Your platforms claimed $5.1M in sales last quarter. Your bank recorded $3.4M.
That’s three ad platforms taking credit for the same shopper. MMM fixes it by working backward from total revenue instead of chasing individuals.
We manage over $500M in media spend at Sequence Commerce across Amazon, Walmart, and multi-channel retail. Here’s how MMM works, what marketing data it needs, and one 2026 finding that changes the answer for commerce brands.
What Is Media Mix Modeling?
Media mix modeling, usually shortened to MMM, is a statistical method that estimates the impact of marketing across every channel using aggregated historical data.
It analyzes two to three years of marketing spend, revenue, and outside factors like pricing and seasonality. Because MMM uses no user-level data, it keeps working after cookies and device IDs disappear.
MMM starts with total weekly sales and asks what explains them. Attribution does the opposite, starting with one shopper and tracing the path forward.
Media Mix Modeling vs Marketing Mix Modeling
People use these interchangeably, and mostly that’s fine. There is a real difference worth knowing.
| Marketing mix modeling | Media mix modeling | |||
|---|---|---|---|---|
| Scope | The full commercial picture | Paid channels only | ||
| Variables | Price, product, distribution, promotion | Spend, placement, and return by channel | ||
| Answers | Did the price increase cost more volume than the ads gained? | Which channel should get the next dollar? | ||
| Origin | The original 1960s discipline | The modern, media-focused narrowing of it | ||
| Used today for | Annual commercial planning | Media budget allocation |
Most vendors selling either one build the same thing. Ask which variables they include rather than which label they use.
A Short History of Media Mix Modeling
The technique dates to the 1960s, when packaged goods brands needed to measure the effect of marketing on television, radio, and print. Those media channels produce no click data, so the only way to estimate marketing impact was statistically.
Digital killed the practice for a while. Why infer anything when you can track every click?
Three shifts in the media landscape brought MMM back: third-party cookie decay, Apple’s App Tracking Transparency, and walled gardens that grade their own homework. A model built on aggregates ignores all three.
How Does MMM Work?
MMM work starts with a regression connecting weekly sales to every input that could have moved them. Media spend by channel goes in, along with price, promotions, seasonality, and inventory. The output is a contribution estimate per channel, plus a baseline covering sales you’d have made regardless.
Five ideas carry the whole modeling process.
The Baseline
Some sales happen with zero advertising. Repeat buyers, brand searches, shelf presence.
Separating that baseline from revenue driven by marketing is the most useful thing MMM does, and it’s what vendors gloss over fastest.
If your baseline erodes while ad-driven sales hold flat, your brand is in trouble and no dashboard will say so.
Adstock and Carryover
An ad from three weeks ago is still working. Someone saw a Sponsored Brands placement, didn’t buy, and came back later.
Adstock spreads a week’s media spend forward across following weeks. Get it wrong and upper-funnel channels look worthless.
Saturation
Your first $10,000 in Sponsored Products buys the cheapest relevant clicks. Your tenth buys clicks you’d have skipped.
The saturation curve shows where each channel stops paying. That’s the real answer to “should we spend more here.”
Control Variables
Sales jumped in November. Was that the campaign or was it Q4?
Regression modeling with controls for seasonality, pricing, and stockouts stops MMM crediting media for things media didn’t do.
Calibration
This step decides whether any of it holds up. A model fit to past marketing data alone finds correlations.
One anchored to real incrementality tests proves something. Ask about calibration before you ask about anything else.
How Mix Models Measure Marketing Performance
Once fitted, mix models produce three outputs. A contribution figure per channel. A return estimate per channel. And a saturation curve per channel.
The third one carries the most weight. It’s what lets you simulate marketing scenarios against the model and find the optimal media mix at any given budget level.
A Media Mix Modeling Example
Here’s a worked example for a mid-market CPG brand doing $4M a year. The figures are illustrative, built to show how MMM models behave rather than reporting a client result.
