Marketing mix modeling glossary: 170+ MMM terms in plain English

Also: aak, ad attribution kit

Apple's successor framework to SKAdNetwork, introduced with iOS 17.4 in 2024, which keeps the aggregated, privacy-first postbacks and adds re-engagement measurement. It runs alongside SKAdNetwork rather than replacing it overnight.

Guide: MMM vs multi-touch attribution → See also: SKAdNetwork, ATT

Also: adstock transformation, adstock model

The transformation that spreads a week's advertising over the weeks that follow, because people remember ads after they've seen them. The simplest version, geometric adstock, carries a fixed share of last week's effect into this week. The idea is usually credited to Simon Broadbent's 1979 paper on how TV advertising works, and pretty much every MMM still uses some version of it.

Guide: Adstock → See also: Carryover effect, Half-life, Decay rate

Also: significance level, type i error rate, false positive rate

The false-positive rate you're willing to accept, conventionally 0.05. It's a choice, not a law: for a cheap, reversible budget decision, accepting a higher alpha in exchange for a faster test can be perfectly sensible.

Guide: Statistical power for marketing tests → See also: p-value, Statistical power

Also: app tracking transparency, ios 14.5

Apple's App Tracking Transparency framework, enforced from iOS 14.5 in April 2021, which requires apps to ask permission before tracking users across other companies' apps and sites. Most people said no, user-level attribution got much patchier, and interest in MMM jumped.

Guide: MMM vs multi-touch attribution → See also: SKAdNetwork, MTA

Also: lookback window, conversion window, click-through window

How long after a click or view a conversion can still be credited to it, such as Meta's default of 7-day click and 1-day view. Longer windows make channels look better without changing what they actually did.

Guide: MMM vs multi-touch attribution → See also: View-through conversion

Also: serial correlation

When the residual in one week predicts the residual in the next, meaning the model is systematically missing something that persists. It makes standard errors look smaller than they are, so channels look more significant than they should.

Guide: MMM diagnostics → See also: Durbin-Watson, Residual

Also: base sales, baseline sales, base

The sales you'd expect with no paid media, driven by brand strength, distribution, seasonality and habit. It's often 60 to 80 percent of the total in a mature brand, and a lot of MMM arguments are really arguments about what belongs in the baseline.

Guide: Baseline sales → See also: Incremental sales, Intercept

Also: blended return on ad spend

Another name for MER, sometimes restricted to paid media spend only. Watch the denominator: whether agency fees, creative and discounts are included changes the number a lot.

Guide: Marketing efficiency ratio → See also: MER

Also: brand lift, brand lift survey

A survey-based experiment that asks exposed and unexposed users about awareness, consideration or ad recall, and reports the difference. It measures attitudes rather than sales, so it's most useful when you have evidence that the attitude moves the sales.

Guide: Conversion and brand lift studies → See also: Conversion lift study

Also: break even roas, breakeven point

The ROAS at which spend exactly pays for itself in contribution, equal to one divided by your contribution margin. At a 40 percent margin it's 2.5; marginal ROAS above that means the next dollar is profitable.

Guide: Breakeven ROAS and contribution margin → See also: Contribution margin, mROAS

Also: budget optimizer, budget optimisation, media mix optimization

The tool that uses each channel's response curve to find the split of a budget that maximises the outcome, usually within constraints you set. Robyn, Meridian and PyMC-Marketing all include one; the answer is only as good as the curves, especially outside the spend range you've observed.

Guide: Marketing budget allocation → See also: Constraints, Equal marginal returns, Response curve

Also: mmm calibration, experiment calibration

Using experiment results to anchor an MMM's estimates, as priors on a channel's ROI (Meridian), as extra lift-test observations in the likelihood (PyMC-Marketing) or as an extra objective the model is scored on (Robyn). It's the most reliable way I know to stop a model's channel split drifting away from reality.

Guide: Calibrating MMM with experiments → See also: Prior, MAPE.LIFT, Lift

Also: cannibalization

When marketing shifts sales you'd have made anyway from one product, channel or period to another, so the total doesn't grow. Promotions that pull forward next month's sales and brand search ads that capture organic clicks are the usual suspects.

Guide: Incrementality → See also: Halo effect, Incrementality

Also: carryover, lagged effect

The part of an ad's effect that turns up in later weeks rather than in the week you paid for it. Adstock is how a model captures carryover; leave it out and you'll undercount slow-acting channels like TV and overcount whatever happens to be running when sales arrive.

Guide: Adstock → See also: Adstock

Also: causal impact, bsts

Google's R package, from Brodersen et al. (2015), that estimates the effect of an intervention with a Bayesian structural time-series model built from control series. It's the standard quick read on "what would sales have been without the campaign?", and tfp-causalimpact is its Python port.

Guide: Synthetic control method → See also: Synthetic control, Interrupted time series

PyMC Labs' Python library for quasi-experiments, including synthetic control, difference-in-differences and interrupted time series. Bayesian by default, with an OLS option.

Guide: Synthetic control method → See also: Synthetic control, Difference-in-differences

Also: data clean room, ads data hub, amazon marketing cloud

A secure environment where an advertiser's first-party data can be matched with a platform's or partner's data without either side seeing the other's user-level records, such as Google's Ads Data Hub or Amazon Marketing Cloud. Useful for reach, overlap and some lift analysis, but you only get aggregated outputs.

Guide: Attribution tools vs MMM → See also: MTA

Also: beta, parameter

The number the model multiplies an input by to get its effect on the outcome. In a linear MMM it's sales per unit of (transformed) spend; in a log-log model it's an elasticity. The coefficient alone doesn't tell you ROI until you account for the transformations.

Guide: Marketing econometrics → See also: OLS, Elasticity

Also: collider bias

A variable caused by two other variables, such as site visits driven by both ads and a news story. Controlling for a collider creates a spurious link between its causes, so it's a control you should leave out even though it correlates with everything.

Guide: Causal inference for marketers → See also: Mediator, Confounder

Also: ci, margin of error

A range produced by a procedure that, across many repeated experiments, would contain the true effect a stated share of the time (often 95 percent). In practice I'd report the interval before the point estimate, because the width is the honest part.

Guide: Incrementality testing → See also: Credible interval, p-value

Also: confounding variable, confounding

Something that drives both your spend and your sales, like seasonality or a product launch, making a channel look more effective than it is. If you can measure a confounder, put it in the model; if you can't, an experiment is the only clean answer.

