Media Mix Modeling for Small Teams: Spreadsheet vs Black Box

Below roughly $50,000 a month in paid spend, spread across fewer than four channels, media mix modeling for small business is usually over-engineering. A clean blended-CAC dashboard plus the occasional geo holdout will move your budget decisions more than a Bayesian model you can't fully audit. MMM earns its place later, and this piece is about spotting the exact moment it does.

I'll show my work, because that's the whole point of the argument. If you're going to tell a CFO you don't need the expensive thing yet, you'd better have the math to back it up.

The threshold nobody wants to give you

Vendors won't hand you a number, because the honest number disqualifies most of their inbound. So here's mine, spelled out, and you're welcome to disagree with the cutoffs.

Monthly paid spend	Active paid channels	What should actually drive your budget calls
Under $20k	1-2	Blended CAC, with platform ROAS sanity-checked against it. An MMM here is fitting a curve to noise.
$20k-$50k	2-3	Blended-CAC dashboard, plus one geo holdout per quarter on your biggest line item.
$50k-$150k	3-5	Add a lightweight open-source MMM as a second opinion. Don't let it overrule a clean holdout.
Over $150k, 5+ channels	5+	MMM starts earning its keep. Triangulate it with incrementality and attribution.

The load-bearing idea: MMM is a statistical model that needs enough variation in your spend, over enough time, to separate one channel's effect from another's and from everything else (seasonality, promos, that PR hit you forgot to log). Small budgets don't generate that variation. You end up with wide confidence intervals dressed up as precision, and a model that confidently tells you Meta drove 34% of revenue when the honest answer is "somewhere between 10% and 55%, we can't tell."

Analytical Alley's 2025 write-up on minimum MMM budgets puts the floor for a basic model around €300,000-500,000 in annual marketing spend, with each measured channel wanting roughly €8,000-10,000 a month before it's worth including. Their sharper rule is the one I'd tattoo on a budget deck: your marketing budget should be about 50 times the cost of the MMM initiative for the optimization gains to pay for the model. A team spending €100k a year that squeezes a 15% improvement out of an MMM saves €15k, and probably spent more than that getting the model built and maintained.

One caveat on those minimums, and it matters. Those thresholds come from firms whose clients are, by definition, the ones who could already afford MMM. That's survivorship bias baked into the benchmark. The €300k floor isn't a law of statistics; it's the spend level below which the vendors stopped taking meetings. Google's free Meridian and Meta's Robyn have dragged the real floor down, but they didn't repeal the math about spend variation. A free tool still can't invent signal your budget never produced.

What MMM actually does, in plain terms

Two ideas do most of the work, and once they click, the whole method stops looking like a black box.

Adstock is the memory of advertising. You run a burst of spend this week, and some of the effect shows up next week, and a little the week after. Adstock is the model's way of saying "advertising doesn't cash out the day you pay for it." A high adstock channel (brand video, say) keeps paying you back for a while. A low adstock channel (a flash-sale search campaign) fires and fades.

Saturation is the ceiling. Your first $10k on a channel buys cheap conversions. Your tenth $10k buys expensive ones, because you've already reached the easy audience. The response curve bends. The 2025 ISJEM paper on adstock and saturation curves lays out the two shapes most models use: an exponential curve (diminishing returns from the first dollar) and the S-shaped Hill curve (a slow start, a steep middle, then a flattening top). The Hill curve is the more realistic one for channels with a threshold effect. You need a minimum push before anything happens, then it takes off, then it stalls.

That's the honest core of MMM. Adstock plus saturation, fit across your channels, so the model can tell you "shift $5k from the saturated channel to the one still on its steep slope." It's genuinely useful. It's also completely wasted if your spend history is three months long and your channels barely moved.

Here's the thing a CFO I worked with said when I first pitched her on an MMM, and she was right: "You're proposing to spend money modeling the diminishing returns of channels we haven't spent enough on to have diminishing returns." We shelved it for a year and spent the modeling budget on a proper geo test instead. Best call we made that quarter, and the test answered the one question the model would have fudged.

What beats it, at your size

If you're under the threshold, you're not measurement-blind. You're just using cheaper instruments that happen to be more honest about their error bars.

Blended CAC as the ground truth. This is the number that can't lie to you, because it doesn't depend on any attribution model. Total marketing and sales spend, divided by net new customers, over a period.

Blended CAC = (all marketing spend + all sales spend) / new customers (same period)

Everything else is a story about why that number moved. Blended CAC is the number itself. When founders run the business day to day, they steer on blended CAC and use channel-level numbers only to decide where to lean in, which is exactly the right instinct. If you want the full dashboard version of this, we walked through it in the unit economics dashboard template; the blended-CAC panel is the one that survives every attribution argument.

The blended-vs-paid gap tells you what you're missing. Paid CAC is ad spend over paid-attributed customers. Blended CAC is everything over everyone. The gap between them is your organic subsidy: brand search, direct, email, referral, word of mouth, the customers arriving without an attributable click. Eightx's 2026 benchmark work on that gap makes the practical point sharp: brands leaning on platform-reported ROAS alone tend to over-credit Meta by 20-40% and under-credit email, SMS, and organic. An MMM can quantify that split for you. So can watching your blended CAC hold steady while you cut paid spend 20% and seeing what actually happens.

