This is an honest guide to Meridian, the open source MMM (Marketing Mix Model) by Google. It is written by a PhD in statistics who, as it happens, has also done the job of a marketing manager.
This is a comprehensive guide for marketing managers and for anyone interested in getting to know the method better. We will try to avoid repeating information you can already find on the official website (https://developers.google.com/meridian). Our effort here is to answer the most pressing questions about MMM, specifically for marketing managers.
What is Meridian in the context of MMM?
Meridian, in the context of Marketing Mix Modeling, is Google’s open source library for building Marketing Mix Models. It is Google’s bet to capture the growing demand for statistical models that tell you which marketing activity is worth your next advertising money.
What is MMM?
A Marketing Mix Model is a statistical model. From one side you provide your business data, the model aggregates and combines all of it, and it gives you back the true value of your marketing activities. And this is exactly where most marketing managers get lost, on what the MMM output and result actually mean.
MMM aggregates your media. Practically, that means it combines your costs and activities across your entire marketing ecosystem. That is the first step. Then it tries to assign the right credit to each activity, and each unit of advertising spend, for your sales.
In theory, an MMM is exactly what every marketing manager is missing. Modern MMMs are also equipped with several other features such as scenario planning and budget allocation, among others.
For more information and a deep dive, go here: https://mar-sci.com/what-is-a-marketing-mix-model/
How does Meridian MMM work?
The official website provides plenty of information here. Again, we will try to give content for marketing managers.
Think of weather forecasts. You open your phone and check the temperature for any given city for the rest of the week. The temperature comes as a range, the lowest and the highest, but you still get the middle point. And, just for the record, we are not meteorologists, but weather forecasts are generated from past data.
For instance, humidity, seasonality, wind speed and many others. A model takes those inputs, tries to understand the impact of each one, and produces the final temperature for tomorrow. Congratulations, you now know how Meridian MMM works.
The weather forecast represents your sales, the business outcome you care about. An MMM works with sales in any currency, or with any other number such as purchases or leads.
The elements that could impact the weather represent your marketing activity. Humidity can be your Facebook cost, wind speed your organic (SEO) sessions, and seasonality your product discounts.
The model has the ability to understand, for any given sale, how much your Facebook cost influenced a consumer, and whether product discounts worked collaboratively with the ads for the purchase.
What this chart shows. This is what an MMM output looks like in practice, on illustrative data. Total weekly sales are split into layers. The grey layer at the bottom is the baseline, the sales that would have happened anyway. Every coloured layer on top is the extra sales one channel added that week. Notice how TV appears in bursts, this is what a campaign on and off pattern looks like once the model separates it from everything else.
Can I trust Meridian MMM?
Great question. The short answer is yes, with a few caveats of course. Would you trust the weather forecast? The answer is yes. Can a weather forecast make mistakes? Absolutely yes.
There is a technical answer to whether you should trust the results of a Meridian MMM. For folks familiar with similar methods, Meridian MMM ships with model diagnostics, measurements related to the model’s ability to do a good and accurate job. There are also ways to validate what the MMM suggests with experimentation. For marketing managers, diagnostics and validation methods mean little to nothing, and that is absolutely fine. What matters is that they exist, and that whoever runs your model actually checks them.
What this chart shows. Just like the weather forecast, Meridian never gives you a single number. Every result comes with a range. Here, on illustrative data, the dot is the best estimate of how much revenue each euro spent returns, and the bar around it is the range the model considers credible. Google Ads is clearly profitable, its whole range sits above the break even line. TV might be profitable, but its range crosses the line, so the honest answer is that the model is not sure yet. A vendor who only shows you the dots, without the ranges, is hiding half the answer.
What is incrementality in the context of MMM?
Data scientists, and the industry overall, often refer to Marketing Mix Models as a method to identify incrementality. Some others refer to it as the true impact of advertising. Frankly, none of it makes much sense at first, but we will try to explain it for marketing managers.
When data scientists talk about incrementality, they essentially talk about the baseline of the MMM. Okay, let’s take a step back. What is the baseline? It is the chunk of sales the MMM says would have happened anyway, without ads. You get a baseline by default with your Meridian MMM.
Once the baseline (sales) is established, you know how much each channel brings on top of it. That extra part is what makes it “incremental”. Look at the first figure again, everything above the grey layer is incremental.
What data do I need for Meridian MMM?
There are three types of data you need for a successful Meridian MMM:
- Marketing data
- Sales data
- Control data
Marketing data. This is the easiest type to obtain as a marketing manager. There are two sub-categories for a successful Meridian MMM. First, you need the cost of each activity, such as Facebook Ads, Google Ads and all the other channels you actively invest in. You should also include marketing activity and media even where you have no cost, such as organic website visits. Second, you need any other marketing activity such as events, promos and anything else that qualifies as marketing activity.