Guide: Confounding variables in MMM → See also: Omitted variable bias, Control variable

Also: budget constraints, spend bounds

Limits you place on a budget allocator, such as "no channel moves more than 30 percent" or "TV stays above its minimum buy". They keep recommendations inside the range the model has seen and the business can execute.

Guide: Marketing budget allocation → See also: Budget allocator, Scenario

Also: channel contribution, incremental contribution

The amount of the outcome a model credits to one driver over a period, usually the coefficient times the transformed input, summed over the weeks. Divide a channel's contribution by its spend and you have its modelled ROAS.

Guide: What is marketing mix modeling? → See also: Decomposition, Baseline

Also: contribution margin ratio, unit margin

Revenue minus the variable costs of delivering it (cost of goods, shipping, payment fees), as an amount or a percentage of revenue. It's what turns a ROAS into profit: at a 40 percent margin, every dollar of incremental revenue is worth 40 cents to the business.

Guide: Breakeven ROAS and contribution margin → See also: Breakeven ROAS, ROI

Also: controls, non-media variable, covariate

A non-marketing input, such as price, distribution, competitor activity or weather, included so the model doesn't credit its effect to media. Good controls are confounders; bad controls are mediators or colliders, and adding them can make estimates worse.

Guide: Confounding variables in MMM → See also: Confounder, Mediator

Also: conversion lift, meta conversion lift, google conversion lift

A user-level experiment run by an ad platform, such as Meta or Google, that randomly withholds ads from a holdout and compares conversions. It's the cheapest experiment to run, but you're trusting the platform to grade its own homework, so read the methodology.

Guide: Conversion and brand lift studies → See also: Brand lift study, Ghost ads, Intent to treat

Also: counterfactual outcome

What would have happened in the absence of the thing you're measuring. You can never observe it directly, so every measurement method is really a way of estimating it: a control group, a synthetic region or a model's baseline.

Guide: Incrementality → See also: Incrementality, Synthetic control

Also: cost per acquisition, cost per action, cac

Spend divided by conversions (customers, sign-ups, orders). Like ROAS, it depends entirely on how conversions are counted, so an incremental CPA from an experiment is often several times the platform-reported one.

Guide: Breakeven ROAS and contribution margin → See also: LTV, Breakeven ROAS

Also: cost per mille, cost per thousand impressions

Cost per thousand impressions. When CPMs swing (Q4 auctions, for example), the same spend buys very different exposure, which is why impressions can be a better MMM input than spend for some channels.

Guide: Reach and frequency → See also: GRP

Also: bayesian interval, hdi, highest density interval

A range that contains a parameter with a stated probability, given the model and priors, such as a 90 percent credible interval for a channel's ROAS. Unlike a confidence interval, it means what most people think a confidence interval means.

Guide: Bayesian MMM → See also: Posterior, Confidence interval

Also: time series cross validation, rolling origin

Repeatedly fitting the model on part of the data and testing on another part, then averaging the errors. For time series you must test on later weeks than you train on (rolling or expanding windows), or the model gets to peek at the future.

Guide: MMM validation → See also: Holdout period, Lambda

Also: causal graph, directed acyclic graph, causal diagram

A diagram of arrows showing which variables you believe cause which. Drawing one before you choose controls is the fastest way I know to spot confounders, mediators and colliders, and to explain your choices to someone else.

Guide: Causal inference for marketers → See also: Confounder, Mediator, Collider

Also: dda, algorithmic attribution, google data-driven attribution

Attribution where an algorithm assigns credit by comparing converting and non-converting paths, rather than following a fixed rule. It has been the default for new conversion actions in Google Ads since 2021 (and, since 2023, the only multi-touch option there), and it's still credit-splitting on tracked paths, not an incrementality measurement.

Guide: MMM vs multi-touch attribution → See also: Shapley value, MTA

Also: retention rate, theta, alpha adstock

The adstock parameter that sets how much of an ad's effect survives from one week to the next. Robyn calls it theta; a value near zero means the effect is gone almost immediately, while a value near one means it lingers for months. It's usually estimated by the model rather than set by hand, within bounds you choose.

Guide: Adstock → See also: Half-life, Geometric adstock

Also: mmm report, recommendation memo

A one or two page write-up that states the recommendation, the evidence, the uncertainty and what would change your mind. It's what decision-makers actually read, so it deserves as much care as the model.

Guide: Communicating MMM results → See also: Scenario, Sensitivity analysis

Also: decomp rssd, decomposition rssd, rssd

Robyn's decomposition root sum of squared distance: how far each paid channel's share of effect is from its share of spend. Lower values favour models whose credit split isn't wildly different from the spend split, which guards against implausible answers but also biases towards the status quo.

Guide: Robyn → See also: NRMSE, Pareto front

Also: decomp, sales decomposition, due-to

Splitting the outcome into the contribution of each driver (baseline, each channel, price, promotions, seasonality) so the pieces add back up to the total. It's the stacked area chart in every MMM report, and the first thing I check against what the business knows happened.

Guide: What is marketing mix modeling? → See also: Contribution, Baseline

Also: lagged adstock, peak delay

An adstock where the effect peaks some weeks after the spend rather than immediately, popularised by Google's Jin et al. (2017) paper. It suits considered purchases where there's a gap between seeing the ad and buying, such as cars, holidays or financial products.

Guide: Jin et al. (2017) → See also: Adstock

Also: did, diff in diff, diff-in-diff

Comparing the change in the treated group before and after an intervention with the change in a control group over the same period. Subtracting the control's change removes anything that hit both groups, like seasonality, leaving the effect of the intervention.

Guide: Difference-in-differences → See also: Parallel trends, Geo test

Each extra dollar in a channel producing less than the dollar before it. It's the everyday name for saturation, and the single most useful idea in budget planning: the question is never whether a channel works, but whether the next dollar works.

Guide: Saturation curves → See also: Saturation

Also: divergent transitions

Warnings from the NUTS sampler that it hit a region of the posterior it couldn't explore properly. A few may be tolerable; many mean the model is badly specified or the priors fight the data, and the estimates may be biased.

