Media Mix Modeling, Explained With a Worked Example

What media mix modeling actually is

Media mix modeling (MMM) is a statistical method that uses historical sales and marketing data to work out how much each channel — TV, paid search, social, display, radio — actually contributed to sales, separate from what would have happened anyway. It answers a specific question: if this campaign hadn't run, what would sales have looked like?

The name comes from the marketing mix — the four Ps (product, price, place, promotion) that Neil Borden and E. Jerome McCarthy described in the 1950s and 60s. MMM applies regression to that mix: it takes several years of weekly or monthly data and fits a model where sales is the outcome and price, distribution, promotions and media spend are the inputs. The model then splits total sales into base sales (what you'd sell with zero marketing — brand strength, distribution, seasonality) and incremental sales (the lift attributable to specific marketing activity).

"Marketing mix modeling" and "media mix modeling" get used interchangeably across the industry now — MMM the acronym covers both. Wikipedia draws a technical line between the broader four-Ps model and the media-only subset, but in practice almost nobody observes it. Search "mmm media mix modeling" or "media mix modeling mmm" and you'll land on the same regression-based method either way.

How it works, with a real number attached

A basic MMM runs in four stages: collect data, build the model, decompose the result, then optimize the budget against it. None of that is exotic — it's multivariate regression, sometimes Bayesian, run against a wide and often messy dataset.

Here's what makes it concrete. Say a brand runs TV, paid search and display over a year. Last-click attribution will credit paid search with most of the conversions, because it sits closest to the purchase, and will likely credit display with almost nothing — someone saw a banner three weeks before they searched and bought, and the click takes all the credit. MMM, and incrementality testing, look at it differently: they measure what happened to sales in the weeks and markets where display spend moved, independent of who clicked what.

That gap between last-click and true incremental contribution isn't theoretical. Across 30 Amazon DSP advertisers we managed in July 2026, the book delivered a 6.04x return on ad spend, 78.4 million impressions at a $4.00 CPM, and a blended $1.42 cost-per-click — none of which last-click alone would surface, because display's job in that mix was to lift search and direct response, not to be the final touch. That's the entire argument for running MMM, or a holdout test, instead of trusting the last click.

MMM vs. multi-touch attribution vs. incrementality testing

Three names come up constantly and they are not the same thing. Multi-touch attribution (MTA) tries to assign credit to individual touchpoints using tracked user-level data — granular, but it degrades as cookies and device IDs disappear. MMM works at an aggregate level, usually weekly and by market, and doesn't need user-level tracking, which is a big part of why it survived the privacy shift better than MTA did. Incrementality testing — holdouts, matched-market tests — is the most direct of the three: you withhold spend from a group and measure the actual difference. MMM estimates what a holdout test proves.

On vendors: Nielsen has run marketing mix modeling for consumer packaged goods brands for decades, built on its syndicated retail sales and media data — a strong fit if you're already a Nielsen data client and want MMM folded into that relationship. Neustar, now part of TransUnion, built its MarketShare practice on similar CPG heritage and tends to suit larger advertisers who want a vendor to run the whole model rather than build it in-house. Google published Meridian, an open-source Bayesian MMM tool that succeeded its earlier LightweightMMM — genuinely strong if you have in-house data science capacity and want to own the model, but it assumes you can staff that work.

None of these three tell you what your Amazon DSP display did on Amazon specifically, because Amazon's on-platform signal doesn't leave Amazon. That's a data-availability limit, not a methodology flaw — any model is only as good as what you can feed it.

The mistakes that break an MMM model

The most common failure is running it on too little history. Under a year or two of weekly data, the model can't separate seasonality from media effect, so it credits December sales to whatever ran that month rather than to the December baseline that would have happened regardless. A second failure: not validating against a holdout or a real experiment, so the model's coefficients are its own opinion about itself, with nothing checking the answer.

Here's a mistake we've made ourselves: trusting last-click on a display campaign because it was the number on the dashboard, and concluding display "underperformed" — when reconciling in Amazon Marketing Cloud later showed it had been lifting search conversion the whole time. Last-click cannot prove incrementality and never could. It can only tell you the order events happened in. That's why we now attach three things to every spend change before it runs: the evidence behind it, a measurement plan, and a rollback trigger.

What to do when the model gives you a bad number

Sometimes the model tells you a channel you like isn't working, or that committed spend should move. Before acting: check whether the input data has a gap or a structural break — a stockout, a price change, a category disruption — that the model will misread as a media effect. Re-run with a shorter and a longer window; if the coefficient flips depending on the window, you don't have an answer yet, you have noise.

If the model and a holdout disagree, trust the holdout. MMM is an estimate; a matched-market or AMC-based holdout is closer to a fact. And if a fix based on the model's output didn't move the number, say so and roll it back — don't keep an expensive change alive just because it was expensive to make.

Where Dr. DSP fits

Dr. DSP is Amazon's Demand-Side Platform — programmatic display, video and audio buying inside Amazon, not the Delivery Service Partner courier franchise — run as a managed product by Fable 5, part of Full Circle, which has managed more than $500 million in Amazon ad spend across 100+ brands. Rather than building a full marketing mix model, the approach reconciles DSP and sponsored ads inside Amazon Marketing Cloud and validates changes with holdouts and matched controls — a narrower, Amazon-specific answer to the same incrementality question MMM asks across the whole marketing mix. There's no published price: a demo, the first 30 days free, and pricing set on a call against real budget and scope, with the Orbit software suite included either way.