Priors are one of the least understood – and most essential – parts of a Bayesian media mix model.
In simple words, a prior is what you know (or believe) about the world before the model sees any data. In marketing, that might include common-sense expectations like: “spending money on advertising probably doesn’t reduce my sales.” Or that impulse-buy channels behave differently than high-consideration ones. In Bayesian statistics, priors are what allows us to translate those assumptions into mathematical form.
So why does this matter? Because real-world marketing data is messy. It’s collinear, noisy, and often too limited to answer every question on its own. If you feed that data into a model without priors, it will still return an answer, but it may be wildly unrealistic.
For example, without any constraints, a regression could spit out a -300% ROI for a channel that everyone agrees has value. Obviously implausible, but you wouldn’t believe how often we see that.
And that’s where priors come in. They don’t (shouldn’t) override the data, but they provide guardrails to help the model produce reasonable results. This doesn’t mean introducing bias or cherry-picking results, but setting up the constraints in a way that represents your knowledge about the world.
The Spectrum: From Uninformative to Informative Priors (And Why Neither Is “Best”)
Marketers often hear the terms “informative” and “uninformative” priors thrown around in MMM conversations. The truth is that there’s no actual hard definition between what is an uninformative set of priors or an informative set of priors. Instead, think of them as two ends of a spectrum.
A truly uninformative prior means the model starts with almost no assumptions, and anything could happen. This can lead to absurd results. Imagine you were building a model to understand humans’ height, and you set it so people could be anywhere from -20 to 300 feet tall – that would be an “uninformative” prior – technically correct but clearly not useful.
On the other side of the spectrum, you have informative priors that impose stronger assumptions. The challenge here is not getting so specific that you end up putting your thumb on the scale.
In practice, most MMMs land somewhere in the middle. You want priors that reflect what’s plausible, but not ones that bias your results. What matters most isn’t how “informative” your priors are in the abstract. It’s whether they’re justifiable and whether the people who will be using your results would agree with them.
The Real Danger: Cooking the Model to Get the “Right” Result
To be clear, it is absolutely possible to bias a Bayesian model using priors. If you tell the model that Facebook ROI is probably 7x, then – surprise – it’s going to pull that way.
And to be honest, the more common issue we see isn’t really manipulation on bad faith. It’s overconfidence.
Analysts and vendors often impose priors that reflect what they hope to see / think will see, not what the data supports. That might mean biasing all estimates toward positive ROI to make performance look better, or using tight priors to “smooth out” noisy channels and hide uncertainty.
That’s why we encourage a simple test: if you showed your priors to a skeptical CFO, would they agree with them? If not, it’s a red flag. Seriously, go show your finance team – this kind of framing builds credibility with them.
In our experience, CFOs care more about whether your priors are reasonable than whether your intervals are narrow. They want to know the model is playing fair and if your forecast range shifts because of a more conservative prior, that’s fine. What matters is that they understand why.
And when in doubt, run a sensitivity check. Use a tighter prior, then a looser one. What changes? What doesn’t? If the conclusions fall apart under slightly different assumptions, they probably weren’t that solid to begin with.
But again, great priors are defensible and transparent.
How Recast Builds Priors: Guardrails, Not Guarantees
At Recast, we use them to keep the model grounded in reality.
The first step is what’s called a prior predictive check. Before we fit the model to any data, we simulate what the priors alone would imply. If those simulations suggest sales could range from -$20,000 to $800 billion for a mid-sized brand, we know something’s off. What we’re looking for here is that the range of possible sales, according to just the simulation that we’re doing, is within an order of magnitude or so of the actual data.
The goal is to avoid obviously implausible outcomes, not to force a specific answer. We’re not baking in a belief that Meta always works or that TV always fails. All we’re doing here is just constraining the model at a high level.
Step 1: Start with Structured Discovery to Surface Business Beliefs
Every model starts with structured discovery and we don’t expect brands to know their exact incrementality numbers — that’s why they’re working with us. But even directional beliefs can be good guardrails.
We ask:
- How much of your business is organic?
This isn’t a fixed number — it’s a belief. For some DTC brands, it might be 20–30%. For legacy brands with strong word-of-mouth, it might be 70%+. - How long is your purchase journey?
Timing shapes incrementality. A mattress brand has a very different lag than a food delivery app. We use this to build realistic decay curves. - What role does each channel play?
We ask how teams think about channels — not just spend. Is YouTube driving awareness? Is paid search just picking up branded queries? - What else impacts performance?
We look for macro and micro factors — Black Friday spikes, pricing tests, iOS changes. These need to be modeled or explicitly excluded.
The goal is not precision. It’s to codify how the business actually operates — so we don’t let the data override common sense.
Step 2: Encode Those Beliefs into Mathematical Priors
Next, we encode those beliefs into mathematically useful priors.
- Base demand:
This is the model’s expectation of what sales would be without any paid media, which may move over time. For mature brands, we assume a higher, more stable baseline. For newer brands, we assume more volatility and a stronger dependence on paid media to drive growth. - Channel ROI/CPA:
Priors help keep ROI estimates plausible, especially when there’s multicollinearity. If Facebook and Google always move together, priors stop the model from assigning absurd results (like 1000x for one, -999x for the other). - Time-shift curves:
Channels don’t work on the same timelines. Display might take weeks to convert. Paid search is immediate. Each is assigned a lag profile based on real-world behavior and the brand’s experience. - Lift test anchoring:
If clients have run clean lift tests, we use them. We can pin a channel’s ROI to that range (e.g., 3.5–4.5x) so the model doesn’t override it with noisy observational data.
It’s important to note that priors aren’t fixed. But they start as explicit, directional beliefs from people who know the brand best.
Step 3: Run Simulations to Validate and Sanity Check Priors
Before we train any model, we validate that our priors even make sense.
- Prior predictive checks:
We simulate data using only the priors. If the model predicts $5M in revenue for a brand that makes $50M, we know something’s off. This catch-before-you-train step helps us prevent wasted modeling cycles. - Parameter recovery checks:
We then simulate “known” data, feed it through the model, and see if it recovers what we put in. This is how we verify that our model can even answer the types of questions marketers want answered. - Configuration compatibility:
Finally, we gut-check: do these priors fit the reality of the business? If a brand says their base is 80% organic and they have a 5x paid ROI, we sanity check that those assumptions don’t imply more revenue than actually exists.
This process makes the model not just more accurate — but more useful. And it builds trust with the marketers who have to use the results.
Recap: Priors Make or Break a Good MMM Model. Here’s How to Get Them Right.
- Priors keep media mix modeling outputs grounded in marketing reality
- But they’re difficult to set, especially with multicollinearity and shifting environments
- Many MMM failures stem from poor or unvalidated prior assumptions
- At Recast, we follow a 3-step process: elicit, encode, validate
- The result: models that reflect the real world, not just noisy data
TLDR:
- Priors are necessary assumptions in Bayesian MMM that keep models realistic – but they must be carefully chosen to avoid bias.
- “Informative” and “uninformative” priors exist on a spectrum and should reflect plausible, defensible assumptions.
- At Recast, we use prior predictive checks to constrain models without favoring any channel or result, ensuring guardrails without distortion.
- Communicating ranges and priors transparently builds trust with finance because clarity is what actually earns you budget.