Guide: Bayesian MMM → See also: NUTS, R-hat

Also: designated market area, nielsen dma

Designated Market Area, Nielsen's division of the US into 210 television markets. DMAs are the usual unit for US geo tests and geo-level MMMs because TV is bought by DMA and most platforms can target them.

Guide: Geo lift testing → See also: Geo test, Geo-level model

Also: donor regions, control units

The untreated regions a synthetic control is allowed to draw from. Every donor must be free of the treatment and of its own shocks during the test, or the counterfactual inherits them.

Guide: Synthetic control method → See also: Synthetic control

Also: indicator variable, binary variable, flag

A column that's 1 when something is happening (a promotion, a holiday, a stock-out) and 0 otherwise. Its coefficient tells you the average lift during those weeks, which is the cheapest way to stop a one-off event from being credited to media.

Guide: Marketing econometrics → See also: Holiday effects, Control variable

Also: durbin watson statistic, dw statistic

A test statistic for first-order autocorrelation in residuals, running from 0 to 4. Values near 2 suggest no autocorrelation; values well below 2 (common in MMMs) mean positive autocorrelation and a missing slow-moving driver.

Guide: MMM diagnostics → See also: Autocorrelation

Also: half-saturation point, half saturation, gamma, ec

The spend (or adstocked spend) at which a Hill curve reaches half of its maximum effect. A channel spending far below its EC50 has room to grow; one spending well above it is deep into diminishing returns. Libraries name it differently, so check which scale it's on (raw spend or a normalised version) before you compare numbers.

Guide: Saturation curves → See also: Hill function

Also: marketing econometrics, econometric modeling

The application of statistics to economic data, and the discipline MMM grew out of. In the UK and Europe, "econometrics" is still what many agencies call their marketing mix models.

Guide: Marketing econometrics → See also: MMM, OLS

Also: three hit theory, effective reach

The idea that an ad needs to be seen a minimum number of times to work, often cited as three after Herbert Krugman's 1972 theory. Later research on recency suggested a single well-timed exposure does most of the work, so treat any fixed number with suspicion.

Guide: Reach and frequency → See also: Frequency

Also: marketing elasticity, advertising elasticity

The percentage change in the outcome for a 1 percent change in an input. An advertising elasticity of 0.05 means 10 percent more spend gives about 0.5 percent more sales, which is a useful sanity check because advertising elasticities in the literature are usually small.

Guide: Price elasticity → See also: Log-log model, Price elasticity

Also: endogenous variable

When an input is correlated with the model's error term, so its coefficient is biased. In marketing the classic case is spend set in response to expected demand: search budgets rise when people are already searching, and the model can't tell the difference.

Guide: Confounding variables in MMM → See also: Omitted variable bias, Reverse causality

Also: equalise marginal roas, equimarginal principle

The rule that a budget is optimally split when the next dollar earns the same in every channel. If search's marginal ROAS is 3 and video's is 1.5, moving money from video to search improves the total until the two meet.

Guide: Marketing budget allocation → See also: mROAS, Budget allocator

Also: excess share of voice

Share of voice minus share of market. In the IPA data analysed by Les Binet and Peter Field, brands with positive ESOV tended to grow share over time. Their 2019 B2B report puts it at about 0.6 points of share growth a year per ten points of ESOV for consumer brands and about 0.7 for B2B, with a wide range by category, so treat it as a sense check on long-term budgets rather than a forecast.

Guide: The Long and the Short of It → See also: Share of voice, Share of market

Also: event study design

A version of difference-in-differences that estimates the effect separately for each week before and after the intervention. The pre-intervention estimates should sit near zero, which makes it a visual check on parallel trends.

Guide: Difference-in-differences → See also: Difference-in-differences, Parallel trends

Also: first touch attribution

Giving all the credit to the first tracked touchpoint. Google retired it, along with linear, time-decay and position-based models, from Google Ads and Google Analytics in 2023.

Guide: MMM vs multi-touch attribution → See also: Last-click attribution

Also: fourier series, sine cosine seasonality

Pairs of sine and cosine waves used to model smooth seasonality without a separate dummy for every week of the year. A handful of pairs captures most annual patterns with far fewer parameters than 52 dummies.

Guide: Baseline sales → See also: Seasonality

Also: average frequency

The average number of times each reached person saw your ad. On a fixed budget, more frequency means less reach, and past a few exposures extra frequency mostly annoys people.

Guide: Reach and frequency → See also: Reach, Effective frequency, Frequency cap

Also: frequency capping

A limit on how many times one person can be shown your ad in a period, set in the ad platform. It's the practical lever for trading frequency back into reach.

Guide: Reach and frequency → See also: Frequency

Also: geo experiment, geo lift, geo testing, regional test

An experiment that turns marketing up or down in some regions and compares them with regions left alone. It works for channels where you can't randomise individuals, like TV, out-of-home or a whole platform, and it's the kind of experiment MMMs most naturally calibrate against.

Guide: Geo lift testing → See also: Matched markets, Synthetic control, DMA

Also: geo mmm, regional mmm, geo-level mmm

An MMM fitted on data by region (state, DMA, country) rather than one national series. It gives the model many more observations and more variation in spend, which is why Google's Sun et al. (2017) paper and Meridian both push it when you have regional data.

Guide: Geo-level hierarchical MMM → See also: Hierarchical model, DMA

Also: geolift r package, meta geolift

Meta's open-source R package for designing and analysing geo tests with augmented synthetic control. It helps choose test markets, run power analysis and read out the lift.

Guide: Geo lift testing → See also: Synthetic control, Geo test

Adstock where each week keeps a constant fraction of the previous week's effect, so the effect decays smoothly and peaks in the week of spend. It has one parameter, which makes it easy to estimate and explain, and it's the default in most libraries.

Guide: Adstock → See also: Weibull adstock, Delayed adstock

Also: ghost ad, predicted ghost ads

An experimental method from Johnson, Lewis and Nubbemeyer (2017) where the ad system records when it would have shown your ad to someone in the control group, without showing it. That lets you compare people who actually saw the ad with their would-have-seen counterparts, at no cost for control impressions.

Guide: Conversion and brand lift studies → See also: PSA test, Intent to treat

Also: turn off test, blackout test, pause test

Switching a channel off, in some regions or for a period, to see how much you lose. It's the bluntest incrementality test and often the most persuasive, as long as you keep a control and don't turn it off during a seasonal dip.