The brand spent $960,000. Platforms claimed $5.1M in sales, which exceeds total company revenue.
| Channel | Spend | Platform-claimed | Modeled incremental | ROAS | Saturation |
|---|---|---|---|---|---|
| Sponsored Products | $420,000 | $2,180,000 | $1,050,000 | 2.5x | Saturated |
| Sponsored Brands | $150,000 | $610,000 | $465,000 | 3.1x | Room to grow |
| Sponsored Display | $90,000 | $340,000 | $153,000 | 1.7x | Over-invested |
| Amazon DSP | $180,000 | $720,000 | $774,000 | 4.3x | Room to grow |
| Meta (off-Amazon) | $120,000 | $1,250,000 | $396,000 | 3.3x | Room to grow |
| Baseline | n/a | n/a | $1,162,000 | n/a | n/a |
What the Numbers Show
Sponsored Products claimed more than double what it delivered. That’s last-click logic catching shoppers who’d already decided.
DSP looked weakest on platform reporting and strongest in the model. Upper-funnel media usually does, because its work lands later, inside someone else’s attribution window. Kochava found the same pattern with TikTok, measuring incremental impact 35% higher under MMM than last-touch (September 2025).
What This Says About Your Marketing Campaigns
The read-across is uncomfortable. Your best-reported marketing campaigns are often your most over-credited ones, because bottom-funnel placements intercept demand that already existed.
Here the decision was to move roughly $130,000 out of Sponsored Products and Display into DSP and Sponsored Brands, holding the total media budget flat. That’s the point of MMM results. Not a new metric on a dashboard, a different budget.
Want to see how your channel mix actually splits? Get a free Amazon PPC audit →
What Data Do You Need to Build a Media Mix Model?
To build a media mix model you need two to three years of historical media data at weekly or daily grain.
The minimum is media spend split by marketing channel, matching sales volume, and controls for price, promotions, and seasonality.
Missing history is the most common reason a model can’t be built at all.
Required, Recommended, and Optional Inputs
| Tier | Data | Why it matters | ||
|---|---|---|---|---|
| Required | Weekly spend across different media channels, 2 to 3 years | Core independent variables | ||
| Required | Weekly sales or units, matched periods | The dependent variable | ||
| Required | Price and discount history | Otherwise promotions get credited to ads | ||
| Recommended | Impressions, reach, frequency | Spend alone hides efficiency shifts | ||
| Recommended | Inventory and out-of-stock dates | Stockouts look identical to ad failure | ||
| Nice to have | Incrementality test results | Anchors the model to causality |
The two-year rule is a hard gate. If you launched eight months ago, nobody can create an MMM model for you.
Amazon’s MMM Data Feed
This is the part almost nobody covers, and it’s the biggest development in years for Amazon-first brands.
Amazon runs its own data feed, supplying privacy-safe aggregated advertising and retail signals already formatted as MMM inputs. You request access through a self-service portal in your manager account.
What the MMM API Provides
In May 2026 Amazon moved the feed’s API to general availability, per the Amazon Ads Advanced Tools Center release notes. That ended a beta restricted to invited partners.
It covers 14 markets: the US, Canada, Mexico, Brazil, Germany, Spain, France, Italy, the UK, Saudi Arabia, the UAE, Australia, Japan, and India. The API is asynchronous, so you submit a report request, it processes in the background, and you retrieve the output. Six endpoints shipped at launch, covering brand group overrides, product and campaign retrieval, and report status.
Most media platforms supply media metrics only: impressions, clicks, spend, attributed conversions. Amazon adds retail signals from actual marketplace sales, already aligned for modeling.
Brand Groups
A brand group is Amazon’s container for related products and campaigns, and the model treats each as a unit. If Amazon’s default grouping doesn’t match your marketing portfolio, override endpoints let you fix it programmatically.
A badly specified brand group produces a badly specified model.
Why It Changes the Math
Without this feed, your modeler is guessing at Amazon using spend exports and sales proxies. Amazon comes out either inflated or invisible.