Geo holdouts for the channel that scares you most. When you genuinely need causal proof (is this channel incremental, or just claiming credit for buyers who'd convert anyway?), you don't need a model. You need a controlled test. Turn a channel off in a few matched regions, leave it on elsewhere, and measure the difference. It's the closest thing to a clean experiment marketing has. Measured's decision tree on when to use incrementality versus attribution versus MMM lands in the same place: for causal proof on a specific channel, a holdout beats a model. If you're running mobile UA specifically, geo holdout and PSA test designs get into the mechanics that don't fit here.

One geo holdout, run properly on your biggest channel once a quarter, will resolve more real budget arguments than a full MMM will at your spend level. And it costs you a few weeks of suppressed spend in a couple of markets, not a modeling engagement.

For most sub-$1M-per-month teams, a plain analytics tool that gives you honest blended metrics and funnels (the category Kixo sits in, alongside the usual product-analytics suspects) closes the loop better than a premature model does.

The assumptions behind all of this

I'd rather you argue with my inputs than with my conclusion, so here they are in the open.

What I'm assuming:

Your paid mix is fewer than five channels and hasn't changed structure in the last quarter. (More channels, or a recent overhaul, shifts the threshold up.)

You have under ~18 months of weekly spend history. MMM wants 2-3 years to see seasonality; short histories overfit.

Your spend on each channel has been stable: you haven't deliberately varied it. Ironically, stable spend is bad for MMM, which needs the variation to identify effects.

You can actually run a geo holdout, meaning you have regional targeting and enough volume per region to read a result.

Break any of these and the threshold moves. A team with three years of history and wildly varying spend across six channels can get a real answer from Meridian at lower total spend than my table implies. Tell me where your situation differs and the recommendation should flex with it.

"But everyone's saying MMM is back"

They are, and they're not wrong about the trend — they're wrong about who it applies to.

The resurgence is real and it's driven by privacy, not hype. As third-party cookies and user-level signal degraded, multi-touch attribution stopped covering the map. Leadgen Economy's 2026 rundown on the cookieless stack pegs the collapse bluntly: MTA coverage fell from 90%-plus to roughly 30-60%, and it's now a tactical layer rather than a single source of truth. MMM works on aggregate data — it never needed user-level tracking — so it survived the privacy shift intact. That's why Google shipped Meridian globally in January 2025 and Meta open-sourced Robyn, and why, per reporting cited in that same coverage, close to half of marketers say they're increasing MMM investment.

All true. All aimed at enterprises drowning in a measurement gap you don't have yet. When you run two or three channels, you can see your marketing. You don't need a model to reconstruct causality from aggregate ghosts, because you can just turn a channel off and watch. The privacy apocalypse that makes MMM a gold standard for a brand spending $2M a month is barely a rain cloud for a brand spending $30k across Meta and Google.

The cookieless case for MMM is legitimate. It just isn't a reason for a small team to buy the model early. It's a reason to keep clean first-party data and blended metrics now, so that when you do cross the threshold, your MMM has trustworthy inputs instead of a Frankenstein of platform exports.

When to actually pull the trigger

Concrete signals you've outgrown the spreadsheet, in rough order of how much they should move you:

You're running five or more paid channels and can no longer reason about cross-channel effects in your head. Your budget meetings have become vibes.
Your monthly paid spend cleared $50k-$100k and small percentage reallocations are now worth real money — a 10% shift is five figures.
You have two-plus years of weekly data with genuine variation in it, so a model has something to chew on.
Platform-reported numbers add up to more than 100% of your actual revenue, and the double-counting is making your channel decisions worse.

Hit three of those four, and MMM stops being over-engineering and starts being the cheapest way to get an answer. Start with open-source Meridian or Robyn before you sign with a vendor — treat the first model as a second opinion against your holdouts, not as gospel. If the MMM and your geo tests agree, trust the direction. If they fight, trust the test, because a holdout is an experiment and a model is a well-informed guess.

The one-line version for your budget deck

MMM is a great answer to a question small teams mostly don't have yet: "with too many channels and no user-level signal, where's the marginal dollar best spent?" Under $50k a month and a handful of channels, you can answer that by looking. Blended CAC tells you if you're winning. The blended-versus-paid gap tells you what's uncredited. A geo holdout tells you what's incremental. That's a measurement stack that fits on one dashboard and doesn't require you to defend a model you can't fully open up in front of your CFO.

Buy the black box when the spreadsheet stops fitting the problem. Not before. And when you do, keep the spreadsheet — it's still the thing that catches the model when it lies.

FAQ

How much ad spend do I need before MMM is worth it? As a working rule, meaningful MMM wants channels each doing roughly $8k-10k a month and a total budget large enough that a 10-15% optimization pays back the modeling cost several times over — Analytical Alley's 50x-of-model-cost guideline is a good gut check. Below about $50k a month across fewer than four channels, blended metrics and geo holdouts do more for less.

Is free MMM software like Meridian or Robyn enough for a small business? The software being free doesn't fix the underlying data requirement. Meridian and Robyn lowered the cost of building a model, not the amount of spend variation and history a model needs to produce trustworthy estimates. Free tools are worth it once you're near the threshold; they don't move the threshold much.

What's the difference between blended CAC and channel CAC? Blended CAC divides all acquisition spend by all new customers and depends on no attribution model, so it can't be gamed. Channel CAC divides one channel's spend by the customers credited to it, and that credit depends entirely on which attribution model you picked. Steer the business on blended; use channel CAC to decide where to lean in.

Does MMM replace attribution and incrementality testing? No — the 2026 consensus is triangulation. MMM gives strategic breadth on aggregate data, incrementality (geo holdouts, PSA tests) gives causal proof on a specific channel, and attribution gives tactical day-to-day granularity. For a small team, you mostly need the last two; the first joins the stack later.