<img height="1" width="1" style="display:none;" alt="" src="https://px.ads.linkedin.com/collect/?pid=9151474&amp;fmt=gif">
Skip to content

Guide

How to choose a marketing mix modelling vendor in 2026

Most marketing mix modelling vendors now use Bayesian models, show saturation curves and promise better budget decisions. So the statistics are rarely what separates them. What separates them is who does the work after the first model is delivered: collecting the data every week, retraining the model, explaining what changed and turning it into a budget decision. This guide covers the questions that expose those differences, and the practical ones about budget, data and time to first results.

Updated November 2026 · about 15 minutes to read

By the Odins team · Updated November 2026 · About 15 minutes

Most marketing mix modelling vendors now use Bayesian models, show saturation curves and promise better budget decisions. So the statistics are rarely what separates them. What separates them is who does the work after the first model is delivered: collecting the data every week, retraining the model, explaining what changed and turning it into a budget decision. This guide covers the questions that expose those differences, and the practical ones about budget, data and time to first results.

The short answer

Pick the vendor whose way of working fits your team, then test them on four things: how they set up the model, how they check it, whether it answers budget questions rather than reporting on the past, and what it takes from you to get started. A framework you run yourself, a consultancy project and a managed platform can all use similar maths and still give you very different results a year later.

1. Start with who will run it

There are four common routes to MMM, and each puts the ongoing work in a different place. The figure shows where the recurring tasks sit in each.

Four routes to MMM, and where the recurring work sits
Open-sourceframeworkConsultancyprojectSelf-servesoftwareManagedplatformhow Odins worksData pipeline, every weekYouVendor, onceYouVendorModel design and priorsYouVendorYouTogetherMonthly retrainingYouNobodyYouVendorChecking forecasts against actualsYouRarelyYouVendorTurning it into a budget decisionYouReportYouPerson-reviewedWork that stays with your teamWork the vendor does, month after monthThe maths is similar across the four. Where the recurring work sits is what differs,and it decides what the model looks like in month 14.
  • Open-source frameworks such as Google Meridian, Meta Robyn and PyMC-Marketing. The code is free and good. Your team owns the data pipeline, the model design, the retraining and the interface the marketing team uses. Companies that do this well tend to have several data scientists on it.
  • Consultancies and media agencies. A team builds a model and delivers a report, often once or twice a year. Good for a one-off answer, harder to keep current between projects.
  • Self-serve software. You get a login and the model runs in the vendor's product. Your team still prepares data, configures the model and interprets the output.
  • Managed platforms. The vendor connects the data, builds and maintains the model, and reviews the recommendations before you see them. Odins works this way.

The question to ask yourself is plain: who in our company will maintain this in month 14? A common story is a model built on Robyn or Meridian by one analyst, which nobody updates after that person moves on. If that is where you are, ask each vendor whether they can learn from the existing work (the channel assumptions, the data cleaning, the findings) rather than start from nothing. A lot of useful thinking usually went into that first model.

The month-14 test

Whatever route you pick, write down who retrains the model, who checks it against actual sales and who presents it to the CFO, a year from now. If any of those three names is "the vendor" and the vendor's answer is "your analyst", you have found the gap before it costs you.

2. Ask how the model is set up

Bayesian is now standard. What matters is what goes into the model before it sees your data, and whether you can see it.

  • How are the priors set? Priors are the starting beliefs about how each channel works. On thin data, they decide much of the result. Ask where they come from: your own past budgets and results, signals from the ad platforms, structured conversations with your team, or defaults from a library. At Odins this typically takes two or three one-hour sessions with the customer.
  • Can you see inside it? Ask to see saturation curves per channel, the prior next to the posterior, the baseline, and the uncertainty range on every number. A vendor that only shows point estimates is hiding the most useful information.
  • Is it independent? Ask whether the vendor also sells media, and whether the model or its defaults come from an ad platform. Open-source code from Google or Meta can be inspected, so the risk is less in the code than in how it is configured and which data it is given. An independent vendor has no stake in which channel wins.
  • Does it change over time? Channel response is not constant. A model that assumes Google Search works the same in July as in November forces the wrong answer on every seasonal decision. Ask whether saturation and seasonality can vary over time. We explain why this matters in time-varying saturation.