With it, Amazon enters on equal footing with every other channel. If you already run Amazon Attribution for off-Amazon traffic, that becomes a cross-check rather than a competing story.
MMM vs Attribution: Which Measures What
For brands spending over $50,000 a month across three or more channels, MMM should drive budget allocation while attribution handles daily optimization. Running only one is how brands over-invest in whichever channel reports most aggressively.
| Media mix modeling | Multi-touch attribution | Incrementality testing | ||
|---|---|---|---|---|
| Data type | Aggregated, no tracking | User-level journeys | Controlled experiment | |
| Answers | Where budget should go | Which touchpoints preceded the sale | Did this cause lift | |
| Covers walled gardens | Yes | No | Yes | |
| Privacy-resilient | Yes | No, degrading | Yes | |
| Time to result | 6 to 12 weeks | Near real time | 2 to 6 weeks per test | |
| Proves causation | No, infers it | No | Yes | |
| Best for | $50k+/month, 3+ channels | Single-channel in-platform tuning | Validating one specific claim |
The tradeoff is coverage against precision. Unlike MMM, multi-touch attribution sees every step, but only where tracking survives, which increasingly means one platform at a time.
Where Data-Driven Attribution Modeling Fits
Data-driven attribution modeling sits between the two. Rather than applying fixed rules like last-click, it learns credit allocation from observed conversion paths.
Better than last-click, but it still depends on user-level tracking. So it inherits every privacy limitation that made MMM relevant again. Treat it as an upgrade to your in-platform layer, not a portfolio-level answer.
How to Optimize Marketing Spend Using All Three
MMM provides the portfolio view. Attribution guides daily optimization. Experiments prove whether the model is telling the truth.
Run them together and each one covers the others’ blind spots. If you’ve run Amazon brand lift studies, those results are exactly what anchors MMM to causality.
Not sure which method fits your margin goals? Talk to our team →
Why Mix Models Undercount Retail Media
On April 7, 2026 the IAB published a white paper arguing that MMM is structurally ill-suited to retail media, causing brands to undervalue one of their most productive channels. Collin Colburn of the IAB and Priyash Shahane of Instacart wrote it, and they identify three specific mismatches.
If Amazon or another retail media network is a major line in your budget, this is the most important section here.
Always-On Spend Has No Variation
MMM needs variation to detect impact. A brand runs TV for four weeks, stops, and sales move. The model reads that signal.
Retail media never stops. Budgets run continuously and rarely vary in ways mix models can detect. No variation means no measurable incrementality, so that marketing activity gets pushed into the baseline and read as background noise.
Buying dynamics compound it. Traditional media plans on CPMs, impressions, and GRPs, which translate cleanly into MMM inputs. Retail media runs on CPC, CPA, and algorithmic auctions that shift several times a day. Those don’t map onto the exposure-based inputs MMM expects, so performance gets structurally underweighted.
Commerce Data Is Fragmented
MMM is typically fed point-of-sale data from a limited set of physical retailers. Retail media drives sales across delivery, pickup, marketplace, and cross-retailer spillover.
The paper cites Instacart: over 2,200 retail banners connected with 310 or more grocery ecommerce sites. The lift is real but has nowhere to land in a model that doesn’t ingest those sources.
Separate research found 50% of MMM users report commerce media as underrepresented in their models (February 2026).
Better Measurement Already Exists
MMM was built for environments where direct measurement was impossible. Retail media provides the opposite: closed-loop, SKU-level, transaction-level data.
Where MMM infers impact, closed-loop reporting proves it. MMM also tends to lump all retail media into a single marketing tactic, hiding campaign-level differences.
What the IAB Recommends
The paper doesn’t say abandon MMM. It says reposition it.
- Use MMM for long-term marketing planning. Portfolio scenarios and budget guardrails. Don’t let it set the exact investment level in closed-loop channels.
- Put incrementality testing at the center for retail media, not at the periphery.
