The Attention Landscape · established evidence

What Media-Mix Modeling Actually Computes (and Why Most Marketers Never See It)

Last reviewed 2026-07-20. Written by Chandranshu Kumar, Founder, Raveneye Global. · 9 min read

Media-mix modeling, MMM for short, is a statistical method for estimating how much of your sales each marketing channel actually caused, and how that effect changes as spend on that channel rises. The two ideas that do the real work are diminishing returns, meaning each additional dollar on a channel usually buys a little less incremental result than the dollar before it, and carryover, meaning a channel's effect on sales does not end the moment you stop spending. Google formalized a Bayesian version of this in 2017, building on statistical foundations that go back to John Little's 1979 survey of aggregate advertising models. In plain terms, MMM is the formal answer to a question most marketers ask informally and answer by habit: where does the next dollar do the most good? A domain-level read of where an audience's attention actually sits is the applied, first-party version of the same question, answered before the dollar is spent rather than only after.

The question MMM is built to answer

Every marketing budget eventually faces the same decision: given a fixed amount of money and several channels competing for it, how should the money be split? Media-mix modeling exists to answer that question with statistics instead of habit. It uses historical data, sales or another outcome alongside spend by channel over time, to estimate a mathematical relationship between how much is spent on each channel and how much outcome results, controlling for other factors like seasonality, pricing, and promotions running at the same time.

The output is not a single number. It is a curve, one per channel, showing the estimated relationship between spend and effect across a range of possible spend levels. That curve is what makes MMM more than a spreadsheet of past ROAS figures: it lets you estimate what would happen at a spend level you have not tried yet.

Diminishing returns: why the tenth dollar is not worth what the first dollar was

The single most important shape in MMM is the diminishing-returns curve. John D.C. Little's 1979 survey of aggregate advertising models, published in Operations Research, cataloged this pattern decades before modern MMM software existed: sales response to advertising is typically concave or S-shaped, meaning the relationship flattens as spend rises. A channel that is highly efficient at low spend often becomes progressively less efficient as more money is poured into it, because the channel eventually runs out of new, reachable people at that price.

This is the mathematical basis for a claim every media planner has heard informally: you can spend too much on a channel, past the point where the next dollar earns its keep. MMM turns that intuition into an estimable number, the marginal return, which is the additional outcome one more dollar of spend on a specific channel is expected to produce at its current spend level.

Carryover: why a channel keeps working after you stop paying for it

The second core idea is carryover, sometimes called adstock. A channel's effect on a buyer does not necessarily end the moment the ad stops running. Someone who saw a video ad last week might still convert this week, influenced by an impression that already happened. MMM models this lag explicitly, so a channel with strong carryover, brand advertising is a common example, does not get unfairly penalized for producing results after the spend already occurred.

Modeling carryover correctly matters because ignoring it systematically undercounts the true value of channels that build awareness before they build conversions, and overcounts channels that look effective only because they capture demand another channel already created.

How the modern version actually works

Google researchers Jin, Wang, Sun, Chan, and Koehler formalized a Bayesian version of this modeling method in a 2017 paper, "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects." Bayesian, in this context, means the model produces a range of plausible answers with associated confidence, not a single point estimate presented as certain fact, which matters because marketing data is noisy and a model that hides its own uncertainty is easy to over-trust.

This 2017 paper is now the methodological basis for widely used open tools: Google's Meridian and Meta's open-source Robyn. Both let a marketer estimate, per channel, the shape of the diminishing-returns curve and the length of the carryover effect, then use those estimates to recommend how a fixed budget should be reallocated toward channels currently under-saturated relative to their marginal return.

Why most marketers never see this machinery directly

MMM requires a meaningful amount of historical spend and outcome data, competent statistical modeling, and time to build and validate, which is why it has historically been the domain of large advertisers with dedicated data-science teams and enterprise budgets. A small or mid-size business rarely has the transaction volume or the internal expertise to run a full Bayesian MMM in-house, and most agencies serving that segment do not offer it either.

What smaller businesses are typically sold instead is last-click or platform-reported ROAS, numbers that are directionally useful but do not account for diminishing returns or carryover, and that a well-known randomized field experiment at eBay found can meaningfully overstate the true incremental effect of paid search, particularly on branded terms. The gap between what MMM actually measures and what most marketers actually see is real, and it is a gap in access to the machinery, not a gap in the underlying economic logic.