The reason the priors matter so much is the shape every channel is modelled with. Each channel gets a response curve: more spend, more sales, with diminishing returns as the channel saturates. With two years of weekly data the model can place that curve reasonably well in the range where you have actually spent. Outside that range, and for channels where spend never varied, the prior is most of what it knows.

The response curve, and the two returns it separates
Where you have actually spentWeekly spend on the channelIncremental salesCeiling: the most the channel can addHalf-saturation pointhalf the ceiling is reached hereToday's spendMarginal return: the slope at today's spend, what the next unit earnsAverage return: total sales over total spend, what a report showsSame channel, two returns. Reports use the average. Budget decisions have to use the marginal.

3. Ask how the results are checked

Every model fits its own history well. That proves little. Ask how the vendor shows that the model predicts sales it has not seen. The honest version works like this: the forecast for the coming period is stored before the period starts, then the actual sales are laid next to it, week by week, and the gap is reported.

Verification: the forecast is stored first, then compared with what happened
Weekly salesWeeks after the forecast was storedForecast stored before the period startsForecast, with its uncertainty bandActual sales, as they arriveA model that fits its own history proves little. Ask to see this chart for every month, in the platform,without the vendor preparing it. Odins aims for forecasts within 10 to 15 percent of actual sales in steady state.
  • Is the forecast compared with actual sales every month, and can you see that comparison yourself?
  • What accuracy do they expect in steady state, and what happens when the model drifts? Odins aims for forecasts within 10 to 15 percent of actual sales in steady state, checked monthly in the platform.
  • Does a person review the output before it reaches you? Automated recommendations from a model that has drifted are worse than none.
  • Will someone meet you regularly to explain what changed and why? Ask who that person is, and how often.

Two words to listen for

"Fit" is how well the model explains the past. "Forecast" is how well it predicts sales it had not seen when the forecast was made. Vendors that only talk about fit are showing you the easy half.

How to read a vendor's accuracy number, and the three tests that mean something, is in how accurate is marketing mix modelling.

4. Insist on decisions, not reports

The point of MMM is not to explain last quarter. It is to answer the next budget question. Two questions matter most:

  • How much should we spend in total? The model should show the return on the next part of the budget, not only the average return. That is what tells you whether a larger budget still pays back.
  • Where should it go? Given a fixed budget, which channels have room to grow and which are saturated, and what is the expected effect of moving money between them?

Ask the vendor to show a scenario live: "What happens to sales if we move a tenth of the search budget to TV next quarter?" Then ask what it does when it is unsure. A good answer is a structured test: which channel to test, for how long, at what cost, and how much narrower the uncertainty will be afterwards. Moving budget toward channels with headroom typically gives 5 to 15 percent more effect from the same budget, so the question is worth asking properly.