- Audit your decision logic. If MMM consistently under-credits commerce media versus experimental results, that’s a measurement mismatch, not a media performance problem.
For context, the IAB forecasts 12.1% growth in commerce media spend for 2026, with the global category approaching $100 billion by 2028.
Three Mistakes Amazon Sellers Make
Most vendors come from a CPG and television background, and they treat Amazon as one line item. For a brand doing 60% of revenue there, that’s a modeling failure before the first regression runs.
Collapsing Amazon Into One Channel
Sponsored Products, Sponsored Brands, Sponsored Display, and DSP behave nothing alike. Sponsored Products saturates early. DSP has a long carryover.
Model them together and you get an average that describes neither.
Ignoring the Halo Effect
Your Meta campaign sends someone to Amazon. Meta doesn’t see the sale, Amazon doesn’t see the source, and the shopper vanishes between systems.
This is the biggest blind spot for brands spending $10k to $500k a month, and it’s exactly what MMM exists to catch. Same dynamic behind why TACoS predicts account health better than ACoS.
Leaving Inventory Out
Amazon sales drop during stockouts. If out-of-stock dates aren’t a control variable, MMM reads that drop as ad failure and tells you to cut a channel that was fine.
Based on our own account work, this is the most expensive error on the list and the easiest to fix, because the data already sits in Seller Central.
If Amazon DSP is meaningful for you, insisting it enters as its own channel is the highest-value thing to demand from a vendor.
Media Mix Modeling Tools
Options split three ways: open-source frameworks you run yourself, a commercial MMM platform you subscribe to, and consultancies building a custom MMM solution per engagement.
Google Meridian
The most-used open-source option, released globally on January 29, 2025 alongside 20-plus certified partners.
Meridian is built on Bayesian causal inference, models reach and frequency rather than raw impressions, and accepts incrementality results as priors. A September 2025 update added pricing and promotion variables, channel-level contribution priors, and binomial adstock decay for longer-term upper-funnel effects.
Other Open-Source Options
PyMC-Marketing offers custom priors, configurable adstock, and time-varying baselines, aimed squarely at people who write code. Meta Robyn is the other well-known framework and doesn’t use a Bayesian approach.
One caution applies to both platform-built tools. When the party providing the model also sells the media, validate the outputs independently.
Commercial Platforms
These handle the pipelines, refresh continuously, and produce output a CFO can read without a statistics degree. You pay for that, and you inherit assumptions you can’t always inspect.
Open-source is only free if you already employ someone who can run it. Budget three to six months of a data scientist’s time, then reread that 8% figure on in-house analytics skills.
Should You Build an MMM Model?
MMM is best suited to brands spending roughly $50,000 a month across at least three channels, with two years of clean history behind them. Below that, the cost of building and maintaining a model usually exceeds the value of the reallocation it recommends.
Three questions settle it.
Do You Have 24 Months of Clean Data?
Not partially. Spend, sales, price, and promotions, weekly, for two full years.
If you migrated platforms last year and lost the old data, you’re not ready. Start collecting now so you are next year.
Do You Spend Across Three or More Channels?
MMM exists to answer “which channel.” If 90% of your marketing budget sits in Sponsored Products, the answer is already obvious.
Can You Act On the Output?
A recommendation to shift 15% into DSP is worthless if nobody can move the money. We’ve watched brands buy models and then ignore them through two full planning cycles.
What to Do Instead
If any answer is no, run structured incrementality tests. Pause a channel in half your geographies for four weeks and measure the gap.
Cheaper, faster, casual. It also builds the test library you’ll use to calibrate a real model later. At that stage, getting your daily PPC monitoring system right and knowing your target ACoS by margin returns more than any statistical exercise.
How to Choose a Media Mix Modeling Company
The vendors worth hiring show their work. They explain baseline assumptions, share their adstock and saturation choices, and let you challenge the priors. Anyone who won’t is selling a black box, regardless of whether the code underneath is open-source.
Ask these five questions before signing.
