Whether your company is ready for marketing mix modeling depends on three conditions: enough first-party conversion volume for the model to find signal (our benchmark is 500 a month), enough ad spend that moving dollars off non-incremental channels pays for the work (roughly $5M a year), and a budget decision waiting on the answer with someone who has authority to act on it.

Most companies frame this question around money, meaning how much they need to be spending before marketing mix modeling is worth it. Spend is the wrong first variable. Whether a model can be built at all depends on transactional volume, and our benchmark is 500 first-party conversions a month, because below that there is not enough signal to separate one channel's contribution from another's no matter how large the budget is. Spend decides something different, which is whether building the model pays for itself.

Underneath both sits the question that actually settles it: what decision are you trying to make? A model earns its cost when a call is waiting on it: how much to put into the upper funnel, whether a channel is carrying its spend, what happens to revenue if the budget moves by a third. And someone with the authority to act once the answer lands. Start from the decision and the technical requirements follow from it, including whether an MMM is the right instrument at all or whether a single incrementality test would answer the question faster and cheaper.

The rest of this page takes each condition in turn, including the cases where the honest answer is not yet.

What minimum ad spend makes marketing mix modeling pay for itself?

No minimum spend is required to build a marketing mix model. If the transaction volume is there, a model can be built for a business spending a few hundred thousand dollars a year. The threshold that matters is economic rather than technical, and it sits at roughly $5M in total annual ad spend.

The arithmetic is easy enough to check. Across our own engagements, moving dollars off non-incremental channels and spending at the level where returns have not yet flattened is worth 20 to 40% of the media budget. At $5M a year the low end of that range is $1M, which covers a six-figure engagement about ten times over. At $20M it is $4M against much the same cost. Below $5M a model still works, and the improvement it produces simply starts to look small next to the cost of producing it.

Be clear about where that range comes from: our own client work, not an industry benchmark, and the result in any given account depends on how misallocated the budget was to begin with. A company already allocating well has less to gain, which is worth knowing before buying anything.

Why does $5M appear so often in this category? Most vendors use it as a qualification cutoff, which is a statement about their cost of serving you rather than about whether a model can be built on your data. The two get conflated, and companies below the line conclude the method does not work for them when what they have been told is that it does not work for the vendor.

When does MMM make sense, and when does it not?

Marketing mix modeling makes sense when a meaningful share of your spend does work that clicks cannot see. Upper funnel video, awareness campaigns, streaming TV, podcasts, out of home, organic social and anything with a lag between exposure and purchase all show up badly or not at all in platform reporting. The more of your budget sits there, the more you are allocating on numbers that describe something else. It also makes sense once you run enough channels that the question is no longer whether one of them works, it is how much each should get.

The clearest signal that you will need one soon is an intention rather than a current state. If you want to move up the funnel, if brand is on next year's plan, if you are weighing television or audio or anything you cannot put a pixel on, you need a way to measure it before the money goes out. Building the model after the budget is committed is how teams end up defending a decision rather than making one.

The case against is just as clear. If the customer journey is short and everything you run sits at the bottom of the funnel, branded search and retargeting and not much else, platform reporting is closer to the truth here than anywhere else. The few things genuinely in question can be settled with a geo holdout test at a fraction of the cost and time. The same applies when a single channel carries most of your spend, since a model built to apportion credit across a portfolio is an expensive way to answer a question about one line item. Start with the test, and build the model when the portfolio gets complicated enough to need one.

How much data history and conversion volume do you need?

500 first-party conversions a month is our benchmark, and 12 months of history is what we prefer. Neither is an absolute floor. Models can be built on less of both, and what you give up is precision: fewer conversions means wider credible intervals, and less history means the model has not seen a full year of your seasonality and will be worse at telling a seasonal swing apart from something you did.

The volume requirement is about signal rather than size. An MMM works at weekly grain, so a year of data is roughly 50 observations, and each one needs enough conversions in it that a change in a channel's spend produces a movement the model can tell apart from noise. Below 500 a month most weeks are too thin for that, and the model returns intervals so wide that every answer is compatible with every decision, which is worse than having no model because it looks like an answer.

On history, much of the category asks for more. Recast publishes an 18-month floor and two to three years is a common request. More data genuinely is better, and it is not the barrier it is usually presented as. Twelve months gives you one full cycle of seasonality, which is enough to model it rather than guess at it, and where a client has had less we have built on less and said plainly which conclusions were provisional.

Does your data need to be in order first?

Your data does not have to be in order before you start, and it does have to be in order before the model is worth trusting. Those are different statements, and confusing them is what stops companies that should be doing this and what produces bad models at the companies that push ahead anyway. Nobody should fit a model to numbers they cannot stand behind, so the question is never whether the data gets cleaned, it is who does that work and in what order.

A model needs spend by channel by week, first-party conversions with their revenue and their dates, and the non-media things that move the business: promotions, price changes, stock levels, CRM sends, product launches. Almost nobody has all of that in one place at the start. It sits across ad platforms, a warehouse, a BI tool and a few spreadsheets somebody maintains by hand, which is the normal starting condition rather than a disqualification.

What matters is whether the underlying records exist and are consistent, not whether they are tidy. Three things genuinely stop a model. No reliable first-party record of conversions, because there is nothing to model against. Spend history that changed definition partway through, where a channel was renamed or two accounts were merged and the numbers either side are not the same measurement. And conversions that cannot be tied to a date, which makes them impossible to line up against the spend that may have caused them. Everything else is work.

