Marketing Mix Modeling (Meridian, Robyn)

Marketing mix modeling (MMM) is a statistical method that estimates how much each marketing channel adds to sales, using weekly or daily totals of sales, spending, prices and seasons rather than individual customer data. Open-source tools such as Google's Meridian and Meta's Robyn have made marketing mix modeling far more accessible to marketing teams.

• Input: Two or more years of weekly sales and spending per channel, plus prices, promotions and seasons.
• Output: Each channel's estimated contribution to sales, its return per rupee, and how returns change as spend grows.
• Privacy: It uses totals, so it works without cookies or user-level tracking.
• Coverage: It can include TV, outdoor, retail and quick commerce, not just clickable ads.
• Tools: Google Meridian and Meta Robyn are open-source packages; paid vendors also offer MMM as a service.

This lesson follows one example: a packaged namkeen and snacks brand from Indore that sells through kirana stores, supermarkets, quick commerce apps and its own website. It advertises on TV, YouTube, Meta and Google search. Most sales happen in shops where no pixel can see them, so attribution tools show only a small part of the picture.

Key Characteristics of Marketing Mix Modeling

• Top-down: It works from totals, while attribution models work from individual journeys.
• Base and incremental: It separates the sales that would happen without marketing, called the base, from the extra sales each channel adds.
• Adstock: An ad's effect carries over into later weeks. A TV campaign in one week still lifts sales in the next.
• Saturation: Each extra rupee on a channel adds less than the one before, because the most responsive people are reached first.
• Periodic: Models are usually refreshed every few months, not every day.

How Marketing Mix Modeling Works

In plain English, the model says that sales in a week equal a base level, plus what each channel added, plus the effect of seasons and prices, plus some random noise.

Weekly sales = Base
             + TV effect + YouTube effect + Meta effect + Search effect
             + Festival and season effect + Price and promotion effect
             + Noise

Each channel effect = strength x saturation( adstock( spend ) )

1. Collect data: The brand gathers two years of weekly sales from distributors, supermarkets, quick commerce and its website, along with weekly spend on each channel, prices, discounts and festival dates such as Diwali.
2. Transform spend: Adstock spreads each week's spend into later weeks, and a saturation curve reduces the effect of very large spends.
3. Fit the model: The tool finds the channel strengths that best explain the ups and downs in sales. Tools built on Bayesian statistics, such as Meridian, also let analysts add prior knowledge, such as results from lift tests.
4. Check the model: Analysts test whether it predicts weeks it has not seen, and whether results make business sense.
5. Use the outputs: Contributions and response curves feed budget planning, as covered in marketing budget planning.

Example: The Indore Namkeen Brand

The model splits a typical week's ₹50 lakh of sales like this (an illustration of how the output reads, not real data):

• Marketing's share: ₹50 lakh minus the ₹32 lakh base is ₹18 lakh from marketing, the sum of 6 + 4 + 5 + 3.
• Return per rupee: TV returns ₹6 lakh divided by ₹8 lakh, or 0.75 rupees of sales per rupee spent, while search returns ₹3 lakh divided by ₹1 lakh, or 3.0.
• Careful reading: These are sales, not profit. After product costs, a channel may need a much higher figure to be profitable.
• Response curves: Search looks best per rupee, but its curve may already be flat, meaning more budget would add little. TV may build the brand's base over months, which a weekly model can understate.
• Next step: Before cutting TV, the brand runs a geo test in a few matched cities, as described in incrementality testing, and feeds the result back into the model.

The Open-Source Tools

• Google Meridian: Google's open-source MMM framework in Python. It uses Bayesian methods, supports calibration with experiment results, and can model data at a regional level.
• Meta Robyn: An open-source MMM package from Meta's marketing science team, used mainly in R. It automates many modelling choices and searches many candidate models, then helps the analyst pick a sensible one.
• Both: Open-source does not mean simple to run. Each needs a skilled analyst, clean data and time to check results.

Benefits of Marketing Mix Modeling

• Covers offline and online: TV, shelves, outdoor and digital channels sit in one view.
• Privacy-friendly: No personal data is needed, which suits cookieless marketing.
• Budget planning: Response curves show where the next rupee is likely to work hardest.
• Includes context: Prices, festivals and competitor activity can be modelled alongside ads.

Limitations of Marketing Mix Modeling

• Needs history and variation: If spending never changes, the model cannot tell channels apart.
• Slow: It reports on weeks and months, not individual ads or creatives.
• Correlation risk: If TV and YouTube always rise together during festivals, the model may split their credit wrongly.
• Skill required: Results depend heavily on choices made by the analyst.

How AI Changes Marketing Mix Modeling

What AI Automates Now

Modern MMM tools automate model search, test many settings for adstock and saturation, and produce budget scenarios. AI assistants can help write data cleaning code, explain outputs in plain words and draft summaries for management, as shown in analyze marketing data with AI.

What Still Needs a Human

People must gather reliable sales data from distributors and apps, record events such as stock-outs and price changes, and judge whether results make sense. Only the team knows that sales fell in one month because a factory shut down, not because an ad stopped.

Risk to Watch

An MMM always produces numbers, even from poor data. A confident AI summary of a weak model can move crores of budget in the wrong direction. Check the model against lift tests before acting on it.

Do It with AI

Use this prompt to check whether your data is ready for an MMM. It works in ChatGPT, Claude or Gemini.

1. Build a weekly table of sales, spend per channel, prices and events.
2. Run the prompt and fix the data gaps it points out.
3. Work with an analyst to run Meridian, Robyn or a vendor model.
4. Check results against a lift test before changing the budget.

Check Before You Use It

• Facts: Confirm every spend and sales figure against finance records before modelling.
• Brand fit: Judge channel results against the brand's long-term goals, not only weekly sales.
• Compliance: Share only aggregate data with AI tools and vendors, and follow the company's data policy.