The Measurement Trinity: MMM, Attribution & Experiments

I tell every cohort the same thing on day one, so I'll tell you too: everything we do in marketing measurement is trying to answer one question. How much is my channel worth? How much is the campaign worth? You're trying to assign credit. That's it. That's the whole job.

The reason people call this a trinity — or a triangle, depending on which corner of LinkedIn you read — is that we answer that single question through three different methods. Marketing mix modeling. Attribution. Experiments. And the reason it's worth knowing all three is that each one is trustworthy in a different situation, biased in a different direction, and almost never gives you the same number as the others. If they're not in the same zip code, you have a problem. If they are, you can start to believe yourself.

Let me walk you through the three, honestly, including where each one quietly lies to you.

The three methods, one question

MMM is a very complicated linear regression. That's the unglamorous truth. It's top-down. You feed it aggregated stuff — paid search impressions, paid social impressions, channel-level budgets, did you do TV or not — and you ask it for a coefficient per channel against your outcome, usually sales. "Paid search had this much influence on sales." It's the only method in the trinity that can see your offline spend at all, which is why brand teams love it. It needs a lot of data and it averages everything over years.

Attribution does the exact same thing, but bottom-up. Instead of budgets moving up and down, you're chasing people all over the internet — which ads did they see, which did they click, which did they convert on. It's granular, it's basically real-time, and it is digital-only. If you work somewhere that's 80, 90, 100% digital, attribution done correctly isn't that far off. If half your budget is above the line — TV, out of home — then you have problems. Attribution simply cannot see those people.

Experiments are the gold standard. A control group, a test group, identical in everything except one sees the ad and the other doesn't. When properly designed you get unambiguous incrementality. But "gold standard" does not mean unbiased — Google's own in-platform A/B tests carry biases, like the fact that you've already won the auction before you get split into test and control. And we will bang on "properly designed" for a week, because statistical power is the poor cousin to statistical significance and it's where everyone cuts corners. Experiments are what you use to validate and calibrate the other two — mostly the MMM.

So three lenses on one question. Each sees a part of the truth and is blind to the rest. The goal is never to crown one winner. The goal is triangulation.

Code is cheap. Intuitions are the moat.

Before we go further, a confession about what this discipline actually is now.

Code is cheap, and it's getting cheaper every single day. LLMs are getting incredibly good at generating correct code. I can have an agent scaffold a Meridian model faster than I can explain the imports. So if the only thing you take from a course like mine is "I can write the Python," you've taken the wrong thing — that's the part the machine does for free.

What you actually need is the theoretical intuition. Why does this variable make sense in the model? Why doesn't that one? When someone asks "can we add CRM campaigns?" — do you know whether you're opening a door you wanted closed? That judgment is the moat. The code is a commodity wrapped around it.

Which brings me to the thing I want tattooed on every analyst who comes through:

Please do not do glorified pivot-table making.

There is so much more to marketing analytics than glorified pivot-table making. And let's be honest — that's about to get automated away from us anyway, so it's not even a good career bet. The interesting, durable, un-automatable work is causal. Because here's the thing almost nobody says out loud: most of what marketing analytics is could be causal analytics. When stakeholders come to us, they're rarely asking us to forecast. They're asking why did this happen? Why did this campaign go up? Why did conversion rate drop? Does this channel work? Those are all causal questions. And a causal question means you can correctly identify cause and effect — which, it turns out, is hard.

DAGs: writing down your assumptions

We've all heard correlation does not imply causation. Fine. So how do you actually split them? You start by drawing your assumptions down, and the tool for that is a DAG — a directed acyclical graph.

Don't let the name scare you. It's nodes (circles — your channels, your sales) and edges (arrows — "X causes Y"). Directed because causation has a direction. Acyclical because you can't loop back to where you started by following the arrows. That's the whole vocabulary.

People sometimes push back — "why do I need to draw a diagram, I know my business." Here's why. Drawing the DAG forces you to write your assumptions down, and writing them down is exactly what surfaces the weird, counterintuitive relationships you'd otherwise miss.

My favorite example. In Australia, ice cream sales and shark attacks are correlated. If you look at the data naively, ice cream appears to cause shark attacks. Obviously it doesn't. The DAG is just badly written. The real structure is summer — summer drives more people to the beach (more shark attacks) and summer drives more ice cream consumption. Summer is a common cause sitting behind both. Draw it properly and the nonsense evaporates.

That's the entire game. There are three shapes you need to recognize.

Confounder — control for it

A confounder is a node with arrows coming out of it into two places, one of which is on the path you care about. The classic in marketing is seasonality. It affects your sales and it affects your TV spend. So when you look at the raw relationship between TV and sales, you're seeing two streams of water flowing together — the direct effect of TV, plus the seasonality riding along underneath. You can think of it like a river: you need to cut it.

