Base sales are the sales a model expects you would have made without the marketing activity it’s measuring, and incremental sales are the extra it credits to that activity. In a marketing mix model the base (or baseline) is usually the intercept plus trend, seasonality and the other non-marketing drivers, while in retail promotion analysis base volume is what a product would have sold in a week without the promotion.
The split looks like a fact on a decomposition chart, a big blue block of baseline with thin colored slices of media on top, but it’s an estimate like any other part of the model. In my experience of building marketing mix models, the conversation that matters most is often about the baseline rather than the channels, because whatever the baseline doesn’t explain is left for media to claim, and whatever it explains too generously is taken away from media. The course lesson on the baseline is titled “‘Not credited to media’ is not a complete explanation”, and that’s the attitude I’d bring to any decomposition. I’ll start with what the baseline contains, then show how much credit can move between base and media when you change what the model includes.
Base sales vs incremental sales
The simplest way to see the split is to write the model down. Google’s 2017 paper on Bayesian MMM by Jin and colleagues uses a standard additive form, where sales in week are:
The paper describes , the intercept, as the baseline sales. The media terms are the incremental contribution of each channel after adstock and a saturation curve, the terms are controls such as price or weather, and is the error. In practice most practitioners, and most tools, report the intercept, trend, seasonality and the control variables together as the base, and the marketing terms as incremental, although which controls go into which bucket is a choice rather than a law.
Incremental here means the same thing as it does in an experiment: the difference between what happened and what the model thinks would have happened without that activity, which is why the baseline is really the model’s counterfactual. The incrementality page covers that idea more generally. The difference from an experiment is that nobody randomly withheld the marketing, so the counterfactual comes entirely from the model’s assumptions about what else drives sales.
What the intercept and trend absorb
The course lesson uses a deliberately simple additive example from its invented Northstar case, which is worth reproducing because it shows that a decomposition is more than a two-way split:
| Part | Weekly revenue |
|---|---|
| Baseline prediction | £100,000 |
| Modeled media contribution | £30,000 |
| Predicted total | £130,000 |
| Observed total | £135,000 |
| Residual | £5,000 |
The residual is the gap between what was observed and what the model predicted, and it isn’t a hidden channel waiting to be assigned to whoever asks first. The intercept itself is just the prediction at zero spend: in a line like predicted sales = 100 + 2 × spend (both in £k), the intercept of £100k is what the model expects with no spend at all, and the slope of 2 adds £40k at £20k of spend for a prediction of £140k. Neither number on its own proves anything causal, and a zero-spend prediction is an extrapolation if the business has never actually spent zero.
What makes the baseline slippery is that it’s the model’s catch-all for everything that isn’t explicitly modeled as marketing. That includes underlying demand and growth, seasonality, holidays, distribution gains, price changes (if price isn’t its own variable), competitor activity, the macro economy, repeat purchasing from customers acquired in previous years and the long tail of past advertising. If a demand change is omitted and a media variable moves at the same time, the media variable can absorb some of it and look better than it is, which is the problem covered in confounding variables in MMM. Conversely, a very flexible time component can absorb variation that genuinely belongs to media, and the model will fit just as well.
Watch credit move between the baseline and media
The widget below builds two years of synthetic weekly sales from a known baseline (a level, underlying growth and seasonality) and a known media effect, then fits a regression with whichever baseline terms you choose. Because the data is invented, you can see the true split alongside the model’s, which you never get to do with real data.
The model's decomposition: baseline plus media, with actual sales as dots
Credit to media each week: this model vs the true effect
Two-year totals
With the defaults (a linear trend and seasonality, matching how the data was built), the model’s media share lands close to the truth. Turn seasonality off and media’s share jumps by more than half, because spend is partly timed to the seasonal peaks and the model has nowhere else to put the peak. Turn the trend off and media picks up some of the business’s growth, because spend grew too. Switching to a flexible trend barely changes the answer here, because the media flights give the model plenty of week-to-week variation to learn from. Push “how much spend follows the season” to 100 percent and that variation disappears. With both trend and seasonality in the model, the estimate now drifts below the true media share (the flexible trend a little further than the linear one), because once spend moves smoothly with time the baseline terms can follow it, while dropping either term sends media’s share far above the truth. In every case the fit stays respectable, and the R² usually improves as you add baseline terms, which is a good reminder that a well-fitting model can still tell the wrong story.