5. The practical questions: budget, data, time

The three practical numbers to ask for
MINIMUM BUDGETAbout EUR 1M a year and upMMM needs variation in spend tolearn from. Below that, a cleancross-channel dataset is usuallythe better first step.DATA HISTORYTwo years of weekly dataEach season is seen twice. Threeyears is better. Six months canwork with strong priors, with wideruncertainty as the price.TIME TO FIRST MODELFour to six weeksFrom kickoff, data collection andstructuring included. The slowestpart is offline media held by youragency, so start that request onday one.Rules of thumb from our own onboarding. Ask every vendor for the same three numbers and for the plan behind them.
  • Minimum budget. MMM needs variation in spend to learn from. As a rule of thumb it starts to pay off with a marketing budget of about EUR 1 million a year, and there is no upper limit; the same model covers budgets in the hundreds of millions. Below that, a clean cross-channel dataset may be the better first step.
  • Data history. Two years of weekly data is a comfortable floor, and three is better. Six months can work when strong priors carry more of the load, with wider uncertainty as a result.
  • Time to first results. Ask for a realistic plan, including getting data out of your media agency. At Odins the first model is typically live four to six weeks after kickoff, and that includes collecting and structuring the data, digital and offline. Once the data is in, the model itself takes two to three weeks. Vendors who quote shorter timelines usually count from the day your data is already collected and prepared, and often do not offer data collection at all. The model is refreshed monthly after that. Data flows in daily and the platform is open any time for reporting, scenarios and forecasts. The model itself is retrained monthly, which is as often as a budget decision changes.
  • Starting small. Ask whether you can start with one market or one brand and add the rest once the first model has proved itself.
  • Your data afterwards. Ask whether the cleaned, structured dataset can flow into your own warehouse, and who owns it.

The data side is usually the slowest part of any plan, whichever vendor you choose. We go through what to collect and how to get it out of your agency in data for marketing mix modelling.

6. Red flags, and how to run a pilot

Most vendor meetings go well. The problems show up a year later, and they were usually visible in the first conversation. These are the signs we would treat as warnings, whichever vendor is in the room.

  • Only fit, never forecast. The accuracy slide shows the model's line on top of your history, and nobody can show a forecast that was stored before the sales came in.
  • An accuracy number without a method. "95 percent accurate" with no answer to how it was measured, over what horizon, or on which model. A number that cannot be explained was chosen for the slide.
  • Point estimates only. Every channel has one return and no range. The uncertainty exists; it is being hidden from you.
  • Every channel equally confident. A channel whose spend never varied is presented with the same certainty as one with three years of tests behind it.
  • Priors you cannot see. The vendor cannot show what the model assumed before it saw your data, or the priors come from a library with no sessions with your team.
  • A timeline that starts after your data is ready. Two weeks to a model sounds good until you learn that collecting and structuring the data is your job and not in the number.
  • Nobody named for month 14. Ask who retrains the model, who checks it against actuals and who presents it to the CFO a year from now. "Your analyst" is a gap, not an answer.
  • Recommendations nobody reviews. Budget moves go straight from the model to you, with no person checking whether the model has drifted.
  • A commercial stake in a channel. The vendor also buys media, or the tool's defaults come from an ad platform with a view on which channel should win.
  • Your old work is worthless. You have an earlier model or a cleaned dataset, and the vendor wants to start from nothing rather than learn from it.

None of these alone rules a vendor out. Two or three together, with no good explanation, usually do.

How to run a pilot that proves something

A pilot is the fairest way to settle the questions above, and it works when the terms are set before it starts.

  • Scope it to one market or one brand. One model on one outcome. Add the rest once it has proved itself. The timeline is the normal one: four to six weeks to a first model with data collection included, then at least two months of forecasts checked against actuals. Plan for about three months in all.
  • Write the success criteria down first. Three are enough: the data pipeline runs without your team's monthly effort; the stored forecasts land within an agreed range of actual sales over the pilot weeks; and the model answers at least one real budget question you have today, with a range on the answer. Agree the range before anyone has seen a result.
  • Ask for a holdout on your own data. Train to a cutoff a year back, forecast a quarter you already know. It is the fastest preview of how the model will behave live. How to read it is in how accurate is marketing mix modelling.
  • Running against an existing model? Keep them apart. If you already have an MMM and want to see whether a new one agrees, give both the same validated dataset and do not let the new vendor use the old model as its priors. The new priors are built the usual way, from platform data, your spend pattern and sessions with your team. Where the two models disagree, the stored forecasts against actuals decide.
  • Agree the exit. What you keep if you stop: the structured dataset, the documented priors and the findings. A vendor who is confident in the result will not object.

A checklist to take into vendor meetings

Twelve questions, with what a good answer sounds like. Take the right-hand column with you; it is the one vendors find hardest to fake.