Guide: Incrementality testing → See also: Geo test, Incrementality test

Also: gross rating point, grps, trp, target rating point

Gross rating points: reach as a percentage of the target population multiplied by average frequency, so 50 percent reach at a frequency of 4 is 200 GRPs. TV is still bought in GRPs (or TRPs for a specific target audience), and they make a better MMM input than spend when prices vary.

Guide: Reach and frequency → See also: Reach, Frequency, CPM

Also: advertising half-life, ad half life

The number of weeks it takes for half of an ad's effect to wear off. It's an easier way to talk about decay with non-analysts than a retention rate: a weekly retention of 0.5 is a one-week half-life, and 0.8 is a little over three weeks.

Guide: Adstock → See also: Decay rate

Also: halo, brand halo, cross-channel halo

When marketing for one product, brand or channel lifts sales of another, such as TV that drives branded search or a hero product's ads that lift the whole range. A model that only links each channel to its own product will undercount it.

Guide: Media mix modeling → See also: Cannibalisation, Mediator

Also: heteroscedasticity, non-constant variance

When the size of the residuals changes with the level of the outcome, for instance bigger errors in peak season. It doesn't bias the coefficients, but it makes the usual standard errors wrong; logging the outcome often fixes it.

Guide: MMM diagnostics → See also: Residual

Also: multilevel model, hierarchical bayesian model

A model where parameters for groups (regions, brands, products) are drawn from a shared distribution, so each group gets its own estimate informed by the others. It's how Meridian handles geo data and how you get sensible estimates for small regions.

Guide: Geo-level hierarchical MMM → See also: Partial pooling, Geo-level model

Also: hill curve, hill saturation, hill equation

A saturation curve borrowed from pharmacology that most MMM libraries use, including Robyn and Meridian. It has two parameters: the half-saturation point (EC50) and a shape parameter that decides whether the curve is simply concave or S-shaped, with a slow start before it takes off.

Guide: Saturation curves → See also: EC50, Saturation

Also: control group, holdout

The people or regions deliberately not exposed to the marketing in an experiment, so you can see what happens without it. The holdout has to be chosen randomly (or matched carefully), or you're back to comparing apples with oranges.

Guide: Incrementality testing → See also: Incrementality test, Randomised controlled trial

Also: holdout validation, test set, out of sample period

The most recent weeks of data kept back from fitting and used to check how well the model predicts. It's the closest thing an MMM has to a reality check, though predicting well doesn't prove the channel split is right.

Guide: MMM validation → See also: Out-of-sample, Overfitting

Also: holidays, event effects

Spikes or dips around specific dates such as Black Friday, Easter or Christmas, which move about from year to year and so aren't captured by smooth seasonality. Robyn uses Prophet's country holiday calendars for these; elsewhere you add them as dummy variables.

Guide: Baseline sales → See also: Dummy variable, Seasonality

Also: hyperparameters

A setting chosen outside the main regression fit, such as adstock decay, saturation shape or the ridge penalty. Robyn searches thousands of hyperparameter combinations; Bayesian libraries estimate them as parameters with priors instead.

Guide: Robyn → See also: Decay rate, EC50, Nevergrad

Also: incremental revenue, incremental conversions

Sales above the baseline that the model credits to marketing activity. These are the sales that wouldn't have happened without the activity, which is what you're paying for.

Guide: Baseline sales → See also: Baseline, Incrementality

Also: incremental, incremental impact

The outcome that happened because of a marketing activity, compared with what would have happened without it. It's the only definition of "working" that matters for budget decisions, and it's usually smaller than what ad platforms report.

Guide: Incrementality → See also: Counterfactual, Lift, iROAS

Also: incrementality testing, lift test, holdout test

An experiment that withholds marketing from a randomly chosen group of people or regions to measure what the marketing caused. It's the gold standard for incrementality, with the trade-off that you give up some sales in the control group.

Guide: Incrementality testing → See also: Holdout group, Geo test, Conversion lift study

Also: itt, intention to treat

Comparing everyone assigned to the test group with everyone assigned to control, whether or not they actually saw an ad. It's unbiased but diluted, because many people in the test group never got an impression.

Guide: Conversion and brand lift studies → See also: Treatment on the treated

Also: constant

The regression's predicted outcome when every input is zero. In an MMM it's the floor of the baseline; when it comes out implausibly low or negative, the model is usually crediting media with sales that belong to trend or seasonality.

Guide: Baseline sales → See also: Baseline

Also: its, pre-post analysis, before and after

Estimating an effect by projecting the pre-intervention trend forward and comparing it with what actually happened. Without a control group it's vulnerable to anything else that changed at the same time, so treat it as weak evidence.

Guide: Synthetic control method → See also: CausalImpact, Counterfactual

Also: incremental roas, incremental return on ad spend

Incremental revenue divided by spend, where incremental means measured against a counterfactual, usually from an experiment or an MMM. It's the ROAS that tells you whether the spend paid for itself.

Guide: Incrementality → See also: ROAS, Incrementality

Also: penalty, regularisation strength

The dial that sets how hard ridge or lasso regression shrinks coefficients. Zero gives you ordinary least squares; very large values shrink everything towards nothing. It's usually chosen by cross-validation rather than by hand.

Guide: Ridge regression for MMM → See also: Ridge regression

Also: lasso regression, l1 regularisation

Regression with a penalty on the absolute size of coefficients, which can shrink some all the way to zero and drop them from the model. That's handy for variable selection but risky in MMM, where a small channel being zeroed out is a business claim, not just a statistical one.

Guide: Ridge regression for MMM → See also: Ridge regression, Regularisation

Also: last click, last touch attribution, last non-direct click

Giving all the credit for a conversion to the final click before it. It flatters channels close to purchase, like branded search and retargeting, and starves the ones that created demand in the first place.

Guide: MMM vs multi-touch attribution → See also: First-click attribution, MTA

Also: incremental lift, uplift

The difference in outcome between the group that saw the marketing and the group that didn't, as an absolute number or a percentage. A 5 percent lift on a 2 percent conversion rate means 2.1 percent in test versus 2.0 percent in control, which is why lift tests need large samples.

Guide: Incrementality testing → See also: Incrementality, Conversion lift study

Also: lightweight mmm, lmmm

Google's earlier Bayesian MMM library in Python, built on NumPyro. It was archived on 19 January 2026, with its README pointing users to Meridian; the last release was 0.1.9 in May 2023.