Why “base” isn’t “organic”
People often hear baseline and translate it as organic, the sales you’d get for free. That’s misleading in both directions. On one side, the baseline usually contains effects marketing caused: the lingering effect of advertising older than the model’s adstock window, brand equity built over years, and repeat purchases from customers marketing originally acquired. On the other side, some things people call organic, like organic search or direct traffic, are partly driven by paid media (a TV campaign that lifts branded searches is a classic case), so if you model them as separate drivers rather than outcomes, you can end up moving media’s effect into the base.
Different tools also draw the line in different places. Meta’s Robyn documentation says plainly that there’s no clear definition of baseline, and Robyn defines it as the non-steerable variables: trend, season and holiday effects from its Prophet decomposition, external factors like competitors and weather, and the intercept. Price and promotions, by contrast, are counted with the steerable marketing variables. That’s a sensible convention for a media budgeting tool, but it means Robyn’s baseline excludes promotions while a retail analyst’s base volume is defined precisely as sales without promotions. Two people can both say “baseline” and mean different numbers, so it’s worth asking which variables are in the bucket before comparing shares.
Brand equity lives in the base
This is where the baseline gets political. Advertising that builds a brand tends to work slowly, over months and years, and much of that effect never shows up as a weekly correlation between spend and sales. In a typical weekly MMM with a few weeks of adstock, the long-term effect of brand building lands in the intercept and the trend, which means the model credits it to the baseline, not to marketing. That’s one reason brand advertisers argue that short-term MMM undervalues them, and it’s the territory of Binet and Field’s work on the long and the short of it.
However, the course lesson makes the opposite point just as firmly: don’t read a smoothly rising baseline as a direct measurement of brand strength. A growing intercept could be brand equity, but it could equally be distribution gains, a price cut, a competitor exiting the market or a trend term absorbing anything slow-moving. If you want to claim that marketing built the base, you need evidence beyond the decomposition, such as brand tracking, a long-running geo test, or a model specifically designed to estimate long-term effects. In my experience the honest version is to say which part of the base you believe marketing contributed to, and how you’d test it.
Base and incremental volume in retail and CPG
Retail and consumer packaged goods have their own long-standing version of this split, used mainly for trade promotions. Base volume is what a product would have sold without the promotion, and incremental volume is the extra sold during it, so promotional lift is incremental volume divided by base. Syndicated data providers typically estimate the base from an average of non-promoted sales over a lookback window before the event (often somewhere between four and 12 weeks), adjusted for seasonality and distribution changes, and then treat anything above that during the promoted weeks as incremental.
That method has the same weaknesses as an MMM baseline, just in a smaller window. The base is a model, so each provider’s choices about the lookback window, seasonality and distribution produce different lift numbers for the same promotion, and you should only compare lifts within one data source. The standard calculation also can’t tell genuinely new demand from pull-forward and pantry loading, where existing buyers stock up during the deal and buy less afterwards, which shows up as a post-promotion dip below base. An MMM that includes a promotion variable and a few weeks of lag (the dummy variables brief covers the mechanics) can catch some of that dip, which is one advantage of modeling promotions alongside media rather than in a separate silo.
Checking a baseline before you trust the media numbers
Because the baseline sets how much credit is left for marketing, it deserves the same review as any channel. These are the questions I’d ask about a decomposition before anyone takes its media returns to a budget meeting:
The last question is the subject of MMM model diagnostics, which covers reading residuals for exactly these misses, and the seasonality brief goes into how to model seasonal patterns. If you want to work through a full baseline review on a realistic case, the Vexpower MMM course has a lesson on understanding the baseline in its invented Northstar case, where the task is to reconcile one week’s prediction before anyone argues about the channels.
