A CMO I worked with got the same question from her board three quarters in a row. "Which channels are actually driving revenue?"
Each quarter she brought a HubSpot attribution report. Each quarter someone pointed out that a third of pipeline was credited to direct traffic and branded search, which is what happens when a buyer hears about you on a podcast and then types your name into Google six weeks later. The report was not wrong. It was just answering a different question than the one being asked.
By the third quarter she had a proposal on her desk. An analytics vendor wanted $85,000 and twelve weeks to build her a marketing mix model. She asked me whether to sign it.
I said no. Eighteen months later she signed a similar contract and I told her it was the right call. Nothing about the vendor changed. What changed was how much data she had.
That gap is what this post is about. Marketing mix modeling is having a real moment right now, partly because cookie-based tracking keeps degrading and partly because Google and Meta both open-sourced credible MMM libraries. A lot of B2B companies are being sold models they cannot statistically support. A smaller number are sitting on enough data to run one and have not noticed.
What marketing mix modeling actually does
Strip the marketing off it and MMM is regression on a time series.
You take a weekly outcome (revenue, pipeline created, new customers). You take a weekly series for every input you control: spend per channel, event budget, headcount, price changes. You take a weekly series for things you do not control: seasonality, competitor launches, macro conditions. Then you fit a model that estimates how much each input contributed to the outcome.
There are no cookies involved. No pixels, no user-level tracking, no identity resolution. That is the entire appeal in 2026. MMM sees aggregate spend and aggregate outcomes, so nothing Apple or Google does to third-party identifiers touches it.
Two mechanics separate a real MMM from a spreadsheet correlation:
Adstock models carryover. Money spent on LinkedIn in week 3 does not stop working in week 3. The model estimates a decay curve so week 3 spend still contributes something in weeks 4 through 9.
Saturation models diminishing returns. Your first $10k on paid search buys better clicks than your tenth $10k. The model fits a curve rather than a straight line, which is the part that lets you answer "what happens if I add 20% to this channel."
Those two together are why MMM can tell you something an attribution report cannot: not who touched the deal, but what would have happened if you had spent differently.
If your real question is "which lead do I work first," you want lead scoring, not a mix model. If your question is "we have $2M next year and I do not trust the attribution report," MMM is the correct family of tools. It is a planning instrument, not an operational one.
Why B2B breaks the standard model
MMM was invented for consumer packaged goods. Procter and Gamble, weekly detergent sales, millions of transactions, television spend that varies by region. Almost every assumption baked into the method comes from that world, and B2B violates most of them.
Volume. Regression needs observations. A company closing 60 deals a year has roughly one deal a week across 104 weeks of history, and a single enterprise deal can be 20% of the quarter. That is not a time series, it is noise with a trend. A B2B company with 1,000+ annual deals and a 3 to 6 month cycle gets usable models. One with 50 target accounts and $1M deal sizes does not, no matter how good the vendor is.
Lag. B2B sales cycles run 6 to 18 months at the enterprise end. Marketing spend in January shows up as closed revenue in November. Standard adstock decay curves assume carryover measured in weeks, and if you model closed revenue with a 12-month cycle you need something closer to four years of history before the estimates settle down. Most companies asking for MMM have been running paid channels for two.
Geography. The strongest MMM designs use geographic variation, and Google's Meridian is built around exactly that. A US-only B2B SaaS company selling to a national buyer list has no meaningful geo split. You lose the cleanest source of variation available and fall back to national time series alone.
The biggest input is not media. For most B2B companies under $50M, the single largest driver of pipeline is how many SDRs and AEs were on the floor that quarter. If you do not put headcount in the model as a variable, the model will quietly attribute a hiring wave to whatever channel happened to increase at the same time. I have seen this exact error produce a recommendation to triple paid social spend. What actually happened was two new SDRs ramping.
Weeks of clean weekly data is the standard minimum for a marketing mix model, which is two full seasonal cycles. Some practitioners will start at 78 weeks if the signal is strong and spend actually varied.
The bar nobody quotes you upfront
Vendors do not lead with data requirements because data requirements lose deals. Here is what the practitioner consensus looks like in 2026.
On history, target 104 weeks or more. Two full seasonal cycles is the standard entry bar, and 78 weeks is the aggressive floor when the signal is unusually clean.
On spend, most analytics shops put the trigger point around $3M in annual media spend before a national MMM makes sense. European practitioners quote €300,000 to €500,000 as a basic-model minimum with €500,000+ for something comprehensive. The number matters less than the reason behind it: below a certain spend level, the week-to-week variation in your budget is smaller than the noise in your revenue, and the model cannot separate the two.