The plain-language translation

Strip away the statistics and MMM is answering one question in a structured, evidence-based way: given everything we know about how each channel has responded to spend so far, where does the next dollar do the most good, right now, at current spend levels? That is the identical question a domain-level read of a market's attention terrain is built to answer, applied to a different input. Instead of estimating response curves from historical ad spend, a Corpus-style read observes where a specific audience's attention currently concentrates and how it is moving, before a dollar is committed, and treats that observation as the starting evidence for where investment is likely to earn the strongest marginal return.

Neither method removes judgment from budget allocation. Both replace habit, the "we've always run it this way" default, with a measured, falsifiable starting point.

The evidence

Key findings, with their sources

Sales response to advertising is typically concave or S-shaped, with marginal effectiveness that changes over time and can be estimated statistically rather than assumed.

established John D. C. Little, "Aggregate Advertising Models: The State of the Art," Operations Research, 27(4), pp. 629-667, 1979.
A Bayesian model for estimating each channel's diminishing-returns curve and carryover (adstock) effect, now the methodological basis for Google Meridian and the widely used open-source tool Meta Robyn.

established Y. Jin, Y. Wang, Y. Sun, D. Chan, J. Koehler, "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects," Google Inc., 2017.
A large-scale randomized field experiment at eBay found that non-experimental (observational, attribution-style) estimates of paid-search ROI are inflated relative to true causal lift, with brand-keyword ads in particular showing no measurable incremental short-term benefit once measured experimentally.

established T. Blake, C. Nosko, S. Tadelis, "Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment," Econometrica, 83(1), pp. 155-174, 2015.
Herbert Simon's attention-scarcity theorem, the theoretical root of treating attention allocation as a constrained-optimization problem that MMM formalizes for advertising specifically.

established Herbert A. Simon, "Designing Organizations for an Information-Rich World," in Computers, Communications, and the Public Interest, ed. M. Greenberger (Johns Hopkins Press, 1971).

Reference

Glossary

Media-mix modeling (MMM): A statistical method that estimates how much of an outcome, such as sales, each marketing channel caused, using historical spend and outcome data.
Diminishing returns: The pattern where each additional dollar spent on a channel produces a smaller incremental effect than the dollar before it, typically shown as a concave or S-shaped response curve.
Carryover (adstock): The lagged effect of advertising, where a channel continues to influence outcomes for a period after the spend itself has stopped.
Marginal return: The additional outcome expected from one more dollar of spend on a specific channel at its current spend level, the key number MMM is built to estimate.

Straight answers

Frequently asked questions

What is media-mix modeling in simple terms?

It is a statistical method that estimates how much each marketing channel actually contributes to sales, and how that contribution changes as you spend more or less on it, so a budget can be allocated based on evidence rather than habit.

Why does MMM matter if my business is too small to run one?

The underlying logic, diminishing returns and carryover, applies regardless of whether you can run the full statistical model. Understanding it helps you ask better questions of any vendor claiming a specific ROAS, and it is the same logic behind reading where a market's attention actually sits before committing budget.

Is media-mix modeling the same as attribution?

No. Attribution typically assigns credit for a specific conversion to a specific touchpoint, often the last click, which a well-known eBay field experiment found can overstate a channel's true incremental effect. MMM instead estimates the aggregate causal relationship between spend and outcome over time, and is generally considered more resistant to that overstatement.

What is Bayesian about Bayesian MMM?

It means the model produces a range of plausible estimates with stated uncertainty, rather than a single number presented as fact, which better reflects how noisy real marketing data actually is.

Provenance

Sources

Little, J. D. C., "Aggregate Advertising Models: The State of the Art," Operations Research, 27(4), 629-667, 1979 (established)
Jin, Y., Wang, Y., Sun, Y., Chan, D., Koehler, J., "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects," Google Inc., 2017 (established)
Blake, T., Nosko, C., Tadelis, S., "Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment," Econometrica, 83(1), 155-174, 2015 (established)
Simon, H. A., "Designing Organizations for an Information-Rich World," in Computers, Communications, and the Public Interest, ed. M. Greenberger (Johns Hopkins Press, 1971) (established)
Webster, J. G., The Marketplace of Attention: How Audiences Take Shape in a Digital Age (MIT Press, 2014) (established)direct.mit.edu

Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.