Marketing Mix Modeling: See What Each Channel Really Earns, Without Cookies
Marketing Mix Modeling estimates how much each marketing channel contributed to sales, using weekly totals (spend, revenue, seasonality, promotions) instead of tracking individual users. It was built for TV and print, when there were no clicks to count. It faded when pixels made user-level tracking possible, and it came back for exactly the reason it was invented: tracking stopped working.
One view: where sales came from, and where the next budget goes
Career figures, not model results. What your own model produces is established by the data audit.
What Marketing Mix Modeling Is, and Why It Came Back
Marketing Mix Modeling estimates how much each marketing channel contributed to sales, using weekly totals (spend, revenue, seasonality, promotions) instead of tracking individual users. It was built for TV and print, when there were no clicks to count. It faded when pixels made user-level tracking possible, and it came back for exactly the reason it was invented: tracking stopped working.
How much did each channel cause?
Not which ad was clicked last, but how much extra revenue each channel produced.
Weekly totals, not people
Spend per channel, revenue, and things like seasonality, discounts and stock-outs.
A likely range per channel
Each channel's return comes back as a range, so you see how sure the model is.
Why user-level tracking got weaker
- iOS tracking prompts and browser tracking prevention
- Cookie restrictions and shorter attribution windows
- Consent banners hiding events, mostly in the UK and EU
- Every platform grading its own homework
Why MMM is worth running again
- Never needed user data, so privacy changes don't hurt it
- Sees TV, print, retail and events next to digital
- Handles small, messy datasets most businesses actually have
- Cheap to run now that compute and modelling tools are affordable
MMM didn't get better than it was in 2010. Everything around it got worse, and the maths got more forgiving of imperfect data. That makes it the most defensible way most businesses can measure channels today, alongside the user-level work on customer lifetime value, not instead of it.
MMM vs Attribution vs Incrementality: Three Different Questions
These three are constantly treated as rivals. They aren’t. Each answers a different question on a different schedule, and a mature setup runs all three. Treating any single one as the only truth is where most measurement goes wrong.
- Data
- User-level; needs tracking and consent
- Covers offline
- No, digital only
- Cadence
- Daily and weekly tweaks
- Weak spot
- Blind to anything untracked; platforms report on themselves
- Data
- Weekly totals; no user data
- Covers offline
- Yes: TV, print, retail, events
- Cadence
- Quarterly budget planning
- Weak spot
- Needs ~2 years of history; must be checked with experiments
- Data
- Region or audience holdouts
- Covers offline
- Yes, any channel you can pause
- Cadence
- Occasional, one question at a time
- Weak spot
- Slow and costly; answers one channel per test
Adstock and Saturation: The Two Ideas That Make MMM Work
A straight line of “spend vs revenue” is close to useless, because advertising doesn’t behave in a straight line, either over time or over volume. Two ideas fix that, and they are the heart of the whole method. Everything else is bookkeeping around them.
Money spent today keeps working next week
An ad seen this week still nudges sales in the weeks after. TV and brand video fade slowly; search fades fast. The model learns how fast each channel fades.
The 10th thousand never earns what the 1st did
Every channel hits diminishing returns. The model draws a response curve per channel and marks where your current spend sits on it.
Each channel's return comes back as a range, and the width tells you something
A narrow range means the data is clear. A wide range that crosses break-even means the data can't tell yet, which is far more useful than one confident number that happens to be wrong. The model can also use what is already known (for example, that a channel's return is unlikely to be negative), which keeps it sensible on short, messy datasets where every channel rises together in peak season.
narrow range
medium range
test this one next
Illustration of how results are read, not client data.
Where does your next budget increment belong?
A short walkthrough of a channel's response curve: where current spend sits, what the flat part means, and how a planner turns that into a reallocation, with the uncertainty still visible.
A Custom Model Built on Your Data, Not an MMM Tool You Rent
There are open-source modelling frameworks and paid MMM platforms, and some are very good. But a framework is not a model, and a platform’s model was not built for your business. What I deliver is a custom marketing mix model: built on your raw weekly data with advanced data science, benchmarked against alternatives before one approach is chosen, and supervised and reviewed by me as data scientist and consultant. It is not a plug-in like an SEO tool.