Guide: LightweightMMM → See also: Meridian

Also: likelihood function

How probable the observed data are under a given set of parameter values. In Bayesian MMM, the likelihood is the part that lets the data speak, and the posterior is proportional to likelihood times prior.

Guide: Bayesian MMM → See also: Prior, Posterior

Also: linear model

Splitting credit equally across every tracked touchpoint. It's fair-sounding but arbitrary, because there's no reason the third ad view deserves exactly as much as the first click.

Guide: MMM vs multi-touch attribution → See also: MTA, Time-decay attribution

Also: log-lin model, semi-log model

A regression where the outcome is logged but the inputs aren't, so a coefficient is roughly the percentage change in sales per unit of input. It makes effects multiplicative, so a channel's lift is larger in weeks when the baseline is already high.

Guide: Marketing econometrics → See also: Log-log model

Also: double log model, log log regression

A regression where both the outcome and the inputs are logged, so each coefficient reads directly as an elasticity. The log also builds in diminishing returns, which is why it was the workhorse of older econometric MMMs.

Guide: Price elasticity → See also: Elasticity, Log-linear model

A one-parameter saturation curve used by PyMC-Marketing, which rises steeply from zero and flattens towards a ceiling. It's simpler than a Hill curve but can't produce the S-shape, which is often fine because few channels have a genuine slow start in weekly data.

Guide: PyMC-Marketing → See also: Hill function

Also: lifetime value, customer lifetime value, clv, cltv

The total contribution a customer is expected to bring over their relationship with you. Subscription and repeat-purchase businesses should judge acquisition against LTV rather than first-order revenue, which an MMM on weekly revenue won't do for you automatically.

Guide: Breakeven ROAS and contribution margin → See also: CPA, Contribution margin

Also: mean absolute percentage error

The average absolute error as a percentage of actual sales. It's easy to explain ("the model is off by 6 percent on average") but it blows up when actuals are near zero, and a low MAPE says nothing about whether the channel credit is right.

Guide: MMM validation → See also: NRMSE, R²

Also: mape lift

Robyn's calibration objective: the error between the model's estimated effect during an experiment window and the lift the experiment measured. When you add experiments it becomes a third objective alongside NRMSE and DECOMP.RSSD.

Guide: Calibrating MMM with experiments → See also: Calibration, DECOMP.RSSD

Also: markov attribution, removal effect

Attribution that models journeys as transitions between channels and credits each channel by its "removal effect", meaning how many conversions would be lost if it disappeared from the paths. It's clever, but it still only sees the tracked paths.

Guide: MMM vs multi-touch attribution → See also: Shapley value

Also: matched market test, market matching

A geo test design that pairs regions with similar historical sales and assigns one of each pair to treatment. It's simple to explain, though with only a handful of regions a single unusual market can swing the result.

Guide: Geo lift testing → See also: Geo test, Synthetic control

Also: markov chain monte carlo

Markov chain Monte Carlo, the family of algorithms that draws thousands of samples from a posterior distribution when it can't be calculated directly. Meridian and PyMC-Marketing both rely on it, which is why Bayesian MMMs take minutes to hours to fit rather than seconds.

Guide: Bayesian MMM → See also: NUTS, R-hat

Also: minimum detectable effect, minimum detectable lift

The smallest true effect a test can reliably detect at your chosen power and significance level. Work it out before the test; if the MDE is a 20 percent lift and you expect 3 percent, the test isn't worth running as designed.

Guide: Statistical power for marketing tests → See also: Statistical power, Sample size

Also: mmm brief, measurement framework plan

A short document that starts from the business decision, then names the outcome KPI, the methods, the data and the timeline for answering it. Writing it before any modeling is the cheapest way to stop an MMM project from answering the wrong question.

Guide: MMM measurement plan → See also: Triangulation, Calibration

Also: media mix model, media mix

A near-synonym for marketing mix modeling, sometimes used for models that focus on paid media rather than the full mix of price, promotion and distribution. Both are abbreviated MMM, and in practice the terms are used interchangeably.

Guide: Media mix modeling → See also: MMM

Also: mediating variable

A variable that sits on the path between cause and outcome, like branded search sitting between a TV campaign and sales. Controlling for a mediator strips out part of the effect you're trying to measure, which is why branded search spend in an MMM often steals credit from TV.

Guide: Causal inference for marketers → See also: Control variable, Collider

Also: marketing efficiency ratio, blended roas

Total revenue divided by total marketing spend, with no attribution at all. It's popular with ecommerce brands because it can't be gamed by tracking, but it mixes baseline sales with incremental ones and hides which channel is doing the work.

Guide: Marketing efficiency ratio → See also: Blended ROAS, ROAS

Also: google meridian, meridian mmm, google mmm

Google's open-source Bayesian MMM in Python, in limited availability from March 2024 and generally available from January 2025 (the latest release is 2.1.0, from September 2026). It supports national or geo-level hierarchical models, ROI priors from experiments, and reach and frequency inputs, and it replaced LightweightMMM as Google's recommended library.

Guide: Google Meridian → See also: Hierarchical model, Calibration, LightweightMMM

Also: geox

Google's Python companion to Meridian for designing and analysing geo experiments, including multi-cell designs. Its results can be handed to Meridian as ROI priors, which closes the loop between testing and modeling.

Guide: Calibrating MMM with experiments → See also: Meridian, Geo test

Also: michaelis menten saturation

A concave saturation curve from enzyme chemistry, equivalent to a Hill curve with its shape parameter fixed at one. It has a maximum effect and a half-saturation point, and nothing else, which makes it easy to explain.

Guide: Saturation curves → See also: Hill function

Also: marketing mix modeling, marketing mix modelling, marketing mix model, mmm marketing

Marketing mix modeling, a regression on aggregated (usually weekly) data that estimates how much each marketing activity and other driver contributed to sales. It doesn't need user tracking, which is why it's back in fashion, and it's only as good as the data and causal thinking that go into it.

Guide: What is marketing mix modeling? → See also: Media mix modeling, Decomposition, Adstock

Also: mobile measurement partner

A mobile measurement partner, such as AppsFlyer or Adjust, that sits between an app and its ad networks to attribute installs and in-app events. Since ATT most of their work involves SKAdNetwork data, modelled conversions and, increasingly, incrementality and MMM features.

