes a test: your ads run in some regions and are held back in others. We design it and feed the result back into the model.
Without weekly price and promotion data, media gets the credit for their effect and every channel looks better than it is.
In plain English
The questions it answers — and the ones it hands back.
The house rule
A number appears only if the model computed it.
It applies to our own screens as much as to your media. No score out of 100, no “7 of 9 checks passed”, no “high confidence” label on a sentence nothing measured.
A model always looks good on the weeks it learned from. So that score never appears alone: it comes with the score on weeks the model was never shown.
You get a verdict — good to model, model with caveats, or don’t model yet — with each problem named. A score is something you chase. A named problem is something you fix.
Sometimes your data leaves a channel less certain than our starting assumption did. You see that exactly as it came out, never rounded to “no change”.
If the model can’t settle on an answer for a channel, its number comes off the screen. A greyed-out number still tells you roughly how big it is, and that is the one thing nobody knows.
Show me the statistics behind these
Shown twice. r² on the weeks the model learned from: 0.901. No weeks were held back on this fit, so there is no second figure to show.
No score. Readiness returns one of three verdicts and the findings behind it. There is no figure to quote, which is the point.
As found. Lowest contraction on the first run: 0.898, on Out of Home. Below zero means the range after the fit is wider than the one assumed before it. It is never clamped to zero.
Removed. No withheld fit in this installation to quote.
Brand manager
A go/no-go you can defend in the meeting.
Verdicts in plain English, one recommended action, and the reason it's recommended written out. No statistics vocabulary required to use it, and none hidden from you either.
Analyst
Every figure traceable to the draw it came from.
Posterior densities, VIF, R̂, ESS, contraction, holdout backtests and response curves — exposed rather than summarised away. Every headline figure deep-links to its method chapter.
Agency lead
A plan that survives the client asking “how sure?”
Scenario compare, an efficiency frontier, and a one-page readout that carries its own caveats — so the difficult question is answered on the page before it's asked out loud.
Find out whether your data can answer the question.
Readiness runs before any model does, and it’s free. If the answer is “not yet”, you’ll get the list of what to fix rather than an invoice.