Anastasia Braitsik is a powerhouse in the digital marketing landscape, recognized globally for her mastery of SEO, content strategy, and the increasingly complex world of data analytics. As privacy changes and signal loss reshape how we understand consumer behavior, she has become a leading voice in navigating the “renaissance” of Marketing Mix Modeling (MMM). Her approach combines rigorous statistical discipline with a deep understanding of organizational psychology, ensuring that data doesn’t just sit in a dashboard but actually drives high-stakes budget decisions. In this discussion, we explore the democratization of sophisticated measurement tools and the hidden challenges that teams face when moving from simple attribution to top-down modeling.
The traditional barriers to entry for Marketing Mix Modeling were once defined by massive price tags and exclusive consulting contracts. How has the emergence of open-source libraries shifted this landscape, and does a “free” tool truly mean the end of the gatekeeping era?
The shift we are seeing is nothing short of a revolution, as the entry point for high-level measurement has moved from a boardroom negotiation to a code repository. For a long time, the $150,000 to $500,000 consulting gate was the only way into MMM, effectively locking out all but the largest enterprise players. Today, nearly half of U.S. marketers—about 46.9%—are planning to increase their investment in MMM because they recognize it as the most reliable measurement methodology available, a sentiment shared by over 27% of the industry. However, there is a massive difference between a free tool and a free model. While any team with R or Python skills can now access the same math used by tech giants, the domain expertise required to tune that math remains rare and expensive. You can download the code for nothing, but the “data archaeology” and the strategic interpretation required to make that code actionable still carry a significant intellectual cost.
Tools like Robyn, Meridian, and PyMC-Marketing offer different levels of accessibility and rigor. When you are looking at these three heavyweights, how do you distinguish their strengths for a team trying to decide where to plant their flag?
Each of these libraries represents a different philosophy of measurement, and choosing the wrong one can lead to a very frustrated data science team. I personally spend a lot of time with Meta’s Robyn because it is the most approachable entry point for those who want automation; it uses Nevergrad for hyperparameter search and provides these beautiful Pareto frontier plots that help you visualize the trade-offs between different model solutions. Then you have Google’s Meridian, which is built on Python and TensorFlow and leans heavily into Bayesian inference with geo-level priors, offering a more rigorous but steeper learning curve. If you are looking for absolute flexibility and something that feels like an academic-grade probabilistic model, PyMC-Marketing is the gold standard, though it demands the highest level of statistical fluency. Each tool essentially asks how much of the “black box” you want to open and how much of the underlying math you are prepared to defend when a CFO starts asking questions.
We often hear that the model is the star of the show, but you’ve mentioned that “data archaeology” is the silent killer of these projects. Could you walk us through the gritty reality of what a team actually faces before they can even run their first script?
The vendor demos always make it look like you just plug in a CSV and get a budget recommendation, but the reality is a grueling six-week scavenger hunt through fragmented systems. To build a model that actually makes sense, you need a baseline of two to three years of weekly data to capture seasonality and spend variations, and that data is almost never in one place. You’ll find that Finance owns the revenue data, a brand team has the TV flight dates in a PDF, an agency has the digital spend, and some trade promotion history is living in a spreadsheet a former employee built in 2021. You aren’t just looking for “digital” spend; you need it broken down into search, social, display, and video to get any real granularity. This phase is often where projects die because it requires a level of cross-departmental cooperation and data hygiene that most organizations simply haven’t maintained.
As AI assistants become more capable at writing code, we’ve seen a rise in what you call “vibe coding” in measurement. Why is it dangerous to rely on AI for the technical scaffolding of an MMM, and where does the “human-in-the-loop” become non-negotiable?
AI is a fantastic co-pilot for handling syntax—it can help you scaffold a Robyn run or debug a Meridian configuration in seconds—but it has zero intuition for the nuances of your specific business. An AI doesn’t know if your Nevergrad optimizer has truly converged or if it’s just stuck in a local minimum, nor can it decide where to sit on the Pareto frontier when you have to trade off between NRMSE and decomposition error. If you let an AI “vibe code” its way through adstock transformation parameters like Weibull shape and scale, you might end up with a model that looks statistically sound but suggests your search ads have a six-month carryover effect, which is absurd. Human experts are the ones who have to diagnose why a model is giving an implausible contribution to a channel and decide whether to fix it with a prior, a data correction, or a complete variable exclusion. The scripting is the easy part; the judgment calls regarding incrementality and calibration are where the real work happens.
How does an expert take a cold, mathematical output and breathe life into it by encoding business context that a machine could never infer from raw numbers?
The most technically perfect model is essentially worthless if it exists in a vacuum away from the actual “boots on the ground” experience of the marketing team. We have to manually account for structural breaks like a pricing crisis in Q3 or the massive troughs and peaks caused by external macro disruptions that the data alone can’t explain. For example, a model might say your TV contribution is 40%, but if you only spent $2 million on TV, an experienced marketer’s intuition will tell them that “feels wrong” and needs a deep dive investigation. We have to encode things like the fact that a branded awareness campaign has a decay spanning months while paid search might only have a three-day carryover. Ultimately, the role of the expert is organizational translation—taking the math and explaining to a CMO or CFO why shifting 15% of the search budget to CTV is a move based on reality, not just a line on a graph.
What is your forecast for the future of Marketing Mix Modeling over the next three years?
I believe we are heading toward a world where MMM is no longer a “once-a-year” strategic exercise but a real-time decision-support system integrated directly into the media buying workflow. As privacy regulations continue to degrade the accuracy of tactical attribution, the 46.9% of marketers currently investing in MMM will likely swell to a vast majority, forcing a consolidation in the vendor landscape where only those who provide deep transparency will survive. We will see the “SaaS layer” on top of open-source tools become much more sophisticated, allowing non-technical brand managers to run “what-if” scenarios while the underlying Bayesian engines, like Meridian, handle the heavy lifting. However, the premium on human talent will only go up; the ability to bridge the gap between data archaeology, statistical rigor, and business storytelling will become the most valuable skill set in the entire marketing department.
