What Data Do You Actually Need to Run a Marketing Mix Model?
Marketing Mix Modelling is only as reliable as the data behind it. Before running your first model, you need an honest assessment of whether your dataset is ready — and a clear plan for what to do if it isn’t.
The data requirements for MMM are not optional or adjustable. They exist because the model needs sufficient variation in your historical spend and outcomes to distinguish each channel’s contribution from background noise and the effects of all other channels.
In this guide, you’ll learn the minimum viable dataset for MMM, what “clean, aggregated weekly data” actually means in practice, which channels are hardest to model and why, how to diagnose your own data readiness, and what to do when your dataset falls short.
The Minimum Viable Dataset for MMM
There are three dimensions of data requirements: volume (how much history), history (how far back), and granularity (how finely resolved).
Volume and History
Minimum requirement: 52 weeks (one year) of data across your KPI and all media channels.
Working requirement: 104 weeks (two years) of data.
Why two years is better: A single year of data may capture only one seasonal cycle. An MMM model that has only seen one Christmas, one summer, or one annual product launch cycle cannot confidently distinguish between the effect of your November advertising and the effect of November itself. Two or more years gives the model the seasonal variation it needs to separate media effects from calendar effects reliably.
Directional threshold: A model with 52–78 weeks of data can produce directional insights — it may correctly identify which channels are relatively over- or under-performing — but the contribution estimates will have wider uncertainty and should not be used as the sole basis for budget reallocation at a precise level.
Granularity
MMM requires weekly aggregated data. Daily data can be aggregated to weekly by summing spend and averaging price/KPI variables. Quarterly or monthly data is not suitable — the temporal resolution is too coarse to detect Carry-Over Effects and within-quarter variation.
If your primary KPI (revenue, sales units) is only available monthly, you have a fundamental data availability problem that cannot be resolved by changing your modelling tool.
What “Clean, Aggregated Weekly Data” Actually Means in Practice
The phrase “clean, aggregated weekly data” hides a significant amount of preparation work. Here’s what it means for each data category.
KPI Data (Revenue or Sales)
Your primary outcome variable should be:
- Your own business outcome — not a platform-reported KPI. Revenue or units sold from your own system, not attributed conversions from Google Ads or Meta Ads Manager.
- Weekly totals — the sum of all revenue or sales in each calendar week.
- Consistent definitions — if revenue includes returns in some weeks and excludes them in others, the inconsistency will confuse the model.
- No imputation — if there are genuine gaps, document this as a context variable rather than filling in guesses.
Media Spend Data
Each channel’s spend must be:
- Aggregated to the same weekly grain — all platforms report on their own date conventions; you need to normalise these to consistent calendar weeks.
- Inclusive of all spend — if a channel has sub-accounts, campaigns, or countries, you need the total for the scope of your model. Partial spend will under-attribute that channel.
- Expressed in the same currency and unit — never mix actuals with thousands, or GBP with USD, in the same column.
- Zero for weeks with no spend — not blank, not NA.
External Variables (Context Variables)
Promotional flags: If your business runs price promotions, discount events, or sales periods, include a binary or continuous variable representing promotional intensity week-by-week. Without this, the model will attribute promotional sales lifts to whatever media happened to be running that week.
Holiday and seasonality: Robyn includes built-in seasonality decomposition, which handles standard seasonal patterns. Add custom flags for events unique to your business — major industry trade shows, annual product launches, or market-specific public holidays that affect your category.
Competitor activity (optional but valuable): If you have data on competitor spend or promotional activity (via media intelligence tools), including it reduces the noise in your model and improves the precision of your own channel attribution.
Which Channels Are Hardest to Model — and Why
Not all channels are equally modellable. Here are the channel types that consistently cause problems in MMM, and the reason why.
New Channels (Under 20 Weeks of Active Spend)
Problem: The model needs variation in a channel’s spend — weeks where spend goes up, and weeks where spend goes down — to estimate its contribution. A channel that has only been running for 12 weeks provides almost no useful variation.
Recommendation: Do not include channels with fewer than 20 weeks of meaningful spend history as primary modelled variables. Include them as context variables or leave them out and note the limitation.
Low-Spend Channels (Under 5% of Total Media Budget)
Problem: Ridge regression applies a regularisation penalty that shrinks small coefficients toward zero. For channels that represent a small fraction of total media investment, the model’s default behaviour is to minimise their contribution.
Recommendation: Treat any channel accounting for less than 5% of total media spend as directional at best. Their contribution estimates are a useful indicator but should not be used as precise measurement.
Channels with Flat, Constant Spend
Problem: If a channel ran at exactly the same spend level every week for two years, there is no variation to model. The model cannot determine whether that channel contributed to sales because there are no weeks where it was absent to compare against.
