Triangulation in marketing measurement: MMM, attribution and experiments

Triangulation is the advice to measure marketing with three methods at once: attribution, experiments and marketing mix modelling. Many measurement tools are built around it, and many marketing teams ask us whether they need it. This guide explains what it is, where it gets hard in practice, and how we handle the same three sources of evidence at Odins.

The short answer

What triangulation means

The word comes from navigation: you fix a position by taking bearings from several known points. In marketing measurement there are three.

• Attribution. Platform pixels, Google Analytics and multi-touch attribution (MTA) follow tracked users and assign credit for each conversion to the touchpoints before it.
• Experiments. Geo tests, holdouts and conversion lift studies change one thing for one group and measure the difference against a control.
• Marketing mix modelling (MMM). A statistical model of how sales respond to spend per channel, built on aggregated history for the whole business, offline included.

Each method sees something the others miss, so the advice is to run all three and compare the answers. The table shows what each one is used for.

Where triangulation gets hard

The three methods often disagree, and that is where the work starts.

• Someone has to pick the number. When attribution, an experiment and the model give three different answers for the same channel, a person has to decide which one the budget follows. In our experience that decision gets made in a meeting.
• Attribution cannot see offline media. It needs tracked users and digital touchpoints. TV, radio, out-of-home and print have neither, so a setup that leans on attribution tends to undervalue them.
• Ad-level answers need more data than there is. Recommendations at ad-group or ad level sound precise, but there is usually too little data to model effects reliably that far down.
• Experiments are scarce. A good experiment answers one question, for one channel, at one spend level. Most teams we meet have run few of them, so the model has to carry the rest.

One model that takes in the other two

A Bayesian marketing mix model does not start from a blank sheet. It starts from prior knowledge about each channel and updates it against your sales history. That gives the other two methods a place inside the model, instead of next to it.

• Platform data and what your team knows inform the priors. The ad platforms hold detailed data on how each digital channel has responded to spend. Your team knows why budgets were set the way they were, and what changed along the way. Together they set the model's starting assumptions for each channel.
• Experiment results go directly into the model. If you have run a lift test or a geo test, what it showed is taken into account when the model is built and trained. Where the model is still unsure about a channel, it says so, and we propose a structured test to settle it.
• Your sales history has the final say. The model learns from how sales moved when spend moved, across every channel at once, and corrects the starting assumptions where the data disagrees.

What comes out is one answer per channel, with a range that shows how sure the model is. You have one model to check and one set of numbers to plan from. Calibrating a marketing mix model with outside evidence in this way is mainstream practice. Google's open-source framework Meridian documents the same approach.

Inside each platform, its own tracking is the best tool

The platforms' own tracking is the best tool there is for optimising inside each platform. Google, Meta and the others see far more signal about their own auctions, audiences and creatives than any outside model can. For deciding which ad, audience or keyword to run, and for bidding, keep using their tracking and their optimisers.

The limit lies elsewhere. Each platform counts conversions by its own rules: its own attribution window, clicks only or views as well, and only the touchpoints it can see. A buyer who touched two platforms can be claimed by both. Each number is right by its own definition, and the definitions differ. That makes platform numbers a poor currency for comparing one channel with another, and they say nothing about channels that cannot be tracked. The budget questions sit across channels, so they need a measure that works across channels. That is the model's job.

When a tool built around attribution fits better

If what you need most is campaign-level insight inside digital channels, day by day, a tool built around attribution will serve that better than a marketing mix model. Several of those tools now include MMM as well. Many of our e-commerce customers keep one for daily operations and use Odins for the investment decisions.

Odins is a managed marketing investment platform. We connect the data, digital and offline, build and maintain the model, and deliver monthly recommendations that our team reviews before they reach you. It fits companies with a marketing budget above about EUR 1 million a year, spread across more than one channel.

Frequently asked questions

Do we need triangulation?

Not for the budget questions. How much to spend and how to split it across channels can come from one model, as long as that model takes in what the platforms and your experiments have to offer. For running campaigns inside a platform, keep using the platform's own tracking.

Do you run multi-touch attribution together with MMM?

No. We do not run a separate attribution model next to the marketing mix model. Data from the ad platforms is used when we set the model's starting assumptions, so what attribution knows is already inside the model. You get one model to check and one set of numbers to plan from.

Do you use our platform conversion data?

Yes. Odins collects spend, impressions and conversions from each platform. The platform's own data, together with what your team knows about each channel, informs the priors. The model then measures each channel's contribution against your actual sales.

Can lift tests and geo tests be used in the model?

Yes. Results from experiments you have already run are taken into account when the model is built. New tests can be designed with the model, which shows where it is least certain and how large a test has to be to learn something. In a single national market, where regions cannot be treated differently, the usual test varies one channel's spend over time instead.

Does MMM replace Google and Meta attribution?

No. Platform attribution and the platforms' own optimisers remain the right tools inside a channel, at campaign and ad level. The model sits above them and sets the budget level and the split across channels that they then execute.

Is this approach unusual?

No. Most modern marketing mix models are Bayesian, and using outside evidence to set their priors is standard practice. What differs between providers is how carefully the priors are set, and whether you end up with one model or several to reconcile.

For the full background on how the model works, read our complete guide to marketing mix modelling. For the questions to ask any provider about priors and verification, see how to choose a marketing mix modelling vendor. To see what this looks like on your own channels, book a demo.