What is a prior?

In a Bayesian marketing mix model every parameter starts as a prior: a probability distribution that says what values are plausible before the model has looked at the data. For a channel that usually means two things, how much sales a given spend produces and where the returns start to flatten, each with a centre and a width. A wide prior says "we are not sure"; a narrow one says "we have good reason to believe this".

Priors are not guesses dressed up as maths. They are the place where everything already known about a channel goes in: past budgets and results, the platforms' own conversion data, the team's experience, and the results of any experiments already run.

Odins builds priors from three sources. Historical spend gives reasonable levels and marginal return per channel. The digital channels give saturation signals: a response curve fitted to the platform's own conversion and impression data. Structured interviews with the team translate experience into numbers, for example what a 25% increase in a channel would be expected to do.

Platform data goes in here too, together with the other knowledge the customer has. It is not used as the answer; it informs the starting point. Results from lift tests and geo experiments go in as priors as well: a point estimate and its uncertainty become the model's starting belief about that channel's return.

Marketing data is thin. One national market gives one data point a week, and the model has to estimate roughly ten activity types from it. A model built in a country with fifty regions can lean on data; a model built for one national series cannot, so the quality of its priors decides whether it is stable or not. That is why prior-setting is a discipline at Odins rather than a default setting, and why the question to ask any MMM vendor is how they set theirs, and whether the approach assumes data the market does not have.

Priors are also what keep always-on activity honest. Paid search and PR run every week, so a data-only model struggles to separate them from the baseline. A prior says how the team believes those channels work, and the model refines that against the data instead of hoping a random variable soaks up the effect.

• Posterior: The model's answer: the prior after the data has had its say, with a most likely value and a range.
• Bayesian marketing mix modeling: A model that starts from what you already know and lets the data correct it, with a range on every answer.
• Calibration: Anchoring a model to outside evidence, usually experiment results. In a Bayesian model, calibration is a prior.
• Saturation curve: The response curve per channel: how sales respond to spend, and where the next unit stops paying.
• Control variable: The non-marketing factors the model explains sales with, so they are not credited to a channel. Fewer than you would think.

Go one level deeper: The technical view: informed priors for thin data. How the model works. Or go back to all terms in the glossary.