Clients ask their media agency the budget question more often than they ask anyone else: should we spend more or less, and where? Platform reports cannot answer it, and an agency that answers with numbers from the channels it buys is grading its own homework. Marketing mix modelling (MMM) is the method built for that question, which is why more agencies want to offer it. This guide covers the routes an agency can take, which platforms agencies use, how a partnership with a managed platform works in practice, how the money works, and how to keep the answer independent.
A media agency has three ways to offer marketing mix modelling to its clients:
Two things decide the route: whether you have marketing scientists who can own models for years, and how many of your clients are big enough for a model at all. For most mid-sized and independent agencies the answer to the first is no and to the second is a handful. That points to a partnership with a managed platform, plus a data offer for the rest of the client list.
These are the options agencies most often look at. We run one of them, so we say so up front.
| Platform | How agencies work with it | Who runs the model | Best for |
|---|---|---|---|
| Odins | Agency partnership: the agency stands in front of the client, Odins runs the data and the models in the background. A data platform for the whole client list and modelling for the clients whose budget justifies it. | Odins, with the agency challenging the assumptions | Mid-sized and independent agencies in Europe without a modelling team |
| Mutinex | An agency partnership programme with platform access, onboarding, training and support from its own marketing scientists. | The agency, with vendor support | Agencies whose clients sit in Australia or the US |
| Recast | Three ways of working: a fully managed service for clients, shared platform access, or a referral programme. | Recast or the agency, depending on the route | Agencies with an analytics team of their own |
| Google Meridian | Open source, with a partner certification. Partners build the data pipelines, set the priors, and run and maintain the models. | The agency | Large consultancies and holding-company agencies with a marketing science team |
| PyMC-Marketing, Meta Robyn | Open-source frameworks. The data work, the verification and the client reporting are yours. | The agency | Teams with strong statistics skills who want full control |
Odins runs the data integration, the Bayesian model and the monthly verification, and reviews every recommendation before it goes out. The agency stands in front of the client. Two offers sit on one platform: structured marketing data for the whole client list, and modelling for the clients whose budget justifies it. Odins buys no media and makes no creative, so it does not compete with the agency for the account. More on the Odins agency partnership.
Mutinex gives partner agencies access to its platform, onboarding and training, and support from its own marketing scientists, aimed at mid-sized and independent agencies that want to compete with holding-company modelling. It operates from Australia and New York.
Recast describes three ways of working with agencies: a fully managed service on behalf of clients, shared access to the platform, and a referral programme. The platform is built with analytics and data science teams in mind.
Meridian is free and Bayesian, and Google certifies partners that deploy it for clients, mostly large consultancies and holding-company agencies. The agency builds the data pipelines, sets the priors, and runs and maintains the models. The trade-offs are covered in Google Meridian: build vs buy.
Two more open-source frameworks. PyMC-Marketing is the most flexible and needs the strongest statistics skills. Robyn gets to a first run faster, but it is not Bayesian and is no longer under active development at Meta. Both leave the data work, the verification and the client reporting to you.
An agency that buys the media and also measures it will be asked whose number it is. The question is fair even when the work is excellent, because the proof of effect comes from the side that is paid to spend. Finance feels it most. It is also the first objection a competing vendor will raise when it pitches your client.
The answer is not to stop measuring. It is to put the model where the client can trust it:
The position this gives you
With that in place, the agency stops defending its own reporting and starts presenting an independent number. That is a stronger position in the budget meeting, not a weaker one. The agency's advice sits on top of the model; the model does not sit on top of the agency's media.
MMM needs scale and variation. A client with a marketing budget below about EUR 1M a year, or with spend that never changes level, gets back mostly the model's starting assumptions. In a typical agency portfolio that leaves a handful of clients where a model pays off.
The rest still have a data problem worth solving. Spend sits in ten ad accounts, TV in a media plan and sales in a finance export, each with its own names. Structuring all of it in one hierarchy per client, digital and offline together, gives reporting, budget tracking and an AI assistant that can answer questions about the client's spend from day one. That is an offer the agency can take to almost every client. The model is added where the volume justifies it, on the same data.
| The data platform | Modelling, added on top | |
|---|---|---|
| For which clients | Almost all of them | The few whose budget and channel mix justify a model |
| What is in it | Every API channel connected, offline and manual spend structured, one hierarchy per client, reporting against budget and targets, data out to the client's warehouse and through MCP | A Bayesian model of what each channel adds, a recommended budget level and split every month, scenarios with confidence ranges, structured tests where the model is uncertain |
| The question it answers | What happened, where, against plan | What worked, what the next unit of budget buys, where the ceiling is per channel, what to change next quarter |
| What the agency does | Picks the channels, decides the structure, delivers offline files | Challenges the assumptions, presents the result, advises the client |
| Useful from | The first weeks | Six to eight weeks after kickoff, then every month |
A practical screen for which clients get a model: enough budget that the split across channels matters, several channels (ideally TV and digital), weekly sales going back about two years, and a marketing lead who has to defend the budget to the CEO or CFO. We go through the data side in data for marketing mix modelling.