On variation, this is the requirement people skip and it is the one that kills projects. If you have spent exactly $40,000 a month on LinkedIn for two years, no model on earth can tell you what $60,000 would do. MMM learns from change. A perfectly stable budget is a perfectly uninformative dataset. When I audit a client's readiness, the first thing I pull is monthly spend by channel for 24 months, and if the line is flat I stop there.
MMM has a data threshold, not a company-size threshold.
A 40-person company with three years of varied spend and 800 deals a year can run a better model than a 400-person company that has spent the same amount on the same three channels every month since 2024.
The three-question test
Before anyone quotes you a price, answer these honestly.
Can you produce 24 months of weekly spend by channel from your own records? Not "could we reconstruct it." Can you export it this week. Most teams discover their 2024 spend lives in a mix of ad platform exports, an accounting system, and one person's memory of what the agency retainer covered.
Do you close enough deals for a weekly series to mean anything? My working line is 300+ opportunities created per year as a floor, and I would rather see 1,000. Below that, model pipeline created or SQL volume instead of closed revenue, because the earlier funnel stage has more events and shorter lag.
Did your mix actually change? Look for at least one channel that doubled, one that got cut, or a large one-off like a conference or a brand campaign. Those are the events the model learns from.
Two out of three and you are borderline. One out of three and you are buying an expensive opinion with confidence intervals printed on it.
What to run instead when you fail the test
This is the part vendors will not tell you, and it is where most B2B companies under $30M should spend their measurement effort.
Run a spend-step test. Pick one channel. Cut it to zero for four weeks, or double it for six. Watch pipeline created, not clicks. This is crude and it costs you real money in the down weeks, but it produces a causal answer for one channel that no correlational model can match. I have run this on branded search with three clients now. Two found that roughly 60% of branded search conversions happened anyway through direct navigation. One found the opposite and kept spending.
Ask the buyer. Put a free-text "how did you hear about us" field on your demo request form and make it required. Self-reported attribution is messy and biased toward recency, and it is still the only channel that captures the podcast, the Slack community, and the conversation with a former colleague. Store the raw answer in the CRM, categorize it monthly (an automation that classifies free text into buckets takes an afternoon to build), and compare it against what your attribution model says. The gap between those two numbers is your dark-social volume, and for most B2B companies it is uncomfortably large. This pairs well with a working multi-touch attribution setup rather than replacing it.
Watch branded search volume as a brand proxy. If podcast sponsorships and events are working, branded search volume rises. It is a lagging, imperfect signal that costs nothing to track in Search Console.
Run a geo test if you actually have geography. Proper design wants 20 or more comparable markets and 4 to 6 weeks of exposure, then a difference-in-differences read on backend revenue rather than platform-reported conversions. Most B2B SaaS cannot meet the market count. Companies with field sales, regional events, or multi-country operations sometimes can.
If you qualify, build it on pipeline
The single biggest design decision in a B2B mix model is the outcome variable, and the default choice is usually wrong.
Modeling closed revenue feels right because it is the number the board cares about. It is also the number sitting furthest from the spend, separated by a 134-day average sales cycle and a win rate that moves for reasons having nothing to do with marketing. A rep leaving in Q2 will show up in your model as a marketing effect.
Model pipeline created instead, dated to the opportunity creation date. Or model SQLs if your opportunity volume is thin. You get more events, shorter lag, and less contamination from sales execution. Then convert to revenue outside the model using your actual stage conversion rates. The model tells you which spend creates pipeline. Your funnel math tells you what that pipeline is worth.
Set the lag structure to match reality. If it takes 30 days from first touch to opportunity creation in your data, your adstock needs to allow for it. Pull the real distribution from your CRM first. Do not accept a vendor default built for e-commerce.
Step 1 and step 2 are where the project actually lives. In every engagement I have run, the modeling took days and the data reconstruction took weeks. If your CRM opportunity data is unreliable, fix that before you think about models. A mix model built on a messy pipeline inherits every problem in it, which is the same reason CRM data quality work pays for itself before any analytics project starts, and why most of our CRM and RevOps engagements begin with the opportunity object rather than a dashboard.
The tools are free now, the data work is not
Three open-source libraries cover most of what a mid-market company needs.
Meta's Robyn uses ridge regression with an automated hyperparameter search. It is the fastest to a first result and the friendliest if your team is more analyst than statistician.
