A recipe generator can produce thousands of dishes. Feed it a cuisine, a set of ingredients, a flavor profile, and it will return something that looks like a recipe. It will be formatted correctly, it will have steps, and if you follow those steps, you will probably end up with food. What it cannot do is taste anything, notice that a technique doesn’t work the way a trained chef would, or develop a method that no one has ever published before. Its output is bounded entirely by what’s already been written down and put on the internet.
The same dynamic is now playing out in marketing measurement. A new category of MMM has emerged, built not through years of research and testing, but through AI-assisted code generation layered on top of existing open-source frameworks. These are called vibe-coded MMMs, and understanding what they are, what they can do, and where their limits lie is important for any marketer who relies on measurement data to make real budget decisions.
Key takeaways
- Vibe-coded MMMs are marketing mix models built using AI coding tools on top of open-source MMM frameworks like Meta’s Robyn, Google’s Meridian, or PyMC-Marketing.
- These frameworks carry inherited structural assumptions baked in from their original design, including forced saturation curves and modeling approaches that can lead to systematically biased budget recommendations.
- Because LLMs can only draw from what exists in their training data, a vibe-coded MMM will reproduce the same modeling concepts and constraints found in public-domain research, nothing more.
- There’s no mechanism for an LLM to develop a novel hypothesis about how marketing works, test it on real data, and iterate over time, which means vibe-coded MMMs are, by definition, recombinations of already-published methods.
- Vibe coding has also lowered the barrier to entry enough that the person who built your MMM may not have a background in marketing science or computer science, which makes it harder to know whether the model’s outputs can be trusted.
- Prescient assessed the open-source and traditional MMM landscape in 2019, determined that no existing model was built to reflect how modern marketing works, and spent years building a proprietary approach from scratch.
- Because Prescient’s methods are proprietary and have not been published, no LLM has been trained on them, and no vibe-coded MMM can replicate what Prescient does.
The rise of vibe-coded MMMs
AI-assisted development has lowered the barrier to building software dramatically, and data science tools are no exception. It’s now technically possible for someone without a background in marketing science or computer science to prompt their way to a working MMM. And without formal training in either field, there’s no guarantee they would know what to look for if something was wrong with the output.
What “vibe coding” actually means in this context
Vibe coding refers to using an LLM-based coding tool to generate functional code by describing in natural language what you want it to do. In the context of MMMs, that might look like asking an AI coding assistant to build a marketing mix model, configure it for ecommerce data, and produce channel-level attribution. The output can look polished. It can pass a basic sanity check. The problem is that functional code and accurate modeling are two entirely different things, and only someone with the right background is equipped to tell them apart.
Why MMMs became a target for this approach
The renewed interest in MMM following iOS 14.5 privacy changes created a surge in demand that far outpaced the supply of qualified practitioners who know how to build and interpret these models. Major tech companies like Meta and Google released their own open-source MMM frameworks, which lowered the floor considerably. The natural next step for a developer-adjacent marketer or a startup looking to offer measurement services was to layer AI-assisted code generation on top of those frameworks. No statistics degree required, no years of research, and no original modeling work.
The open-source foundation and what it inherits
Most vibe-coded MMMs are built on one of a small handful of publicly available frameworks: Meta’s Robyn, Google’s Meridian, and PyMC-Marketing are the most common. These are legitimate tools developed by serious researchers. But they were designed for broad applicability, which means their defaults reflect assumptions about how marketing works in general, not how any specific modern brand’s marketing mix actually behaves.
What these frameworks were designed to do
The structural assumptions built into these frameworks include separable baseline-media decomposition, which assumes that baseline demand and the effects of marketing spend are independent of each other. They also use fixed parametric saturation curves, functions like Hill, Michaelis-Menten, or Weibull, that force diminishing returns by design regardless of whether the underlying data actually supports that pattern. Most operate at the channel level rather than the campaign level, and they apply static adstock transformations that may not reflect how the impact of any given campaign actually decays over time. A vibe-coded MMM built on top of these frameworks inherits all of these design decisions, even if the person who built it doesn’t fully understand what they mean.