This is the form of collaboration we set up with partner agencies. It is built so that the agency does the parts where client knowledge beats technique, and the platform does the parts that are machinery.
Onboarding a client has four steps. Two of them are machinery and one is where the agency does the job.
A partnership works when each side does the part it is best placed to do. This is the split we start from. It shifts as the agency takes more on.
| Area | The first clients | When it runs by itself |
|---|---|---|
| Ad platform connections | The platform sets them up, monitors them and flags when access needs renewing | The platform |
| Onboarding a new client | The platform together with the agency | The agency |
| Structure and hierarchy | The agency, with advice from the platform | The agency |
| Offline and manual spend | The agency owns it, the platform helps closely | The agency, through a pipeline |
| Sales data | The agency opens the door at the client, the platform sets up the feed | The platform |
| The model | The platform builds it, retrains it monthly and checks every recommendation | The platform, as close to the agency as it wants |
| Client meetings | The platform joins the early model discussions | The agency in front, the platform behind |
Two rows deserve a decision early. Offline spend is the one recurring task on the agency side, and it is good agency work: media plans and spot reports for TV, radio, print and outdoor each month, which the agency already holds. And client meetings: the platform joins the first rounds to explain the model, then the agency stands in front and the platform supports from behind.
Once a client has a model, the work settles into a rhythm: a monthly reading of the model, moves in the plan, and a quarterly strategy conversation against the budget. The monthly update can come from either side, so settle who sends it before the first one goes out.
Because the agency does part of the work, the commercial model follows the division of labour. In our partnerships it has three parts, and the numbers are set per partnership:
The principle is that the agency is paid for the clients it brings in and the work it takes on, and the client pays one price for a platform and a model it can trust. The terms tighten as the volume grows, and we say that up front so it does not come as a surprise later. What the agency charges the client for its own advice on top is the agency's business.
The client gets one view of marketing spend and results, a model of what each channel adds, a recommended split and budget level every month, and scenarios with confidence ranges it can take to the board. Moving budget toward channels with headroom typically gives 5 to 15 percent more effect from the same budget, from moving spend rather than adding to it.
The agency gets an independent number under its advice and a seat in the budget conversation it used to lose to finance. It also gets a new service line without hiring a modelling team, and a reason for clients to stay.
Two conversations, not one
The model opens two different conversations with the client. The mix conversation: same budget, different split, nobody asks for more money. And the level conversation: if the client wants to hold a target cost per new customer, what does the budget have to be? The second is a board matter, and it is the one agencies rarely got to have before.
The wider checklist for any MMM vendor is in how to choose a marketing mix modelling vendor.
Pick one client you know well, with a real budget decision coming up. Agree who does what on that client, connect the data, build the first model together and present it to the client side by side. Decide the form of the partnership once both sides have seen it work on real numbers. Planned well, the client goes into its next media year with the model in hand.
Yes, through a managed platform. The platform runs the data integration, the modelling and the verification, and the agency owns the client relationship and the advice. Building on an open-source framework needs a marketing science team that can own models for years.
Odins, Mutinex and Recast have agency partner models. Google Meridian, PyMC-Marketing and Meta Robyn are open-source frameworks an agency can build on. Odins runs the models for the agency and pairs them with a data platform for the whole client list.
It can look like one. The fix is an independent model, run by a party with no media business, with forecasts stored before the actuals arrive and visible to the client. Then the agency presents the number instead of defending it.
Clients with a marketing budget above about EUR 1M a year, several channels and about two years of weekly sales history. Smaller clients still benefit from structured data and reporting across all their channels.
With Odins, through a share of the licence revenue on the clients the agency brings and serves, and a lower platform price when the agency does the integration work. The exact terms are set per partnership. The agency's own fee for its advice is separate.
Not with Odins. The agency stands in front of the client. Odins joins the early model discussions and the quarterly meetings to explain the model, and supports from behind.
Read more about the Odins agency partnership, or start with our complete guide to marketing mix modeling. To talk about a first client together, book a meeting.