Sales data. Sales data has two sub-categories, similarly. The first one is, of course, the variable you care about. It can be sales in any currency, or any other business outcome such as leads or purchases. We highly recommend adding other data that can impact sales, such as promos or discounts. You could also include the availability of your products.
Control data. Controls are where most marketing managers get confused and conclude that MMM is impossible. Controls can make or break an MMM, since they impact the rest of the model. We know very well that different control data can generate different MMM results, and not only for Meridian MMM, for any MMM. Read more here: https://mar-sci.com/control-variables-marketing-mix-modeling/.
Does MMM require only cost data to operate?
No. MMM does not require only cost data, such as spend, to operate. Although we do recommend having a cost variable wherever you spend money on an activity. Why? Because that is how you learn what every euro you spend on each channel brings back.
An example of data for Meridian MMM
| Date | Facebook spend | Facebook impressions | Google Ads spend | TV spend | Organic sessions | Promo running | Sales |
|---|---|---|---|---|---|---|---|
| Week 1 | 4,200 | 610,000 | 6,800 | 0 | 12,670 | No | 92,000 |
| Week 2 | 4,500 | 655,000 | 7,100 | 0 | 12,020 | No | 101,500 |
| Week 3 | 3,900 | 588,000 | 6,500 | 15,000 | 13,340 | Yes | 118,700 |
One row per week, one column per input, and your sales at the end. That is the entire shape of the dataset. If you can build this table, you can build an MMM.
How much should I invest in marketing before trying Meridian MMM?
You might hear different stories about how much you need for sufficient data for a Meridian MMM. We would typically suggest at least 24 months of weekly data and more than 1 million in advertising spend.
In case you invest more than 1 million but you invest heavily (more than 80 percent) in a single channel, such as Facebook or Google, then Meridian MMM will not be able to help you.
The best fit for Meridian MMM is 3 to 5 million in advertising spend, across multiple channels.
What this chart shows. Two advertisers with the same total budget, on illustrative data. Brand A puts 85 percent of its money into one channel. The model has almost nothing to compare against, so it cannot separate what that channel did from what would have happened anyway. Brand B spreads its spend across five channels that move up and down at different times, and that variation is exactly what the model learns from. If your spend looks like Brand A, run experiments first, not an MMM.
How can I use Meridian MMM as a marketing manager?
There are three ways:
- Use Colab
- Run it locally
- Use MarSci’s free UI
Use Colab
Colab is the web based notebook by Google, and you can access the demo of Meridian here: https://colab.research.google.com/github/google/meridian/blob/main/demo/Meridian_Getting_Started.ipynb
You run it in the browser, cell by cell, with no installation. It works out of the box with Google’s demo data. To swap in your own data you will need someone comfortable with Python, but even running the demo once is worth it. You will see with your own eyes what the inputs and outputs of an MMM look like.
Run it locally
Meridian is a Python library, so it can run on your own machine or your company’s servers. Your data science team installs it with a single command, pip install google-meridian, and works with it like any other Python package. Fair warning, this is the data science route. Meridian is a Bayesian model and training one takes real computing power, Google recommends running it on a GPU. If you have analysts or a data science team in house, this is their job, not yours. Your job is to bring the data and the questions.
Use MarSci’s free UI
If you want to try Meridian MMM without notebooks or Python, click here and join the platform: https://mar-sci.com/solutions-marketing-mix-modeling/ or reach out at info@mar-sci.com
Q&A
What marketing channels can I include in MMM?
- Online paid, such as Facebook and Google Ads
- Online organic, such as organic social media traffic, SEO (organic) traffic and CRM
- Offline channels, such as events
- Offline paid traditional advertising channels, such as TV, radio and print
Can I include promos in Meridian MMM?
Yes you can. Promo is actually a great data input for an accurate MMM.
Is Meridian MMM free?
Yes it is. Although you might get charged by your vendor or agency for the time they put into customizing the model.
Is Meridian MMM the best MMM?
Frankly, yes. Not because it is Google’s solution, but compared to other methods, Meridian MMM is one of the best options out there.
Can I use reach and frequency for MMM?
Yes you can. You can use Meridian MMM with and without reach and frequency information for your marketing channels.
What is Bayesian theory in the MMM context?
Meridian MMM is a Bayesian model. It starts with a default hypothesis about each channel’s efficiency and updates those numbers the more it learns. The model gives you an answer on each channel’s efficiency today, and the answer improves as more data becomes available.