Recommendation: Channels with near-constant spend should be flagged as potential problems before the model runs. If the spend is truly invariant, it may need to be treated as a component of the Organic Base rather than a separately modelled media variable.
Very Short-Lifecycle Campaigns (Under 4 Weeks)
Problem: A TV burst that ran for three weeks provides three data points. Even with Carry-Over modelling, the model has almost no signal to work with.
Recommendation: Include short campaigns as binary context variables (1 = campaign running, 0 = not) rather than attempting to model Spend Efficiency.
How to Diagnose Whether Your Data Is Ready
Run this checklist before starting your MMM configuration.
Data Readiness Checklist
History and Volume
- At least 104 weeks of continuous weekly data (2 years)
- No gaps of more than 2 consecutive weeks (and any gaps are documented)
- KPI data sourced from internal systems, not platform-reported metrics
Media Spend
- All channels aggregated to the same weekly date grain
- Currency and units consistent across all channels
- Zero values (not NA) for weeks with no spend
- All relevant sub-accounts and campaigns included in each channel total
External Variables
- Promotional periods flagged (at least major events)
- Any known market disruptions documented (competitor launches, external shocks)
Channel Eligibility
- All included channels have at least 20 weeks of spend history
- No channel has constant, invariant spend across the full period
- Channels with under 5% of total budget are flagged as low-confidence
What to Do When Your Dataset Is Too Small or Too Sparse
| Problem | Option 1 | Option 2 |
|---|---|---|
| Too few weeks of history (< 52 weeks) | Wait 6–12 months and re-evaluate | Run a directional pilot model with explicit uncertainty caveats |
| New channel (< 20 weeks of spend) | Exclude from model, include as context variable | Document as a limitation; reintroduce in next model refresh |
| Channel spend is too low (< 5% of budget) | Merge with related channel | Exclude and note the gap |
| KPI data only available monthly | Use weekly proxy (e.g., website sessions, organic traffic) with caution | Consider an annual model rather than weekly |
| Missing promotional data | Reconstruct from internal records or invoice history | Include as an unmodelled seasonal component with caveats |
The honest conclusion in some cases is: the dataset is not yet ready for a reliable MMM. Running a model on thin data produces misleading outputs faster than not running a model at all. Building trust with clients or stakeholders means being transparent about data limitations rather than presenting an unreliable analysis as a definitive view.
Frequently Asked Questions
How much data do I need to run a marketing mix model?
A minimum of 52 weeks (one year) of weekly data is required for a basic MMM run, but two years (104 weeks) gives the model the seasonal variation it needs to reliably separate media effects from calendar effects. With fewer than 52 weeks, outputs should be treated as directional only.
What format does MMM data need to be in?
MMM data must be formatted as a weekly time-series table: one row per week, with columns for the KPI (e.g., revenue), each media channel’s spend, and any external variables (promotions, seasonal flags). Dates must be in a consistent format (YYYY-MM-DD), spending figures must be in consistent units and currency, and zero must be entered (not NA) for weeks with no spend on a channel.
Can I run MMM with data from just one or two channels?
Yes, technically — but the analysis is more limited. With only one or two channels, MMM can still estimate contribution and Spend Efficiency, but it cannot decompose the interaction between multiple competing channels or identify which channel is cannibalising another. The more channels included (with sufficient spend history each), the more useful the strategic reallocation recommendations become.
What is the minimum spend level for a channel to be reliably modelled?
There is no universal minimum spend threshold, but as a practical guideline, channels representing less than 5% of your total media budget are consistently under-attributed by ridge regression models. For a business spending £500,000 per year on media, a channel receiving £25,000 or less may not produce reliable MMM contribution estimates. These channels can still be included, but their estimates should be flagged as low-confidence.
Can MMM model offline channels like TV and out-of-home?
Yes — one of MMM’s core advantages over digital attribution is its ability to model offline channels. TV, radio, out-of-home, and print spend can all be included as weekly spend variables, with Carry-Over Effects applied to capture their delayed impact on sales.
What if I don’t have two years of data?
Run the model with the data you have, but set realistic expectations with your client or stakeholder. With 52–78 weeks, you may see correct directional channel rankings — the model may correctly identify which channels are relatively over- or under-performing — but precise percentage contribution estimates and budget optimiser recommendations should be treated with additional caution. Document the limitation explicitly in your output.
Know Your Data Before You Run Your Model
The most common cause of unreliable MMM outputs is not a modelling error — it’s a data problem that wasn’t caught before the model ran.
MMM Pilot validates your data automatically on upload, flagging format issues, date gaps, and potential channel eligibility problems before you invest time configuring a model. Upload your dataset to see where it stands.