The full comparison of commercial MMM platforms is on MMM SaaS vs Cognitive Intelligence.
| Rented MMM platform | Custom model (Cognitive Intelligence) | |
|---|---|---|
| Who builds it | The vendor: one product for many brands | Built for your business, on your data, by me |
| What it knows | Generic defaults across clients | Your channels, seasons, promotions and market |
| How it's checked | Fits history; trust the dashboard | Held-out periods, then real experiments |
| Uncertainty | Often a single number per channel | A range per channel, with the reasons |
| Offline & odd channels | Whatever the connectors support | Anything you can put a weekly number on |
| Ownership | Model and history stay with the vendor | You own the model, code and outputs |
| Who reviews it | Nobody, unless you pay for services | Supervised and reviewed by me every refresh |
Different approaches recover channel contributions differently on the same data. The differences are big enough to change a budget decision, which is why candidates are tested on your dataset instead of picked by reputation.
Calibration: Why an Unchecked MMM Is Just a Curve Fit
A model fitted to history has found correlations. It hasn’t proven that more spend on a channel causes more revenue, and since every channel rises and falls with the seasons, correlation is least trustworthy exactly here. Calibration with real experiments is what turns the output into a number a CFO can act on.
Run a geo-lift test
Pause or raise one channel in matched regions while others stay the same. The gap in sales is a causal answer, not a correlation.
Pick the least certain channel
Tests cost real revenue, so they go where the model's range is widest. The range tells you which one that is.
Feed the result back
The measured effect becomes known information in the next fit, so the model is anchored to something proven.
Repeat every quarter
One test per quarter, rotating channels, tightens the whole model over time. Slow by design.
If an MMM comes with no calibration plan, ask how they know it's right. Usually the answer is "it fits the history well", which is much weaker than it sounds: any flexible enough model can fit history.
How Your Marketing Mix Model Gets Built, in 6 Steps
Most of the elapsed time is data work, not modelling. With clean weekly inputs the modelling itself takes days of analyst time. With messy inputs the same project takes weeks, because data problems look like model problems. The engagement process covers how this is scoped.
- 1
Define the outcome
Revenue, orders or qualified leads: one model, one outcome. A model asked to explain several explains none well.
Modelling - 2
Assemble the dataset
Weekly spend per channel, the outcome, and controls: seasonality, promotions, price changes, stock-outs. Clean labels matter more than any modelling choice.
Data work - 3
Fit fade and saturation
Learned per channel, using known information where it exists. Alternative approaches are benchmarked on the same data.
Modelling - 4
Test on unseen weeks
Held-out periods, not just a good fit on history. A model that recreates the past and fails on new weeks is overfitted, the default failure of MMM.
Modelling - 5
Calibrate with an experiment
The step that turns a description into a causal claim, and the one most often skipped.
Modelling - 6
Deliver and refresh
Planning outputs delivered where your planners work, then refreshed every quarter as the business and market move.
Data work
What Data Marketing Mix Modeling Actually Requires
MMM has a lower privacy burden than user-level modelling but a higher history burden. The constraint isn’t sensitive data; it’s having enough weeks, with enough change in spend, to tell channels apart.
Weekly spend by channel
Every channel with budget, offline included: TV, print, radio, outdoor, events, retail media. A channel left out doesn't get zero credit; its effect is wrongly handed to whatever it moves with.
The outcome series
Weekly revenue, orders or leads from a source whose definition didn't change midway. A tracking migration inside the window is a real problem and is handled explicitly.
Control variables
Seasonality, promotions, price changes, stock-outs and wider market factors. Without them, the model credits channels for what the calendar caused.
~104 weeks = two seasonal cycles
Seeing each peak season twice lets the model separate "it was Q4" from "it was the campaign". Less history can work with modest goals; much less usually can't. Equally important is spend variation: a channel funded the same every week gives the model nothing to learn from.
What You Get: 4 Planning Outputs, Not a Dashboard
MMM outputs are planning instruments. Each one answers a question a marketing director or CFO is already asking, and each is delivered as something a planner can work with, not a static PDF.
Contribution breakdown
How much of the outcome each channel drove, plus a baseline: what you'd have sold with no marketing. That baseline is often the most sobering number in the model.
Response curve per channel
Spend against extra sales, with your current spend marked. Shows which channels still have room and which are flat.
Scenario planning
What happens to the forecast under a given reallocation, compared as ranges, so an "improvement" inside the noise is visible as such.
Calibrated ROI ranges
Return per channel as a range, with experiment-checked channels flagged. You know which numbers are proven and which are inferred.
For revenue forecasting beyond channel planning, see sales forecasting; for one view across every channel and data source, see Omnichannel Data Intelligence.
Who Marketing Mix Modeling Is Built For (and Not For)
MMM suits organisations spending across several channels, with a couple of years of consistent history, where budget allocation is a real recurring decision. Where it isn’t a fit, the audit says so and points to what would help more, usually incrementality tests and clean conversion tracking.