Guide: Attribution tools vs MMM → See also: SKAdNetwork

Also: refresh, model update

Re-fitting an MMM with new weeks of data, usually monthly or quarterly. A good refresh keeps the story stable unless something real changed; if a channel's ROAS swings by half every refresh, the model wasn't identifying it in the first place.

Guide: MMM scenario planning → See also: Structural break

Also: marginal roas, marginal return on ad spend

The revenue from the next dollar of spend, read from the slope of the response curve at your current spend. Because of diminishing returns it's lower than average ROAS, and it's the number that should drive budget changes.

Guide: Marginal ROAS → See also: ROAS, Response curve, Diminishing returns

Also: multi-touch attribution, multi touch attribution, user-level attribution

Multi-touch attribution splits credit for each conversion across the tracked touchpoints in a user's journey. It's granular and fast, but it only sees trackable digital clicks and views, it assigns credit rather than measuring incrementality, and privacy changes have made the journeys much less complete.

Guide: MMM vs multi-touch attribution → See also: Last-click attribution, Data-driven attribution, Shapley value

Also: collinearity, correlated channels

When two or more inputs move together so closely that the model can't separate their effects, such as search and social that always go up in the same weeks. The total is estimated fine, but the split between them is unstable, and small data changes can swing it wildly.

Guide: Multicollinearity in MMM → See also: VIF, Ridge regression

Meta's open-source Python library for gradient-free optimisation. Robyn calls it (through R) to search adstock, saturation and ridge hyperparameters across thousands of trials.

Guide: Robyn → See also: Robyn, Hyperparameter

Also: normalized root mean square error, normalised rmse

Root mean squared error divided by the range (or sometimes the mean) of the actuals, so it's comparable across datasets. Robyn uses it as its prediction-error objective alongside DECOMP.RSSD.

Guide: MMM validation → See also: DECOMP.RSSD, MAPE

Also: no-u-turn sampler

The No-U-Turn Sampler, the default MCMC algorithm in Stan, PyMC and Meridian. It explores the posterior efficiently by simulating a particle rolling around the probability surface, and tunes itself so you don't have to set step sizes by hand.

Guide: Bayesian MMM → See also: MCMC, Divergences

Also: ordinary least squares, linear regression

Ordinary least squares, the standard regression that picks coefficients to minimise the sum of squared errors. The original marketing mix models were OLS on transformed media, and it's still the right place to start before anything fancier.

Guide: Marketing econometrics → See also: Coefficient, Residual

Also: ovb, missing variable

The error you get when a variable that affects sales, and is correlated with an included input, is left out of the model. Its effect leaks into whatever it's correlated with, so a missing promotion calendar can make your biggest promo-week channel look like a genius.

Guide: Confounding variables in MMM → See also: Confounder, Endogeneity

Also: uber orbit, orbit ktr, ktr

Uber's open-source Bayesian time-series library in Python. Its Kernel-based Time-varying Regression (KTR) model lets channel coefficients drift over time, and Uber has written about using it for MMM, but it has no built-in adstock, saturation or budget optimiser.

Guide: Open-source MMM libraries → See also: Time-varying coefficient

Also: out of sample accuracy, oos

Performance measured on data the model didn't see during fitting. In-sample fit always improves as you add variables; out-of-sample error is the number that tells you whether those variables helped.

Guide: MMM validation → See also: Holdout period

Also: overfit

When a model fits the noise in its training data rather than the pattern, so it looks brilliant in-sample and falls apart on new weeks. With two or three years of weekly data and a dozen channels, an MMM has far fewer observations than it would like, so overfitting is the default risk.

Guide: MMM validation → See also: Holdout period, Cross-validation

Also: p value, statistical significance

The probability of seeing a result at least this extreme if there were really no effect. It isn't the probability that the effect is real, and a p-value just above 0.05 isn't evidence that the campaign did nothing.

Guide: Statistical power for marketing tests → See also: Alpha, Confidence interval

Also: parallel trends assumption, common trends

The assumption behind difference-in-differences: without the intervention, treatment and control would have moved in step. You can't test it for the test period, but you can look at the pre-period, and if the lines already diverge, the estimate is suspect.

Guide: Difference-in-differences → See also: Difference-in-differences

Also: pareto frontier, pareto optimal models

The set of models where you can't improve one objective without making another worse, for example lower NRMSE only at the cost of higher DECOMP.RSSD. Robyn gives you the models on its Pareto front and leaves the final pick to the analyst.

Guide: Robyn → See also: DECOMP.RSSD, NRMSE, Nevergrad

Also: shrinkage across groups, borrowing strength

The compromise between fitting one estimate for every region (complete pooling) and a fully separate estimate for each (no pooling). Regions with lots of data keep their own estimate; regions with little data get pulled towards the average.

Guide: Geo-level hierarchical MMM → See also: Hierarchical model

Also: platform attribution, in-platform roas, self-reported conversions

The conversions an ad platform credits to itself, using its own attribution rules. Every platform counts the same customer, so platform totals routinely add up to more than your actual sales.

Guide: Incrementality → See also: Incrementality, View-through conversion

Also: u-shaped attribution, position based

A rule-based model that gives most of the credit to the first and last touchpoints (often 40 percent each) and splits the rest across the middle. Also called U-shaped attribution.

Guide: MMM vs multi-touch attribution → See also: First-click attribution, Last-click attribution

Also: posterior distribution

The updated distribution of a parameter after combining the prior with the data. With lots of informative data the posterior is driven by the data; with weak data it looks a lot like the prior, which is worth checking before you present it.

Guide: Bayesian MMM → See also: Prior, Credible interval

Also: pre-test period, training period

The weeks before an experiment starts, used to check that test and control regions behave alike and to fit a synthetic control. A longer, stable pre-period gives you a much better counterfactual.

Guide: Geo lift testing → See also: Parallel trends, Synthetic control

Also: price elasticity of demand, ped

The percentage change in volume sold for a 1 percent change in price, usually negative. An elasticity of -2 means a 10 percent price cut lifts volume about 20 percent. Leaving price out of an MMM is one of the quickest ways to bias the media estimates.

Guide: Price elasticity → See also: Elasticity, Log-log model

Also: prior distribution, mmm priors

In a Bayesian model, the range of values you believe a parameter could take before seeing the data, written as a probability distribution. A good prior encodes things you genuinely know (ROAS isn't negative, adstock doesn't last a year); a careless one quietly decides the answer.

