MMM, attribution, and incrementality each answer a different question and carry a different bias. The operator's move isn't picking one — it's giving each a job, and letting them calibrate each other.
Stop asking "which measurement method is right?" Ask "which one answers this decision?"
Every measurement method is a lens, and every lens distorts. MMM smooths over the tactical detail it can't see. Attribution over-credits the last clickable touch and quietly loses signal to privacy. Incrementality is the truest read you can get — and it's too slow and expensive to run on everything. Pick one and crown it, and you inherit its blind spot as your strategy.
So don't crown one. The 2026 stack treats all three as instruments on the same dashboard — each trustworthy for the thing it's built to read. Tap a vertex below to see what each one is actually for.
"Half the money I spend on advertising is wasted; the trouble is I don't know which half."
The edges are the calibration: incrementality discounts attribution's credit, MMM sets the spending envelope, incrementality validates the model.
The fastest way to sound junior in a measurement conversation is to argue that one method "beats" the others. They aren't competing; they're specialized. Here's the division of labor I run.
| Method | The question it owns | Its bias / blind spot | Cadence |
|---|---|---|---|
| MMM | How much should we spend in total, and how should the envelope split across channels? | Can't see what it can't measure; smooths over tactics; needs history | Monthly / quarterly |
| Attribution | What do I change this week — and what do the ad platforms optimize toward? | Credits correlation, not cause; leaks signal to privacy; recency-biased | Continuous |
| Incrementality | Is this channel actually causing lift, or would it have happened anyway? | Expensive, narrow, episodic; one channel at a time | A few times a year |
Attribution isn't just a reporting tool you can demote and ignore — it's the only operational one. It runs in real time, and the same touch-level signal feeds the optimization engines inside Meta, Google, and Apple. So even where attribution is a biased measure, it's still a live input to performance. You don't get to throw it out; you get to discount it. It's the steering wheel, not the map.
You can't trust an instrument you can't picture. Here's what each method is really doing under the hood — and exactly where its bias creeps in. Three diagrams, three jobs.
Attribution stitches a user's touchpoints into a path and assigns credit for the conversion. Last-click gives 100% to the final touch — simple, and wrong in an obvious direction. Data-driven / multi-touch (MTA) spreads credit using the observed pattern across many journeys. Either way, it's correlational: it can only credit touches it saw, which is why Apple's App Tracking Transparency and cookie loss — which hide the early view-through touches — bias it systematically toward the last clickable step.
MMM never touches a single user. It takes aggregate weekly data — what you spent per channel, plus price, promotions, seasonality and a base-demand trend — and fits a statistical model that asks: when this input moved, how did sales move? Two transforms make it realistic: adstock (advertising's effect carries over into later weeks) and saturation (each channel bends into diminishing returns). The outputs are a decomposition of sales into base + each channel, and a response curve per channel — which is exactly what a budget optimizer needs. Because it's privacy-safe and top-down, it sees brand and offline effects attribution misses — but it's coarse and needs years of clean history.
An incrementality test is a real experiment. You split the audience — or, in a geo holdout, whole regions — into a test group that sees the ads and a control / holdout that doesn't, then measure conversions in both. The control tells you what would have happened anyway; only the gap above it is truly caused by the ads. That gap, divided by the total, is the channel's incrementality factor — e.g. a result of 0.6 means 40% of attribution's claimed conversions were going to convert regardless. It's the closest thing to ground truth, which is why it's the referee. The price: it's expensive, runs one channel at a time, and the read is only as good as the holdout.
They will disagree — that's normal, not a failure. MMM says social is worth 0.8x what attribution claims; a geo test says paid search is 60% incremental. The rule isn't "average them." It's: match the number to the decision you're actually making right now.
And the tiebreaker behind the tiebreaker: a fresh, relevant incrementality test wins. Causal evidence beats correlation, so when you have a recent holdout result on the channel in question, it becomes the referee — you re-weight MMM and attribution toward it, not the other way around. The catch is in the words "fresh" and "relevant": a six-month-old test on a different geo doesn't get to overrule today's data.
"Give each a job" is the default, not the ceiling. When a decision recurs and the methods are stable, the sharper move isn't switching instruments per question — it's fusing them into a single composite KPI and steering on that. I call it the blended read: one number that already has the other methods baked in.
Why $67, not $40? Attribution's $40 only counts the conversions it could see and credit. The incrementality factor of 0.6 says roughly 40% of those would have converted anyway — so the true cost of a genuinely incremental subscription is $40 ÷ 0.6 ≈ $67. MMM's $80 efficiency ceiling confirms you're still buying profitably. One number, all three lenses — and it updates daily off attribution while only needing a re-base when a new test or model refresh lands.
Blend when the decision is recurring and the inputs are stable — daily/weekly spend steering, pacing to a target CAC, channel guardrails. Switch back to "give each a job" when the decision is novel, the stakes are large, or a method just moved (a fresh test, a model refresh, a new channel) — because a blended number can hide a disagreement you actually need to see. The blend is for cruising; the individual instruments are for when the dashboard lights up.
Triangulation isn't three reports in three tabs. It's a loop where each method corrects the next. Run it in this order:
A geo holdout says paid search is ~60% incremental. That 0.6 becomes a discount you apply to attribution's claimed conversions for that channel. Now attribution is anchored to a causal read.
The model says total efficient spend is ~$X and social should be a smaller slice than it looks. That's the budget boundary — how much, and the rough split — that the tactical layer has to live inside.
Day to day, you allocate inside MMM's boundary using discounted attribution — and feed the platforms the signal they optimize on. Fast decisions, but bounded by the slower, truer layers above.
When a channel scales hard or a discount looks stale, you queue the next incrementality test — and the loop tightens. Measurement is a maintenance schedule, not a one-time project.
You don't need a six-figure vendor to play the MMM role. Plenty of MMMs already refresh monthly — but you can still do ad-hoc, directional modeling between the outputs, at any cadence, or when you can't afford one at all. A lightweight response-curve fit won't be precise, but done honestly it's directionally useful and fast — and a timely directional read usually beats a perfect number that arrives after the decision. (Full piece on this coming in the series.)
What "triangulation" looks like on an actual week, month, and quarter — so it's an operating rhythm, not a philosophy.
The methods above aren't proprietary — they're built on open tooling and public research. If you want to go from understanding them to running them, start here.
Set a saturation curve per channel and watch the optimizer split a fixed budget — the allocation logic from this piece, made interactive.
Open the tool →Directional media-mix modeling between the vendor outputs — or when there's no vendor at all.
Read it →Three forecast models that should disagree — and why a range beats a number.
Read it →You can run a handful of clean tests a year — how to choose them like a portfolio.
Read it →The other thing I write about: how to get cited by ChatGPT, Perplexity, and Google's AI answers.
Switch topics →If you're building a measurement stack, pressure-testing an allocation, or just want a second set of eyes on what your three numbers are really telling you — let's talk.
Last updated June 2026 · Part of an in-progress series on growth measurement & budget allocation.