Know what really sells.
Marketing mix modelling (MMM) uses your own sales history to work out how much each channel actually adds, after seasonality, pricing and promotions are taken out. We use it to move budget from channels that are saturated to the ones where the next rupee still earns, then check every move with an experiment.
The usual problems, fixed properly.
Most budget plans run into the same handful of issues. Here is what changes when ivia takes over.
Most budget plans
- Budget split by last year's plan or whichever platform report looks best
- Meta, Google and Amazon each claiming credit for the same sale
- Video and brand spend cut because click-based reports cannot see its effect
- A channel scaled long after the extra rupee stopped paying back
- Marketplace, offline and word-of-mouth sales left out of the picture
With ivia
- Each channel's incremental contribution, measured against real sales
- One model across Meta, Google, Amazon, offline and organic, with no double counting
- Carryover (adstock) modelled, so slow-burn brand media gets fair credit
- Response curves showing where each channel saturates and what the next rupee returns
- Results checked against holdout and geo-lift tests before budget moves
Three ideas behind every model.
An MMM looks at months of weekly sales next to everything that might have moved them. It separates what each channel added from what would have happened anyway, then learns two things about every channel: how long it keeps working, and when it stops paying back.
Base vs incremental
Part of your sales comes from your brand, word of mouth, seasonality and repeat customers. The model separates that base from what each channel adds on top. Only the added part counts as a channel's return, which is why MMM numbers are usually lower, and more honest, than platform dashboards.
Carryover
Ads keep working after the week they run. People remember, search later, buy next payday. The model learns this decay, called adstock, for every channel.
Diminishing returns
The first rupees reach the most interested people. Each extra rupee reaches someone a little less likely to buy, so returns flatten. The model finds where each channel's curve bends.
Everything in Marketing Mix Modelling.
Platform reports tell you what each channel claims. MMM tells you what it actually adds. We model your sales history against every channel, price, promotion and season, then use the response curves to move budget to where the next rupee earns most, and confirm each move with an experiment.
- Data audit and one clean weekly dataset across channels
- Bayesian MMM with carryover and diminishing returns per channel
- Validation on held-out weeks and calibration with lift or geo tests
- Response curves, marginal ROI and a budget optimiser
- Scenario plans for bigger, smaller or reshaped budgets
- Regular model refreshes as new data and test results arrive
Our MMM playbook.
Data audit & collection
Ideally two to three years of weekly sales and spend by channel, plus prices, promotions, seasonality and other drivers. If the data is not ready, we say so and build it first.
Build the model
A Bayesian model built with open-source frameworks such as Google Meridian, Meta Robyn or PyMC-Marketing, with carryover and diminishing returns for every channel.
Validate & calibrate
Accuracy checked on weeks the model has never seen, then calibrated with lift or geo-holdout tests so channel effects match real experiments.
Optimise & plan
Response curves feed a budget optimiser: what to cut, what to scale and what a bigger or smaller budget would do, shown as ranges rather than false precision.
Test, refresh, repeat
Recommended shifts roll out as controlled tests. Results flow back into the model, which is refreshed as new data arrives.
Numbers that matter.
Targets are agreed in writing before launch, then reported every month in plain language, next to blended results across every channel you run.
Incremental ROI
Incremental revenue ÷ spendRevenue that would not have happened without the channel, divided by what it cost. Not what the platform claims.
Marginal ROI
Return on the next ₹1The return on the next rupee, not the average one. This is the number that decides where budget should go.
Carryover
Effect that lasts beyond the weekHow long a channel keeps working after the money is spent. Brand video often keeps selling for weeks.
Where should the next rupee go?
Budget lab
Move weekly spend between three channels and watch where the next rupee earns the most. The curves are illustrative, not your data. Your own model would draw them from your sales history.
| Channel | Spend | Avg ROI | Next ₹1 | |
|---|---|---|---|---|
| Google Search | — | — | — | — |
| Meta Ads | — | — | — | — |
| YouTube & video | — | — | — | — |
Filled dot: your split. Ring: the best split for the same budget. Avg ROI = revenue ÷ spend. Next ₹1 = marginal ROI, the return on one more rupee. A channel can have a fine average and a poor next rupee: that is the money MMM finds.