Guide: Bayesian MMM → See also: Posterior, Likelihood

Also: facebook prophet, meta prophet

Meta's open-source forecasting library, which decomposes a time series into trend, seasonality and holiday effects. Robyn uses it to build those baseline components before fitting the media model.

Guide: Robyn → See also: Trend, Seasonality, Holiday effects

Also: public service announcement test, placebo ad

An older lift-test design that shows the control group an unrelated charity ad instead of yours, so both groups are equally "reached". It works, but you pay for the placebo impressions, and ad delivery algorithms can target the two ads differently.

Guide: Conversion and brand lift studies → See also: Ghost ads

Also: pymc3

The open-source probabilistic programming library in Python that PyMC-Marketing is built on. You describe a model with priors and a likelihood, and it samples the posterior for you.

Guide: PyMC-Marketing → See also: PyMC-Marketing, NUTS

Also: pymc marketing, pymc mmm

PyMC Labs' open-source Bayesian marketing library in Python, with an MMM, customer lifetime value models and more (version 1.0, in August 2026, made the multidimensional MMM its only MMM class). It's the most flexible of the main libraries, with many adstock and saturation options and lift-test calibration, which also means more choices you have to justify.

Guide: PyMC-Marketing → See also: PyMC, Logistic saturation

Also: rhat, gelman-rubin, convergence diagnostic

A convergence check that compares several MCMC chains: if they agree, R-hat is close to 1.0. Values above about 1.01 to 1.05 (depending on whose guidance you follow) mean the chains haven't settled and the results shouldn't be trusted yet.

Guide: Bayesian MMM → See also: MCMC, Divergences

Also: r squared, r-squared, coefficient of determination, adjusted r squared

The share of variation in the outcome that the model explains, from 0 to 1. MMMs with trend and seasonality routinely score above 0.9, so a high R² is table stakes rather than proof; adjusted R² penalises extra variables, and out-of-sample R² is better still.

Guide: MMM validation → See also: MAPE, Overfitting

Also: rct, randomized controlled trial, a/b test

An experiment where units are randomly assigned to treatment or control, so any difference in outcome can be attributed to the treatment. Randomisation is what removes selection bias, which is why RCTs beat observational methods when you can run them.

Guide: Incrementality testing → See also: Holdout group, Selection bias

Also: net reach, unique reach

The number or share of people who saw your ad at least once in a period. Reach grows with diminishing returns because the easy-to-reach people are reached first, which is one mechanism behind saturation.

Guide: Reach and frequency → See also: Frequency, GRP

Also: regularization, shrinkage, penalised regression

Any technique that pulls estimates towards a default (usually zero) to stop the model overfitting noise. Ridge and lasso do it with a penalty; Bayesian models do it with priors, which is the same idea with a more explicit story.

Guide: Ridge regression for MMM → See also: Ridge regression, Prior

Also: residuals, error term

The gap between actual and predicted outcome in a given week. Plotting residuals over time is the best diagnostic I know: a run of large positive residuals usually points to something you've left out, like a promotion or a stock-out.

Guide: MMM diagnostics → See also: Autocorrelation, Structural break

Also: marketing response curve, spend response curve

A chart of spend against modelled outcome for one channel, combining its saturation curve and coefficient. It's the picture you need to read average and marginal returns, and it's what a budget allocator is really optimising over.

Guide: Saturation curves → See also: Saturation

Also: simultaneity

When the outcome drives the input rather than the other way around, for instance budgets set as a percentage of last month's sales, or search ads that only serve when people already want to buy. It makes channels look more effective than they are.

Guide: Confounding variables in MMM → See also: Endogeneity

Also: ridge, l2 regularisation, tikhonov

Regression that adds a penalty on the squared size of the coefficients, shrinking them towards zero so correlated channels don't get wild, offsetting estimates. It's what Robyn fits under the hood. It trades a little bias for a lot less variance.

Guide: Ridge regression for MMM → See also: Regularisation, Lasso, Lambda

Also: return on ad spend, average roas

Revenue divided by ad spend. On its own it's ambiguous: a platform-reported ROAS counts attributed revenue, an MMM ROAS counts modelled incremental revenue, and neither is profit, so always ask which ROAS someone means.

Guide: Marginal ROAS → See also: iROAS, mROAS, MER

Also: meta robyn, facebook robyn, robyn mmm

Meta Marketing Science's open-source MMM package for R. It fits ridge regression and uses Nevergrad to search adstock and saturation hyperparameters, then gives you a Pareto front of candidate models to choose from. The latest CRAN release is 3.12.1, from July 2025; development has slowed since then, and a Python port (robynpy) is in beta.

Guide: Robyn → See also: Ridge regression, Nevergrad, Pareto front, DECOMP.RSSD

Also: return on investment, marketing roi, mroi

Strictly, profit from the investment divided by the investment, so it can be negative. In MMM reports "ROI" is often used loosely for revenue per dollar (which is ROAS), so check the definition before comparing numbers; Meridian, for instance, reports ROI as incremental revenue per dollar.

Guide: Breakeven ROAS and contribution margin → See also: ROAS, Contribution margin

Also: sample size calculator

How many users, conversions, regions or weeks a test needs to reach its target power. For ad effects the required numbers are often startlingly large, which Lewis and Rao's work on the economics of ad measurement made very clear.

Guide: Statistical power for marketing tests → See also: MDE, Statistical power

Also: media saturation, saturation effect

The point where more spend in a channel buys less extra outcome, because you've already reached most of the people likely to respond. Models capture it with a curve that flattens as spend rises, and it's the reason the optimal budget is never "put everything in the best channel".

Guide: Saturation curves → See also: Diminishing returns, Hill function, Response curve

Also: what-if scenario, scenario plan

A specific plan (budget, channel split, prices, timing) run through the model to forecast its outcome. Comparing two or three realistic scenarios is usually more persuasive to a finance team than a single "optimal" allocation.

Guide: MMM scenario planning → See also: Sensitivity analysis, Budget allocator

Repeating patterns in demand by time of year, month or week, like a December peak for gifts or a January spike for gyms. Because marketers tend to spend more in peak season, seasonality is the most common confounder in an MMM.