Google's Meridian is Bayesian and hierarchical, and it is genuinely better when you have regional data. Google added support for non-media variables and channel-level priors in late 2025, and shipped Meridian GeoX for geo-incrementality in May 2026. It is also the most technically demanding of the three, and much of its advantage assumes a geo split most B2B companies do not have.
PyMC-Marketing is the most flexible and the most downloaded. If you want custom priors, custom lag structures, or a model that reflects a strange business, this is where you end up.
The libraries removed the six-figure software line item. They did not remove the analyst, and they definitely did not remove the data engineering. Budget for a person who can defend the model in a board meeting, because a number nobody can explain gets ignored the first time it disagrees with someone's intuition.
How to use the output without wrecking the team
MMM results arrive as contribution percentages and response curves. Two rules for what happens next.
First, treat it as a budget allocation tool and nothing else. The moment MMM output starts driving individual channel owners' bonuses, you have created a strong incentive to argue with the model instead of the market. Reallocations driven by MMM typically produce 10% to 25% efficiency gains without raising total spend, and that is a planning-cycle number, not a monthly performance review number.
Second, move budget in increments and check. If the model says paid social is at 40% of saturation and search is past it, shift 15% and watch what happens over the next quarter. Do not shift 60% on the strength of one model run. Every mix model is an estimate with real uncertainty, and the honest ones publish the confidence intervals alongside the point estimates. Ask to see them.
The mistakes I keep seeing
Running MMM on closed revenue with a 9-month sales cycle and 24 months of data. There are not enough independent observations there to learn anything, and the model will still return a confident-looking chart.
Leaving sales headcount out. This one is almost universal and it corrupts everything downstream.
Treating the output as attribution. MMM does not assign credit for deals. If someone asks "which channel closed the Acme deal" and you point at a mix model, you have misunderstood the tool.
Building it once. A mix model is a living thing that needs refitting quarterly as the mix changes. A one-off report is a very expensive snapshot of a business that has already moved.
Skipping the validation test. If the model says something surprising, test it with real money on one channel before you rebuild the plan around it. The surprising results are the ones most likely to be an artifact.
Not sure whether your data can support a mix model?
We audit spend history, CRM pipeline data, and channel variation, then tell you plainly whether MMM is worth it or which cheaper test would answer the same question.
Book a measurement audit →Frequently asked questions
How much does marketing mix modeling cost in 2026?
Vendor engagements run from roughly $30,000 for a lightweight annual model to $150,000+ for continuous enterprise setups with quarterly refits. Building in-house with Robyn, Meridian, or PyMC-Marketing costs nothing in software and typically 4 to 8 weeks of a data person's time for the first model, with most of that spent reconstructing historical spend rather than modeling.
Can a company spending $500k a year on marketing run MMM?
Sometimes, and less often than vendors suggest. The general trigger point for a national model is around $3M in annual media spend. Below that, what matters more is whether your spend varied and whether you create enough opportunities weekly for the series to be readable. A company at $500k with genuinely lumpy spend and 800 opportunities a year is a better candidate than one at $2M spending the same amount every month.
Does MMM replace multi-touch attribution?
No, they answer different questions. MTA tells you which touchpoints appeared in a buyer's journey and is useful for operational routing and campaign optimization. MMM estimates what each channel contributed in aggregate and is useful for budget planning. Most mature teams run both and use incrementality tests to check them against each other. Our attribution model guide covers where each one breaks.
What data do I need before starting a marketing mix model?
At minimum: 104 weeks of weekly spend by channel including agency and event costs, weekly outcome data (pipeline created or closed revenue) from your CRM, and control variables covering sales headcount, pricing changes, product launches, and seasonality. Missing the control variables is the most common cause of nonsense results.
How long before a mix model produces something usable?
If your data is clean, four to six weeks to a first model. In practice, data reconstruction is the long pole and eight to twelve weeks is more realistic for a first-time build. Ongoing refits take days once the pipeline exists, which is the argument for building the data plumbing properly the first time rather than treating it as a one-off project.
Where this leaves you
Most B2B companies under $30M in revenue should not buy a marketing mix model this year. They should fix their pipeline data, add a required "how did you hear about us" field, and run two honest spend-step tests on their largest channels. That combination costs almost nothing and answers 80% of the board's question.
The companies that should run one usually already know it. They have three years of history, real variation in spend, enough deal volume for a weekly series, and a specific decision waiting on the answer. If that is you, start with the data audit and pick a library. If it is not you yet, keep collecting clean weekly data, because in two years it will be.
Want help figuring out which group you are in? We do this as a fixed-scope audit. Get in touch and we will look at your spend history and CRM data before anyone talks about models.