Three views. One answer.
No single method tells the whole truth, so we never rely on one. Experiments calibrate the model, the model sets the budget, and platform data runs the day-to-day. Where they disagree, the experiment wins.
Platform attribution
Ads Manager, Google Ads, GA4- Answers
- Which ad, keyword or creative is working right now
- Speed
- Daily
- Blind spot
- Each platform over-claims; offline, cross-device and slow-burn effects are missed
Marketing mix model
Built on your total sales- Answers
- How much each channel adds, how long it lasts and where budget should go
- Speed
- Refreshed monthly or quarterly
- Blind spot
- Cannot judge a single ad; needs history and spend that varies
Experiments
Geo-lift, holdouts, conversion lift- Answers
- Did this channel cause these sales, and how many?
- Speed
- A few weeks per test
- Blind spot
- One channel, one period at a time
Is MMM right for you yet?
MMM needs history and variety in your data. If you are not there yet, we will tell you, and start with tests that give causal answers today.
MMM readiness check
Tick everything that is true for your business today.
Start with experiments.
An MMM would mostly be guessing today. Holdout tests give you causal answers now while the data builds up.
Good questions.
Still unsure? Ask Ivvy, our assistant, or message the team directly.
A statistical model that links your sales over time to everything that drives them: spend on each channel, prices, promotions, seasonality and more. It estimates how much each channel adds on top of the sales you would have made anyway. It uses aggregated weekly data, not individual user tracking, so cookie loss and privacy changes do not break it.
Attribution follows clicks and views of individual people and gives credit to the touchpoints it can see, so it misses offline sales, cross-device journeys and slow-burn brand effects, and every platform tends to over-claim. MMM works top-down from your total sales, so it sees every channel at once but cannot judge a single ad. We use both, plus experiments to settle disagreements.
Ideally two to three years of weekly sales and spend by channel. Google's Meridian guidance, for example, recommends at least two years of weekly data for regional models and three years for national ones. Spend also needs to have varied over time. If you are not there yet, we start with clean tracking, structured spend tests and geo holdouts, and build the model as history builds up.
Yes, and this is where MMM is most useful. Total sales across your website, marketplaces and offline can be the outcome the model explains, and marketplace ads sit in the model as channels alongside Meta and Google. If you have sales by state or city, we can build a regional model and run geo tests.
It is an estimate, so we always show a range, not a single magic number. We check accuracy on weeks the model has not seen, calibrate it against lift or geo tests, and treat every recommendation as a test to confirm before scaling. Where the model is unsure, we say so.
The first model usually takes a few weeks once the data is collected. After that we refresh it on a regular schedule, typically monthly or quarterly, and after any major test, so it keeps learning from what actually happened.
Read the playbooks.
ROAS vs MER vs CAC: Which Number Should Run Your Business?
Add up every platform's ROAS and they often claim more revenue than you made. The metrics that can't double-count — and how to use them.
→Tracking That Actually Works: Pixel, Conversions API, GA4 and Consent
Every optimisation decision sits on top of your tracking. The setup we use for Meta, Google and GA4 — and the mistakes we fix most often.
→Amazon + Meta + Google: A Full-Funnel Plan for D2C Brands
Amazon, Meta and Google each do a different job. How they work together, how to split budget, and how to measure the halo between them.
→Better together.
Analytics & Tracking
GA4, Tag Manager, Meta Conversions API, Google enhanced conversions and Looker Studio dashboards: one honest view of what is working.
ExploreMeta Ads
Facebook and Instagram campaigns built on consolidated structures, broad targeting and a steady creative-testing rhythm, optimised for purchases or qualified leads.
ExploreGoogle Ads
Search, Shopping, Performance Max and YouTube: capturing high-intent demand with clean conversion data, tight negatives and brand-safe guardrails.
ExploreLet's look at
your measurement.
Start with a free, no-obligation audit. We'll show you what's working, what's wasting money, and what to fix first.