Guide: Baseline sales → See also: Fourier terms, Trend, Holiday effects

When the people (or weeks, or regions) exposed to an ad differ from those who weren't, in ways that affect buying. Ad platforms target people likely to convert, so comparing exposed with unexposed users overstates lift, sometimes by a multiple.

Guide: Experiments vs observational methods → See also: Randomised controlled trial, Endogeneity

Also: tornado chart, what-if analysis

Checking how much a recommendation changes when you vary the assumptions behind it, such as adstock, saturation or margin. If a modest change in one assumption flips the answer, say so in the recommendation.

Guide: MMM scenario planning → See also: Scenario

Also: shapley attribution, shapley values

A method from cooperative game theory (Lloyd Shapley, 1953) that credits each player with their average marginal contribution across every possible order of joining. In attribution, the players are channels, and it's the logic behind several data-driven models.

Guide: MMM vs multi-touch attribution → See also: Data-driven attribution, Markov chain attribution

Also: som, market share

Your share of category sales. Comparing it with share of voice is the basis of the excess share of voice rule of thumb.

Guide: The Long and the Short of It → See also: ESOV

Also: sov, advertising share of voice

Your share of total advertising in a category, by spend or impressions. It's a competitive measure: the same budget is loud in a quiet category and inaudible in a noisy one.

Guide: The Long and the Short of It → See also: ESOV, Share of market

Also: skan, skad network

Apple's privacy-preserving framework, introduced in 2018, that reports app-install attribution to ad networks in aggregated, delayed postbacks with limited conversion values. It tells you something about campaigns but nothing about individual users.

Guide: MMM vs multi-touch attribution → See also: ATT, AdAttributionKit

Also: contamination, leakage

When the treatment leaks into the control group, for example people in a control region seeing TV from a neighbouring test region or shopping across a border. Spillover shrinks the measured lift, so pick regions with clean boundaries.

Guide: Geo lift testing → See also: Geo test

Also: spurious correlation

A regression that finds a strong, significant relationship between two series that have nothing to do with each other, simply because both trend. Granger and Newbold described it in 1974, and it's why a channel that grew alongside the business can look like the cause of the growth.

Guide: Marketing econometrics → See also: Stationarity, Confounder

Also: stan language, rstan, cmdstan

A probabilistic programming language with its own NUTS sampler, callable from R, Python and elsewhere. Many in-house and vendor Bayesian MMMs are written directly in Stan.

Guide: Bayesian MMM → See also: NUTS, MCMC

Also: stationary series, non-stationary

A time series is stationary when its mean and variance don't drift over time. Most sales and spend series aren't, because they trend, and regressing one trending series on another invites spurious results unless you include the trend.

Guide: Marketing econometrics → See also: Spurious regression, Trend

Also: power, power analysis

The probability that a test detects an effect of a given size if it's really there, conventionally targeted at 80 percent. Underpowered marketing tests are the norm, and a "no significant lift" result from one tells you very little.

Guide: Statistical power for marketing tests → See also: MDE, Sample size, Alpha

Also: regime change, break point

A point where the relationship between inputs and sales changes, such as a pricing overhaul, a new competitor or a pandemic. A model that assumes one relationship across the break will average the two periods and describe neither.

Guide: MMM diagnostics → See also: Residual, Dummy variable

Also: synthetic control method, scm, augmented synthetic control

A counterfactual built as a weighted blend of untreated regions that together track the treated region's pre-period sales. Developed by Abadie and colleagues in the 2000s, it's what Meta's GeoLift uses, in an augmented form, to read geo tests.

Guide: Synthetic control method → See also: Donor pool, CausalImpact, Counterfactual

Also: test-and-learn, experimentation program

An ongoing programme of experiments, each designed to answer a specific budget question, with results fed back into planning and the MMM. The value comes from the cadence and the decisions, not any single test.

Guide: Incrementality testing → See also: Incrementality test, Calibration

Also: time decay

A rule-based model that gives more credit to touchpoints closer to the conversion. Like the other rule-based models, it encodes an assumption rather than measuring anything.

Guide: MMM vs multi-touch attribution → See also: Linear attribution, Position-based attribution

Also: time varying effects, tvc

A channel effect allowed to drift over time instead of staying fixed for the whole dataset, closer to how channels behave as creative and competition change. Uber's Orbit KTR model is built around it, and PyMC-Marketing supports time-varying media effects too.

Guide: Open-source MMM libraries → See also: Orbit

Also: tot, treatment effect on the treated, att effect

The effect on the people who actually received the treatment (saw the ad), rather than everyone assigned to it. If only 40 percent of the test group was reached, the effect on the treated is roughly the intent-to-treat effect divided by 0.4.

Guide: Conversion and brand lift studies → See also: Intent to treat

The slow, long-run direction of sales, up or down, from things like distribution growth, brand building or market change. Leave it out and any channel whose spend grew over the same period will steal the credit.

Guide: Baseline sales → See also: Seasonality, Stationarity

Also: measurement triangulation

Checking one measurement method against others with different weaknesses, such as an MMM against a geo test and platform data. When three imperfect methods agree you can act with confidence; when they disagree, that's where the next experiment should go.

Guide: Marketing measurement framework → See also: Unified marketing measurement

Also: umm, unified measurement

Combining MMM, experiments and attribution into one consistent view rather than three dashboards that disagree. In practice it usually means calibrating the MMM with experiments and using attribution for fast, within-channel optimisation.

Guide: Marketing measurement framework → See also: Triangulation, Calibration

Also: view through, vtc, post-view conversion

A conversion credited to an ad the person saw but didn't click. Some of these are genuine, but many people would have bought anyway, which is why view-through numbers are where platform reporting and incrementality tests disagree most.

Guide: Incrementality → See also: Attribution window, Incrementality

Also: variance inflation factor

The variance inflation factor measures how much a coefficient's variance is inflated because that input can be predicted from the others. A common rule of thumb treats a VIF above 5 as a warning and above 10 as a problem, though the real test is whether the estimates are stable.

Guide: Multicollinearity in MMM → See also: Multicollinearity

A more flexible adstock built on the Weibull distribution, where the decay rate can change over time instead of staying constant. Robyn offers a CDF version (decay that speeds up or slows down) and a PDF version that can push the peak effect several weeks after the spend. The extra flexibility costs you a second parameter to estimate.

Guide: Adstock → See also: Geometric adstock