Meridian vs. Robyn vs. Hiring an MMM Expert: What the Free Tools Can and Cannot Do

When I built Upwork's first marketing mix model, covering an $88M annual media portfolio, the statistics were the smaller part of the job. Most of the work was getting clean, consistent data from every channel into one place, agreeing on what we were actually measuring, and then explaining results that did not match what each ad platform was reporting about itself. That experience is why I am glad Google and Meta have made marketing mix modeling free, and also why I am not surprised when teams download the tools and get stuck.

If you have tried Meridian or Robyn and the project has stalled, you are in good company. Both are serious, well-built tools. Neither one removes the hard parts of MMM, and the newest AI tools do not remove them either. Here is how the two compare, what AI can now do to help, what most of these models are missing, and where you still need someone who has done this before.

1. What are Meridian and Robyn?

Both are free, open-source marketing mix modeling tools built by the two largest ad platforms. Meridian comes from Google. It became available to everyone in January 2025, runs in Python, and uses a Bayesian approach, which lets you build in what you already know from past experiments before the model looks at your data. In February 2026 Google added a no-code Scenario Planner on top of it, so marketers can test budget splits and see estimated returns without writing code.

Robyn comes from Meta's marketing science team. It was built in R, and there is now a Python version that Meta describes as a beta translated with the help of a language model. Robyn uses ridge regression with an automated search across thousands of model versions, then gives you a set of candidate models to choose from.

2. How do Meridian and Robyn differ?

The practical difference is in how much you have to decide up front and how much the tool decides for you. Meridian asks you to set priors, which are your starting assumptions about how each channel performs. That is powerful if you have good experiment data and know how to use it, and risky if you do not, because the priors shape the answer. Robyn automates more of the search and hands you several models that fit the data about equally well. Someone still has to choose among them, and that choice can change your budget recommendation.

Meridian is generally the better fit for teams with Python skills, geographic data, and lift tests they want to build in. Robyn has a longer track record in the open-source community and suits teams that already work in R. Either can produce a credible model in the right hands. Either can produce a confident-looking wrong answer in the wrong ones.

3. Why do so many teams get stuck with the free tools?

Most teams stall in the same places. The first is data. MMM needs at least two years of weekly data, spend and results for every channel, and the sales numbers from your finance system rather than what each platform claims. Pulling that together and making it consistent usually takes longer than the modeling.

The second is the model itself. Channels that always move together, such as search and social spend that rise at the same time every holiday season, are hard for any model to separate. Too little variation in spend makes estimates unstable. Settings that look like minor technical choices can swing the result for a channel from hero to waste.

The third is trust. The first model almost always disagrees with what the ad platforms report. Someone has to decide whether the model is right, the platforms are right, or neither, and then explain that to a CFO who will ask hard questions.

4. Can AI pull your marketing data for you now?

Yes, and that genuinely helps with the first problem. Tools like Supermetrics now connect directly to ChatGPT and Claude, so you can pull spend and results from Google Ads, Meta, LinkedIn, TikTok, GA4, and other platforms in one conversation instead of exporting files one by one. I wrote about this in my recent post on how to use AI in marketing now. AI can also write and debug the Python or R code to load the data, run Meridian or Robyn, and chart the output.

That saves real time. It also creates a new risk, because it makes it easy to feed a model only the data that is easy to pull.

5. Why can you not build a marketing mix model from inside one channel?

A marketing mix model only works if it sees the whole mix. Ad platform data is one piece. The model also needs your actual sales or revenue from your own systems, pricing and promotions, seasonality, offline channels like TV, events, direct mail, and PR, organic and referral traffic, and outside factors like competitor activity and the economy. Leave those out and the model will hand their credit to whatever channels you did include, which usually means paid digital looks better than it is.

The same logic applies to who runs the model. Every ad platform's own reporting gives itself credit, and each one grades its own homework. Meridian and Robyn are open-source and built to model every channel, but the person setting the priors and choosing the final model should not be someone whose budget, commission, or agency fee depends on one channel winning. You cannot measure the mix fairly from inside one piece of it, whether that piece is a data source or a job title.

6. What outside data are most marketing mix models missing?

In my experience, this is where I add the most value to a model, and it has little to do with the software. It comes from pulling in data that an AI tool will not have access to and usually will not think to ask for. Your sales do not move only because of your marketing. They move with the economy, the weather, the news, and everything else your customers are dealing with that week. If the model cannot see those forces, it will give their effect to whatever marketing happened to be running at the same time.

Economic conditions are the first thing I look at. Consumer sentiment, unemployment, interest rates, gas prices, and indicators specific to your category, such as housing starts for a home services company, can all shift demand. Most of these are free to pull from public sources like the Federal Reserve Bank of St. Louis's FRED database. A brand that grew during a strong economy can easily credit its paid social program for growth that really came from customers having more money to spend.

Weather is the second. For anything seasonal or tied to location, such as home services, apparel, beverages, or outdoor activities, a cold snap or a rainy month can move demand more than a budget change does. NOAA publishes historical weather data by location, and building it in by region often explains spikes and dips nobody on the team could account for.

