Marketing Mix Modeling for B2B SaaS

MMM came back into fashion because cookies stopped working, and the vendors selling it did not check whether B2B SaaS has the data to support it. A consumer brand running MMM has millions of weekly purchases across a dozen channels. A 30M ARR SaaS company has maybe 60 closed won deals a month, a sales cycle measured in months, and three channels doing 85 percent of the spend. Those are not the same problem.

What MMM actually needs before it can tell you anything

Three conditions, all of them non-negotiable, and most mid market SaaS companies fail at least two.

Enough observations. MMM regresses an outcome against spend series and control variables. Weekly data over three years gives you 156 rows. That sounds fine until you count what the model has to estimate: a coefficient and an adstock parameter and a saturation parameter for each channel, plus seasonality, plus trend, plus controls for price changes and product launches. Eight channels at three parameters each is 24 before you’ve modelled a single control. The statistical term is degrees of freedom, and you run out of them fast.

Meaningful spend variation. A regression can only measure the effect of something that changed. If you spent 40,000 a month on LinkedIn every month for three years, the model has nothing to work with and will assign that channel’s effect arbitrarily to whatever else moved. Variation has to be real: on and off periods, step changes, regional differences. Teams that ran stable budgets and then commissioned MMM are asking a model to identify an effect from data containing no experiment.

A stable product and price. If you repositioned, changed pricing, launched a second product or moved upmarket during the modelling window, your outcome series contains structural breaks the model will attribute to media. That is the single most common source of nonsense results in B2B MMM, because sub-50M SaaS companies change their pricing roughly every 18 months.

The data density reality check for a company under 50M ARR

Run the arithmetic on your own business before anyone builds anything. Here is the comparison that makes the problem concrete.

The PLG row is the interesting one. If you have hundreds of weekly self serve signups, you can model that intermediate outcome with far better resolution than you can model closed won revenue. The catch is that signups and revenue are not the same target, and a channel that drives cheap signups can drive terrible customers. Model signups if you must, then check the quality split separately rather than assuming it holds.

Sales led companies have the hardest version of the problem. Forty opportunities a week is a small number, the variance between weeks is large, and the lag between spend and closed revenue can be six months. By the time the outcome lands, the model is fitting a relationship between this quarter’s revenue and last year’s spend, across a period when you also changed your ICP. The honest answer at that size is that the data cannot separate the channels, and any model that claims to is fitting noise with confidence.

The options ladder, from spreadsheet to vendor

There are four rungs, and most companies should stop on the first or second.

Option	Cost	Time to first result	Skill needed	Honest fit
Spreadsheet regression on aggregate spend	Free	1 to 2 days	Comfortable with regression basics	Sanity checking, never budget allocation
Meta Robyn (open source, R)	Free licence, 40 to 80 analyst hours	4 to 8 weeks	R, data engineering, statistics	Teams with a data scientist and 3 years of data
Google Meridian (open source, Python, Bayesian)	Free licence, 60 to 120 analyst hours	6 to 12 weeks	Python, Bayesian modelling, priors	Same, plus willingness to set informed priors
Commercial vendor (Recast, Prescient and similar)	$40K to $150K per year	8 to 16 weeks	Analyst to interpret and challenge	$50M ARR and above with 5+ funded channels

Robyn, from Meta, uses ridge regression with automated hyperparameter search. It’s fast to iterate and it will happily give you an answer from a thin dataset, which is both its convenience and its danger. Meridian, from Google, is Bayesian, which means you specify prior beliefs about each channel’s effect and the model updates them against the data. For small datasets that is genuinely better, because informed priors stop the model from producing wild coefficients when the evidence is weak. It also means the output partly reflects your assumptions, which you have to disclose honestly when you present it.

The spreadsheet rung deserves more respect than it gets. Regress monthly pipeline against monthly spend by channel with a lag term, look at the residuals, and you will learn whether any relationship exists at all before committing 80 hours. If nothing shows up in the crude version, a sophisticated model is unlikely to rescue it.

Reading adstock and saturation curves without over-reading them

Two outputs get quoted in every MMM readout and both are routinely misread.

Adstock is the assumption that spend this week keeps working next week, decaying over time. The model estimates a decay rate. With 150 observations and correlated spend series, that parameter is weakly identified, meaning it can shift a lot when you add one quarter of data. Treat a reported half-life of 3.4 weeks as “somewhere between two and six weeks” and cross-check it against your known sales cycle. If the model says your LinkedIn adstock decays in four days and your average cycle is four months, something is wrong with the specification, not with reality.

Saturation curves describe diminishing returns. The dashboard will show a curve and tell you that you’re at 60 percent of the saturation point on paid search, so you could add 40 percent more budget. Check what range of spend the model actually observed. If your search budget never went above 30,000 a month, the shape of the curve above 30,000 is extrapolation, and the model cannot know whether the next 20,000 works. Vendors rarely shade the unobserved region on the chart. Ask them to.

Why three incrementality tests a quarter beat one model

Here is the position: most SaaS companies under 50M ARR should not buy MMM. They should build a testing cadence instead, and the comparison is not close on cost or on decision quality.

An eight week geo holdout on one channel costs you the withheld spend in the holdout regions, plus a few days of analyst time to design and read it. Call it under 10,000 dollars of opportunity cost for a meaningful channel. Three of those per quarter gives you twelve causal readings a year for roughly the price of a junior contractor, and each one is an experiment rather than a coefficient. When a holdout says paid social produced no detectable lift in the withheld regions, that is a fact about your business. When a regression coefficient says the same thing, it’s a claim that depends on the specification.

The detailed test design, including sample size maths and how to pick matched regions, is in incrementality testing for B2B SaaS, with the statistical treatment in the guide version. If you’re weighing this against a touchpoint model, the multi-touch attribution versus incrementality testing comparison sets out where each one breaks.

When MMM does earn its place, and how to calibrate it

Three situations make a model worth building even at moderate size. You run significant offline or unmeasurable spend, such as events, podcast sponsorships, out of home or field marketing, where no click data exists at all. You operate across enough regions that geo tests are cheap and frequent, giving you real calibration inputs. Or you’re above 50M ARR with five or more funded channels and a genuine budget allocation problem that no single test can resolve.

Even then, the model is only as good as its calibration. The correct pattern is to run incrementality tests, then use those measured lifts as priors or constraints in the model, rather than letting the regression free-run. Meridian supports this directly through its prior specification. Robyn supports it through calibration inputs. A model that has never been checked against an experiment is a curve fitted to coincidence, however good the fit statistics look.

That table is the real decision. MMM is strongest on the unmeasurable and on interaction effects. Tests are strongest on everything you can switch off. Most sub-50M SaaS budgets are dominated by channels you can switch off.

What to do instead this quarter

Before commissioning anything, count your rows. Pull weekly spend by channel and weekly outcomes for as far back as your data is trustworthy, then count how many weeks you have and how many channels show real variation. Take that one page to whoever is pitching you MMM and ask them what confidence interval they expect on a channel coefficient given those inputs. The quality of that answer tells you whether they’re selling a model or a feeling.

Then set up the underlying data properly regardless, because both paths need it. A clean weekly spend and outcome table in your warehouse is the prerequisite for modelling, for testing, and for the reporting you already owe the board. Start with the marketing tracking plan template and the SaaS marketing dashboard template, check what the tools in your attribution stack can actually export, and benchmark your channel mix against published sales and marketing spend benchmarks. The wider measurement context sits in SaaS metrics and analytics, and if you’re rebuilding the plumbing anyway, the B2B SaaS marketing tech stack guide covers where warehouse-first setups pay off.