AskA good answer
Who maintains the data connections, and what do you need from us each month?The vendor does, and watches them. From you: nothing recurring beyond approving access.
How are offline channels such as TV, radio and outdoor collected?Through a pipeline for the agency's files, not a spreadsheet you fill in every month.
How do you set priors, and can we see them next to the posteriors?From your history and your team's knowledge, in a few working sessions. Yes, per channel, in the product.
Can saturation and seasonality change over time in your model?Yes, and here is a channel where it mattered.
How do you show that the model predicts unseen sales, and how often?Stored forecast against actuals, every month, visible to you.
Does a person review recommendations before we receive them?Yes, and this is who.
Can the model answer how much to spend in total, not only how to split it?Yes: the marginal return at today's budget, with a range.
How do you handle channels where the model is uncertain?A structured test with a cost, a duration and what it will narrow.
What can we learn from or reuse from our earlier MMM work?The channel definitions, the data cleaning and the findings. Not the old code as it is.
How long until the first model, and what is on our side of that plan?A week-by-week plan, with your tasks named. Four to six weeks from kickoff is realistic when data collection is part of the plan; ask whether the vendor's number starts before or after your data is prepared.
Can the structured data flow into our own warehouse?Yes, and you own it.
Do you sell media or have commercial ties to any channel in the model?No.

Where Odins fits, and where it does not

Odins is a managed marketing investment platform. We connect the data, digital and offline, build and maintain Bayesian models, and deliver monthly recommendations that our team reviews before they reach you. Scenarios and forecasts you run yourself in the platform, or from Claude, ChatGPT or Copilot through the Odins MCP, at any time. Companies like CDON, Nettbil, Hyre and Megaflis run their marketing budgets on it.

It is not the right choice for everyone. If you have a strong data science team that wants to own the code, an open-source framework is a sensible route, and our build versus buy guide for Meridian covers the trade-offs. If your marketing budget is well below EUR 1 million a year, start with the data. For a wider view of the market, see our overview of MMM providers.

Frequently asked questions

What is the minimum marketing budget for MMM to be worth it?

As a rule of thumb, around EUR 1 million a year in marketing spend. There is no upper limit; the same model covers budgets in the hundreds of millions. The model needs enough variation between channels and over time to separate their effects. Below that, the uncertainty ranges get wide and a structured cross-channel dataset is often the better first investment.

We built a model with Robyn or Meridian and nobody maintains it. Can a vendor take over?

Ask each vendor directly. At Odins we do not run your old code as it is, but we learn from it: the data cleaning, the channel definitions and what the earlier work found about how your channels behave. The new model is then connected to live data, retrained monthly, and open for scenarios and forecasts any time.

Consultancy or software?

A consultancy gives you a project and a report. Software gives you a model you can query any time, but self-serve tools leave the data and interpretation work with you. A managed platform sits between the two: software you can use every day, with a team doing the data and model work.

How long does it take to get results?

Plan for four to six weeks to a first model from kickoff, with data collection and structuring included; once the data is in, the model itself takes two to three weeks. The slowest part is usually offline media data held by agencies, so start that request early. Reporting on the structured data is useful weeks before the model is ready. Vendors who quote shorter timelines usually count from the day your data is already collected and prepared, and often do not offer data collection at all.

How do we know the model is right?

You cannot know from the fit alone. Ask for stored forecasts compared with actual sales, month after month. In steady state, Odins aims to be within 10 to 15 percent of actuals, and when the model drifts, a person looks at why before any recommendation goes out.

For the full background on how MMM works, read our complete guide to marketing mix modelling. If you are comparing vendors now, book a demo and bring the checklist.

The Odins letter

Where marketing budgets work hardest, once a month.

One short email: a question about marketing investment, answered with numbers, and what it means for the next budget decision.

  • One email a month
  • Written by the team that builds and runs the models

Protected by reCAPTCHA: Google's Privacy Policy and Terms apply.