That work has a name, and it is worth doing whether or not you ever build a model. A semantic data layer is the definition, held in one place, of what a conversion is, what a channel is and what revenue counts, so every model, report and agent reads the same number. Teams without one usually find out the hard way, when they try to figure out what is working and the first week produces three different revenue figures and a meeting about which one is right. Teams with one find every piece of analysis after it cheaper, and they are in a position to put agents on top of their data rather than people.

We build that layer with clients who do not have one, as part of the engagement rather than a precondition. It is worth putting the same question to any vendor you evaluate, because one that assumes your warehouse is already in order will hand that problem back to you in month one.

Do you need a measurement vendor, or can you build this yourself?

You can build it yourself. The real question is whether you want to own measurement as a standing capability rather than run it as a project. The code is no longer the obstacle, since Meridian, Robyn and PyMC-Marketing are open source and genuinely good. The judgment is. Models go wrong on judgment calls: priors that encode what you already know without quietly deciding the answer, adstock and saturation shapes that match how each channel actually behaves, and a specification that can be identified at all given how correlated your channels are. They go wrong silently. A model with badly chosen priors returns a confident number and no error message, which is why validation design matters as much as the model itself.

Getting that first model right is difficult, and it is still only the start. Around it sits infrastructure that has to run whether anyone is watching or not: a pipeline that refreshes the data and retrains on a schedule, and back-testing against held-out weeks so you know whether this week's version is better or worse than last week's. Then scoring and gates that stop a degraded model reaching anyone, and monitoring for the renamed field or schema change that quietly corrupts an input. Teams that skip this end up with a model that was right in month one and no way to say whether it still is.

The second half gets less attention and is harder. A posterior distribution is not a decision. Somebody has to turn marginal returns and saturation curves into what to do with next week's budget and reconcile that against what the platforms and the incrementality tests are saying. Then put it in front of the people who own the channels in a form they will act on, and carry it into the ad accounts. That translation is most of the value and almost none of the open source code. Building the model in house is realistic. Building the engine that keeps it accurate and turns it into action every week is the part that quietly becomes a team.

Build it yourself when you have data science capacity you can keep, when the answers matter enough that you want the argument happening in house, or when your business is unusual enough that a standard model would need heavy adaptation anyway. Buy when the constraint is time or people, when you need the first real decision in weeks rather than the six to twelve months an in-house build typically takes, or when you want someone accountable for the number rather than a model nobody owns.

A third answer is worth naming. You can buy the capability and keep the ownership, and the question that separates vendors on this is whether the model, the data layer and the business context they build are yours at the end of the contract. If they live on the vendor's platform and leave when you do, what you have rented is an opinion. We wrote a companion piece on the build vs. buy decision.

An MMM readiness checklist for your business

Work through these honestly. The first four decide whether a model can be built at all, the last three decide whether it is worth building.

• You record at least 500 first-party conversions a month, with dates and revenue attached.

• You have roughly 12 months of spend history by channel, and the definitions did not change partway through.

• You can tie your conversions to a date and line them up against spend on a common time axis.

• You know what a conversion, a channel and revenue mean across the business, or you are prepared to define them once as part of the work.

• You spend around $5M a year or more across all channels.

• You put a meaningful share of that spend into work clicks cannot see, or you intend to move that way within the next year.

• You have a decision waiting on the answer, and somebody with the authority to act on it.

Most of the first four can be answered in an afternoon by whoever owns your warehouse. If you fail one of the last three, the honest answer is not yet, and a geo holdout test on the channel you are least sure about will teach you more this quarter than a model would.

FAQ

Is my company too small for marketing mix modeling?

Probably not for technical reasons. If you record 500 first-party conversions a month you have enough signal to build a model, whatever your spend. The size question is economic: below roughly $5M in annual ad spend, the improvement a model produces starts to look small next to what it costs to produce.

What if we only have six months of data?

A model can be built on six months, and it will not have seen a full year of your seasonality, so it will be weaker at telling a seasonal swing apart from something you did. That is workable as long as the seasonal conclusions are treated as provisional and revisited once a full cycle is behind you.

Can we do this without a data warehouse?

Yes, though somebody has to define what a conversion, a channel and revenue mean, and get them onto a common time axis. That definition is the semantic layer, and whether it lives in a warehouse or somewhere simpler matters less than whether it exists and everyone reads from it.

Do we need incrementality tests as well as an MMM?

They answer different questions. A test gives you the truth about one channel at one spend level at one moment, while a model tells you how the whole budget behaves and keeps telling you as things change. Tests calibrate the model, which is why running both produces a better answer than either on its own.

How do I know whether I need a vendor or can do this myself?

Build in house if you have data science capacity you can keep and you want measurement as a standing capability. Buy if the constraint is time or people, or if you need a real decision in weeks rather than the six to twelve months an in-house build typically takes.

Does MMM work for businesses with few but high-value conversions?

The 500 a month benchmark is about having enough weekly observations to separate signal from noise, so a business with 200 large transactions a month is harder to model regardless of how much revenue those transactions carry. Start with incrementality testing on the channels you are least sure about and revisit the model as volume grows.

Find out where your MMM readiness stands

If you are not sure which side of these lines you fall on, the fastest way to know is to show someone your actual numbers. A 30-minute conversation covering your conversion volume, your history and the decision you are trying to make will tell you whether a model is worth building for you, including when the answer is that it is not yet. Book a call.