Marketing mix modeling for trade and media spend

The trouble: these levers fire at the same time, in overlapping regions, with lagged effects. A TV campaign might lift sales for weeks. A price promotion spikes volume immediately but can pull forward demand (people stock up, then buy less later). Untangling this by eye is impossible.

What MMM actually measures

MMM works at the aggregate level. It takes historical sales data (usually weekly, by region or store cluster) and regresses it against every marketing and non-marketing driver you can measure.

Typical inputs:

• Media spend by channel and week
• Trade promotion depth and timing
• Price and competitor price
• Distribution (how many stores carry the product)
• Seasonality and weather
• Macroeconomic factors

The model outputs a contribution decomposition: what share of sales came from base demand versus each marketing lever, plus an ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → (return on investmentreturn on investmentReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition →) figure per channel.

Two concepts make MMM realistic:

Adstock (carryover): advertising effect decays over time rather than vanishing when the ad stops. A TV flight keeps working for several weeks.

Saturation (diminishing returns): the tenth dollar in a channel works less hard than the first. Beyond a point, more spend barely moves sales. This is the single most important thing MMM reveals, because it tells you where you are overspending.

Where AI changes the game

Classic MMM used simple linear regression and took analysts months to build. AI-driven MMM improves this in concrete ways:

Bayesian methods handle the "small data, many variables" problem FMCG always faces. You rarely have enough clean weeks of data to isolate every channel. Bayesian models let you bring in priors (informed starting assumptions, for example that TV ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → is usually positive) so the model stays stable instead of producing nonsense coefficients.

Automated feature handling: machine learning helps model non-linear saturation and interaction effects (for example, TV making a promotion work harder) without hand-tuning every curve.

Speed and refresh: what took a quarterly consulting engagement can now run monthly or on demand, so budgets adjust to what is actually happening.

Google's open-source library Meridian is a good example of where the field sits in 2026. It is a Bayesian MMM framework built for exactly this kind of analysis. You can explore it here: Meridian on GitHub.

Here is the conceptual shape of a modern MMM, simplified:

You do not need to code this. You need to understand that the model separates a base (sales you would get anyway) from incremental sales driven by spend, then divides incremental sales by cost to get ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → per channel.

MMM versus attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition →: use both

People confuse these. They answer different questions.

• MMM is top-down. It uses aggregate sales and works for all channels, including offline ones like TV and in-store displaysdisplaysThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.View full definition → that you cannot track click by click.
• Multi-touch attribution (MTA) is bottom-up. It follows individual user journeys (this person saw the ad, clicked, bought) and works only for trackable digital touchpoints.

Retail media sits awkwardly between them. It is digital and trackable, but it lives inside a retailer's walled garden, so you often get the retailer's own attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → numbers, which tend to flatter the retailer. Treat retailer-reported ROASROASReturn on Ad Spend (ROAS) measures the revenue generated for every unit of currency spent on advertising, calculated as revenue divided by ad cost.View full definition → (return on ad spendreturn on ad spendReturn on Ad Spend (ROAS) measures the revenue generated for every unit of currency spent on advertising, calculated as revenue divided by ad cost.View full definition →) with healthy skepticism and validate it against your MMM.

The 2026 best practice is a layered approach: MMM for the strategic budget split across all channels, attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → and experiments for tactical decisions within digital.

🎬 [VIDEO: "Marketing Mix Modeling Explained" — youtube.com — a clear walkthrough of MMM inputs, outputs, and how ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → curves guide budget decisions]

Solving the snack brand's problem

Back to our snack brand. Here is how MMM would resolve the promo-versus-ads dispute.

The model looks across many weeks and regions where the promo and the ads did *not* perfectly overlap. In regions where digital video ran without a promo, sales still rose modestly. In regions where the promo ran without ads, sales spiked hard but then dipped afterward (the pull-forward effect).

The decomposition might reveal:

• The promo drove most of the immediate volume, but a large chunk was existing buyers stockpiling, so the *incremental* profit was thin once you subtract the discount cost.
• The digital video drove a smaller but stickier lift, with better margin because there was no price discount attached.

Now finance has an answer. The 18 percent lift was mostly promo-driven, but the promo was less profitable than it looked. The recommendation: shift some trade dollars toward media, and make the next promo shallower but paired with a display, since MMM showed displaysdisplaysThe total number of times an ad or piece of content is displayed, regardless of clicks. Each display counts as one impression, even to the same person.View full definition → and promos amplify each other.

Reading the output without being fooled

MMM is powerful but easy to misuse. Watch for:

Correlation is not proof. If you always advertise heavily in December, the model may credit ads for what is really holiday seasonality. Good models control for this; bad ones do not.

Validate with experiments. The gold standard is a geo experiment: turn spend up in some regions, hold it flat in others, and measure the difference. This gives a real causal read that you can feed back to calibrate the MMM. Meta and Google both publish free guidance on incrementality testing worth reviewing.

Do not over-trust a single ROI number. Every estimate has a range of uncertainty. A Bayesian model gives you that range (a credible interval). If TV ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → could be anywhere from 0.8 to 2.5, that is a very different decision than a tight 1.9 to 2.1.

Turning insight into budget action

The whole point of MMM is reallocation. The saturation curves are your mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition →.

Practical workflow:

1. Find the flat parts. Any channel where the curve has gone flat is overspent. Move money out.

2. Find the steep parts. Channels still on the steep part of their curve can absorb more spend profitably.

3. Set guardrails. Do not slam an entire budget into one channel based on one model run. Reallocate in steps (for example, shift 10 to 15 percent), measure, and re-run.

4. Separate trade decisions. Trade spend has strategic strings attached (shelf space, retailer relationships) that pure ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → cannot capture. Use MMM to inform trade negotiations, not to blindly cut a retailer who also controls your distribution.

A realistic outcome for our snack brand: hold total budget flat, move a slice of trade money into retail media (which was still on a steep return curve), keep a leaner promo calendar, and commit to a geo test next quarter to confirm the media lift is real.

Key Takeaways

• MMM separates base sales from incremental sales, then divides incremental sales by spend to give ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → per channel. This is what settles "did the promo or the ads do it" arguments.
• Saturation curves are the money-maker. They show which channels are overspent (flat) and which can absorb more (steep). Reallocate toward the steep curves in steps.
• Use MMM and attribution together. MMM for the strategic all-channel split (including offline and trade), attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → and geo experiments for tactical digital decisions.
• AI, especially Bayesian MMM, makes this faster and more stable