Good fit
- Multi-channel advertisers where the quarterly split across search, social, video, retail media and offline is a real open question
- Brands with offline or brand spend: TV, outdoor, sponsorship, retail
- Teams whose platform ROAS doesn't add up: every platform's claimed revenue sums to more than you actually made
- Privacy-constrained or regulated industries where user-level tracking is limited
- Marketing leaders defending budget to a CFO or board with evidence, not screenshots
Not a fit yet
- Advertising on only one or two channels
- Less than about two years of consistent history
- Spend that never changes week to week
- Wanting a weekly optimisation tool (that's attribution's job)
Where AI Agents Fit Into an MMM Programme
A SaaS tool is someone else’s generic model. An AI agent is your own model, kept running. MMM is unusual because the model refreshes quarterly, so most of an agent’s value sits around it: watching the data feed and the results between refreshes. The agents are autonomous, built from data science and marketing work, and every action is reviewed. More on the AI agents hub.
What the agents do between refreshes
- Flags weeks that fall outside the predicted range, so a surprise is caught early, not at quarter end
- Detects a structural break (new product line, price change, market shock) that means the model needs re-fitting early
- Names the next channel to test, the one whose range is now widest
- Checks the weekly data feed: missing weeks, renamed channels, changed definitions, before they quietly corrupt the model
What stays with a person
- Deciding whether a number is believable. A model can say outdoor returned four times its spend. Whether that's real or a seasonal artefact is a judgement about the business.
- Choosing what to test. Holdout regions cost real revenue. Which channel is worth that cost this quarter is a commercial call.
- Refusing to reallocate on thin evidence. When two scenarios overlap inside their ranges, the right move is usually to wait and test. No optimiser recommends doing nothing.
- Calling it off. If the data can't support a credible model, someone has to say so instead of shipping one anyway.
Marketing Mix Modeling: Frequently Asked Questions
What is the difference between MMM and attribution?
Attribution tracks individual users and shows which touchpoints came before a sale; it needs user-level data and sees digital only. Marketing Mix Modeling uses weekly totals to estimate how much each channel caused, offline included. MMM sets the quarterly budget; attribution fine-tunes inside it week to week.
How is MMM different from Google Analytics?
Google Analytics reports what happened on your website among users it could see. MMM estimates what each channel caused across your whole marketing programme, including TV, print, retail and digital activity that consent or tracking prevention hid from Analytics.
Do you use an MMM tool or platform?
No. Each model is custom-built on your raw data with advanced data science, benchmarked against alternative approaches before one is chosen, and supervised and reviewed by me at every refresh. You own the model, the code and the outputs.
How much data do we need?
Roughly two years of weekly data, so the model sees each season twice. Just as important is variation in spend. Less history can work with narrower goals; the audit checks this against your actual data instead of applying a fixed rule.
How long does an MMM build take?
Mostly as long as the data takes. With clean, consistent weekly inputs the modelling is a matter of days; with scattered or inconsistently labelled data it takes considerably longer. The audit tells you which situation you’re in before any timeline is committed.
Can MMM work for a small business?
Usually not, and it’s more useful to say so plainly. MMM needs several channels, real spend variation and about two years of history. A business on one or two channels will get more from incrementality tests and clean tracking.
Is MMM privacy compliant?
It works on weekly totals with no user-level records and no cookies, which makes it well suited to privacy-constrained and regulated industries. Your consent framework and legal basis for collecting the underlying data remain yours to own.
Can you guarantee better ROI from the model?
No. Results depend on product, pricing, market and competition. What is committed to is a model checked on unseen weeks and real experiments, with honest ranges, so budget decisions rest on evidence instead of platform self-reporting.
Do you work with brands outside Pakistan?
Yes. Delivery is remote from Lahore, for brands across Pakistan, the UK, USA, UAE, Canada and Turkey. Each model is fitted on the market it will be used in. Working hours overlap comfortably with the Gulf and the UK, and partly with US mornings.
AI-driven digital marketing intelligence consultant. 12+ years in digital marketing, 100+ clients across 15 industries and 6 markets, a Master's in Computer Science and MPhil / MS Data Science (AI). Every marketing mix model is built, checked and reviewed by me personally, with no hand-off to an account manager.
Where to Go Next
Find Out Whether Your Data Can Support a Credible Model
A data audit checks your channel history, spend variation and control variables, and tells you plainly whether an MMM would produce something worth acting on, including when incrementality tests would serve you better first.