How do you cut the river in a regression? You control for it — you add the variable to the model. If you don't, your TV coefficient gets biased upward, picking up the compound effect that isn't really TV. So: it's a confounder, it causes both, we must control for it.

Mediator — be very careful

A mediator has an arrow coming in and an arrow going out. It sits on the path and passes the effect along. Branded search is the textbook case: your advertisement drives people to search your brand, the brand search drives a purchase through paid search. Brand search is mediating the effect of your advertising.

Here's where people get dogmatic — "control for confounders, remove mediators." I don't think that's a hard rule. If you naively throw branded search into a model that already has paid search and TV, the mediator steals signal from them and your TV and paid-search coefficients shrink. So my rule of thumb is softer: always control for confounders; be very careful controlling for mediators. Only control a mediator if you specifically want the direct effect rather than the total effect. Know what you're measuring before you add it.

Collider — never control for it

A collider is the opposite of a confounder: a node with two arrows flowing into it that is not your outcome. The quintessential collider is the outcome itself — everything flows into sales and nothing leaves it, which is exactly why you never put sales in as a predictor. The rule here is absolute: never control for a collider. Open that door and you manufacture associations that don't exist.

If you want the one-line version of all of this: control for common causes, not consequences. Confounders are common causes — close those backdoors. Colliders and (usually) mediators are consequences — leave them alone. This isn't kitchen-sink econometrics where you throw every variable you have at the model. The whole point of the DAG is to decide what belongs, so you don't get omitted-variable bias on one side or open spurious paths on the other.

And to be fair — there's no mathematical rule for drawing the right DAG. The same set of variables can support several different causal structures, and reasonable practitioners disagree. Recast, for instance, doesn't treat seasonality as a confounder and won't control for it. The point isn't that there's one true diagram handed down from the heavens. The point is that drawing it forces the conversation, and gives you the language to have that conversation with a vendor instead of nodding along.

When each method is trustworthy — and how it's biased

Now layer the DAG thinking back onto the trinity, because each method's blind spots are really just different DAG problems.

MMM is trustworthy when you have the data for it — two years of weekly data at an absolute minimum, three years in the sweet spot. Less than that and don't even spend the money; you'll get pretty point estimates wrapped in confidence intervals so wide they're useless, and you'll fool yourself. Its bias is structural: it averages effects over years, the carryover and saturation parameters are genuinely hard to identify, and if your DAG is wrong — if you forgot a confounder — your coefficients are wrong even when the model fits beautifully. Which is the trap I'll come back to constantly: a model can be accurate and still have every channel coefficient wrong, as long as the channels are correlated enough.

Attribution is trustworthy in a digital-heavy world and near-useless the moment real money goes above the line. Its bias is direction-of-sight: each platform sees its own conversion funnel and nobody else's, so it grades its own homework and skews credit toward itself. Treat in-platform numbers as an upper limit, not the truth — it's the best that platform can possibly claim for itself.

Experiments are the closest thing to ground truth, which is why they're what we use to calibrate the rest. Their bias is that they're a snapshot — a few weeks, one season, one slice of geography — while the MMM is trying to describe years. A short test won't capture a twelve-week carryover effect on a branding channel; shut TV off for four weeks and you simply haven't waited long enough for the effect to accrue. And an underpowered experiment isn't evidence at all — it's noise wearing a lab coat. My personal rule: I do not calibrate a model with an underpowered experiment. I'd rather trust a sensible default than inject confident-looking garbage.

Triangulation is the goal

So no, you don't pick a side. MMM, attribution, and experiments will never give you the same point estimate, and anyone who tells you they should is selling something. What you want is for them to land in the same neighborhood, and for each to cover the blind spots of the others — attribution for tactical weekly digital decisions, MMM for the top-down picture including offline, experiments to anchor and audit the whole thing. Use the right tool for the question it's actually built for. Don't use MMM for weekly tactical optimization; that's attribution's job. Don't trust a digital-only attribution model to tell you whether TV works; that's what the experiment and the MMM are for.

The methods are just three lenses. The skill — the thing the LLM can't do for you, the reason this isn't glorified pivot-table making — is knowing which lens to trust for which question, and being able to draw the DAG that tells you why.

---

If this way of thinking is new, or if you've been doing the pivot-table version of this job and you can feel it about to be automated out from under you, that's exactly the gap I built my course around — the intuitions, not the syntax. I keep a Discord open for past cohorts and I'm always happy to look at a real DAG over a messy real dataset. You can find the course and reach me at marketingscience.dev. Come with questions. Ideally bring your own data.