Why these assumptions matter
Forced saturation is a good example of why this is a real problem and not just a theoretical one. Prescient’s own research has found that linear response patterns often outperform saturating parametric forms in both explanatory power and predictive accuracy (look for the research paper coming soon to our website). When a model forces saturation by construction, it can systematically underestimate scalable opportunities and push budget recommendations toward underspending. Brands using these models might be leaving revenue on the table not because their campaigns are saturated, but because the model was built to assume they would be. A vibe-coded MMM built on open-source frameworks doesn’t just inherit this limitation; it inherits it silently, with no flag to the user that the output is shaped by this structural assumption.
What LLMs can and can’t know
This is the most important thing to understand about why vibe-coded MMMs have a hard ceiling on their quality. LLMs are trained on publicly available data: research papers, open-source repositories, documentation, blog posts, and public codebases. Their knowledge of marketing mix modeling is bounded by what has been written, published, and indexed.
LLMs remix; they don’t research
An LLM can generate syntactically correct, logically coherent MMM code because the patterns are well-represented in its training data. What it cannot do is develop a new hypothesis about how marketing works, test that hypothesis against real brand data, backtest it for accuracy, and iterate on it over years. The output of vibe coding is always a recombination of existing methods, assembled from what the model has already seen. This is not a critique of how well a developer prompts or how carefully they iterate. It’s a structural reality of how large language models work.
The knowledge ceiling problem
Even a highly skilled developer using the best available AI coding tools is working within a knowledge ceiling defined by what has been published. That ceiling corresponds almost exactly to the open-source MMM frameworks those tools have been trained on. This means the inherited structural constraints of Robyn or Meridian are not a starting point to improve on through clever prompting. They are the outer boundary of what a vibe-coded MMM can produce. Whatever limitations those frameworks carry, a vibe-coded MMM carries them too.
How Prescient was built differently
In 2019, Prescient’s founding team evaluated the MMM landscape, open-source frameworks and traditional regression-based approaches alike, and reached a clear conclusion: none of them were built to reflect how modern marketing actually works. That wasn’t a reason to pick one and customize it. It was a reason to build something entirely new.
Starting with the right question
Rather than asking how to implement an existing MMM, the team asked what a model would actually need to look like to capture campaign-level attribution, halo effects across organic search, branded search, and direct traffic, non-linear and non-saturating response functions, and model updates that happen daily rather than monthly or quarterly. Those requirements drove the architecture. The team assessed the available approaches, determined that none of them met the bar, and started from scratch.
Years of research, not prompts
The models Prescient uses today are the result of years of data science research, testing on real brand data, and ongoing refinement by a team with deep expertise in both statistics and marketing. Prescient’s approach has been benchmarked against open-source baselines, and the results show substantially lower attribution error across every evaluation dimension. That research has since been formalized, but the work that produced it spans years of iteration that no LLM has access to and no prompt-driven workflow can replicate.
Proprietary by design, not by accident
Because Prescient’s methods have not been published, they are not part of any LLM’s training data. A vibe-coded MMM built today, regardless of how sophisticated the development process, cannot reproduce Prescient’s approach. This isn’t a matter of access or gatekeeping. It’s simply how LLM knowledge works: models can only draw from what has been made public. Prescient’s research hasn’t been.
Questions to ask before trusting any MMM
If you’re evaluating a new measurement tool, including one that may have been vibe-coded, these are worth asking directly.
- What open-source framework, if any, is this built on?
- What saturation assumptions does it make, and are those assumptions configurable based on your data?
- Can it report at the campaign level, or only at the channel level?
- How were the modeling choices validated against real brand data?
- Who built this, and what is their original research contribution to the field?
The answers won’t always be easy to get, but the questions themselves will tell you a lot about whether the people behind the tool understand what they’ve built.
Where Prescient comes in
Prescient was built by a team of data scientists and marketers who looked at the existing MMM landscape and decided it wasn’t good enough. The result is a proprietary model that operates at the campaign level, updates daily, and is built to reflect the actual complexity of modern marketing, including halo effects, non-linear response patterns, and the ways upper-funnel activity influences lower-funnel performance. None of those capabilities come from an open-source framework. They come from years of original research.
If you’re evaluating your current measurement approach and want to understand what an MMM built from the ground up actually looks like in practice, we’d love to show you. Book a demo to see Prescient in action with real data.