Then there are one-time shocks. Years ago I built models that had to account for the impact of COVID, which changed shopping behavior, media prices, and channel performance almost overnight. A model that does not account for an event like that will either treat it as noise or, worse, credit the marketing that happened to be running during it. Stockouts, supply problems, price increases, a competitor's big launch, and a story that goes viral all belong in the same group.

PR is one of the most common blind spots I see. When a story about your company runs, or your founder lands on a popular podcast, branded search and direct traffic jump. If the model cannot see the PR, it hands that lift to paid search, which already looks strong because it captures people who were going to search for you anyway. I ask clients for a record of press coverage, earned media mentions, and major announcements, even if it is just a spreadsheet of dates.

Linear TV is the other big one. Many brands still spend real money on broadcast and cable, but that spend usually lives in agency invoices and media plans, reported in ratings points or impressions, rather than in an ad platform an AI connector can reach. Leave it out and you lose one of the largest lines in the budget, and its effect gets scattered across search, social, and direct traffic. Getting TV into the model usually means collecting weekly flight dates and spend by market from the agency and lining them up with everything else. Streaming and connected TV can have the same problem when they are bought through a separate platform.

7. What statistical checks should you run before you build the model?

This is the step that gets skipped most often, and it is where a lot of bad models start. If you ask a general AI assistant to load your data and run Meridian or Robyn, it will often go straight to fitting the model. Meridian now includes an exploratory data analysis report that flags some problems, but its default checks are set to catch only extreme cases, such as two variables that are almost perfectly correlated. Robyn checks that your inputs are formatted correctly and handles trend and seasonality for you, but most of the diagnostic work is left to the analyst. The checks I run before modeling are what tell me whether the model can answer the question at all.

I start by plotting every series over time. That catches missing weeks, gaps where spend dropped to zero, outliers, and sudden jumps that turn out to be a tracking change or a reporting error rather than real performance. Next I check how much each channel's spend actually varied. If a channel spent about the same amount every week, no model can tell what it contributed, and it is better to know that before the model gives you a confident-looking number.

Then I test for multicollinearity, which means channels that rise and fall together. I look at the correlation matrix and variance inflation factors for every variable, and I set much tighter limits than the tools use by default. When two channels move in lockstep, the choice is to combine them, add an outside variable that separates them, or run an experiment that breaks the pattern. I also test whether the sales series and the main drivers are trending over time, using stationarity tests like the Augmented Dickey-Fuller test, because two things that both grow over time will look related even when they are not.

I look for structural breaks, which are points where the relationship between marketing and sales changed, such as a pandemic, a pricing change, a website relaunch, or a switch in how conversions were tracked. I check lags with cross-correlations to see how long it takes for spend in each channel to show up in sales, which gives me a starting point for the carryover settings rather than accepting the defaults. And I count the number of variables against the number of weeks of data, because a model with too many inputs and too little history will fit the past well and predict the future badly.

After the model is fit, the checks continue. I look at whether the errors follow a pattern over time, how well the model predicts weeks it was not trained on, and whether the results hold up when I change reasonable assumptions. A model that only works under one set of choices is not ready to move budget.

8. What can AI still not do in marketing mix modeling?

AI can write code that runs. It cannot tell you whether the model that comes out of that code is right. It does not know that your sales jumped because of a pricing change, a distribution deal, or a competitor's stockout. It cannot decide which factors belong in the model, what reasonable starting assumptions are for your category, or when a result is statistically tidy and commercially absurd.

It also cannot design the experiments that keep a model honest. The strongest MMM programs use geo tests or holdout tests to check the model's estimates, then feed those results back in. Deciding what to test, where, and for how long is a judgment call about your business. So is the last step, which is the one that matters most: turning model output into budget decisions that leadership will actually approve and act on.

9. When should you use Meridian or Robyn on your own?

The free tools are a good choice when you have an in-house data scientist or analyst who is comfortable in Python or R and knows how to run the checks above, at least two years of clean weekly data across channels, meaningful variation in spend over that time, and enough marketing investment that a better allocation is worth the effort. If that describes your team, start with whichever tool matches your team's language, and plan to validate the results with at least one controlled test before moving serious budget.

10. When does it make sense to hire an MMM expert?

It makes sense when the project has stalled, when the results do not make sense to anyone in the room, when you need a model your CFO and board will trust, or when you do not have the in-house skills to build and maintain one. It also makes sense when you already have a model and are not sure it is right. I have rebuilt models that looked polished and were quietly crediting the wrong channels, usually because the data checks were skipped or the outside factors were never in the model.

An outside expert can build the model with you, audit and rebuild one you already have, or coach your team so they can run it themselves going forward. Whichever path you choose, the goal is the same: a model that sees your whole business, that someone with no stake in any one channel has checked, and that leads to budget decisions you can defend.

If you have tried Meridian or Robyn and hit a wall, or you want a second opinion on a model you already have, that is exactly the kind of work I do. You can book a free 30-minute introductory call here.

Dr. Linda Orr, PhD, is the founder of Orr Consulting, a fractional CMO and marketing analytics practice serving brands from $5M to $500M in revenue. She built Upwork's first enterprise marketing mix model, covering an $88M annual media portfolio, and spent about 20 years as a marketing professor before moving fully into consulting.