- “Walk me through the baseline decomposition.” What share of sales is baseline versus media-driven, and how did you land there? A vague answer means the number was a default nobody examined.
- “How are you modeling Amazon and retail media?” Push back if the answer is “as one retail media channel.” Then ask what they make of the IAB’s April 2026 findings. A partner who hasn’t read it isn’t current.
- “What priors are you setting, and why?” Bayesian models start with assumptions, and those assumptions shape results.
- “How do you calibrate against experiments?” Ask specifically how incrementality results feed in as priors.
- “Who acts on the output?” A vendor hands you a report. An agency running your media can implement the reallocation and measure what happened. Be cautious when the party doing the measuring benefits from the result.
Conclusion
MMM answers what dashboards can’t: what each channel actually contributed, once you strip out the sales you’d have made anyway. Two years of clean marketing data, real spend across three or more channels, and the authority to act on the result decide whether it’s worth building.
For commerce brands there’s a newer wrinkle. The IAB’s April 2026 research found MMM undercounts retail media, so use it to guide portfolio strategy while closed-loop measurement and incrementality govern retail media decisions specifically.
At Sequence Commerce, Jake Gilbert and the advertising team have found the biggest gains rarely come from the model itself. They come from splitting Amazon’s ad types into separate channels and accounting for the off-Amazon halo that no platform reports.
If you want a clear read on where your Amazon budget is working, get a free Amazon PPC audit from Sequence Commerce →
Frequently Asked Questions
What Is a Media Mix Model?
A media mix model is a statistical model that estimates how much each marketing channel contributed to sales using aggregated historical data. It analyzes two to three years of media spend, revenue, pricing, and seasonality to separate baseline demand from advertising-driven demand.
How Accurate Is Media Mix Modeling?
MMM is directionally accurate rather than precise. A well-built model reliably identifies which channels are saturated and which have room, within a stated confidence range. Accuracy depends heavily on marketing data quality and history length.
How Long Does the MMM Process Take?
Most first builds take six to twelve weeks from data handover to usable output. Collection and cleaning consume the majority of that time, not the modeling itself. Brands with clean centralized data move faster.
Can Small Brands Use MMM?
Small brands usually shouldn’t start here. Below roughly $50,000 in monthly media spend, the cost of building and maintaining a model exceeds the value of the reallocation it produces. Incrementality testing gives smaller brands causal answers faster and cheaper.
Does Amazon Provide Data for Media Mix Modeling?
Yes. Amazon operates a data feed supplying privacy-safe, aggregated advertising and retail signals already formatted as MMM inputs. Advertisers request access through a self-service portal in their manager account.
Why Does MMM Undercount Retail Media?
According to an April 2026 IAB white paper, three structural mismatches cause it. Retail media runs always-on, so it lacks the spend variation MMM needs to detect impact. Commerce sales are fragmented across sources most models don’t ingest.
Jake Gilbert serves as the Chief Revenue Officer (CRO) at Sequence Commerce, where he spearheads all lead generation initiatives and drives the agency’s growth strategy. With a keen analytical mindset, Jake transforms complex marketplace data into actionable insights that deliver measurable results for clients. His expertise spans conversion rate optimization, customer acquisition funnels, and performance marketing specifically tailored to Amazon and e-commerce ecosystems.
At Sequence Commerce, Jake has developed proprietary reporting frameworks that accurately track and attribute revenue across multiple marketing channels, allowing clients to optimize their ad spend with precision. His data-driven approach to revenue optimization has helped numerous brands scale their Amazon presence while maintaining profitable growth metrics. Jake regularly analyzes emerging marketplace trends and platform algorithm changes to refine client strategies and maintain competitive advantages in the increasingly complex e-commerce landscape.
Before joining Sequence Commerce, Jake built successful sales operations for SaaS companies targeting the e-commerce sector, giving him unique insights into both the technical and strategic aspects of online selling. He holds certifications in advanced analytics and digital marketing, and frequently shares his expertise through the Sequence Commerce blog and industry webinars.