Marketing mix modelling for mid-market B2B: is it within reach?

Marketing budgets are under pressure while the cost of misallocation rises. Gartner's 2025 CMO Spend Survey, covering 402 marketing leaders in North America, the UK and Europe, found budgets flat at 7.7% of company revenue, with 59% of CMOs reporting they lack the budget to execute their strategy. Paid media accounts for 30.6% of marketing budgets, so the largest line is also the one most exposed to attribution error. When budgets are flat, every pound moved on the evidence of a flawed report is a pound taken from something that may have been working.

Cost-cutting is already reaching the support structure. Gartner found that 39% of CMOs plan to cut back on agency budgets, which in many mid-market firms removes the analytical capacity that produced the reports. The CMO is left with platform dashboards, each of which credits itself, and a sales team that attributes wins to relationships. Nobody has a view of how the channels work together, or what would happen to pipeline if one were cut by a fifth.

Buyer behaviour compounds the problem. Gartner's survey of 632 B2B buyers, fielded in August and September 2024, found that 61% prefer an overall rep-free buying experience, and that 73% actively avoid suppliers who send irrelevant outreach. Much of the influence on a B2B purchase therefore occurs where no click or form fill is recorded: in peer conversations, search, events and content consumed anonymously. Last-click reports see only the final, visible step and systematically undervalue the activity that created the demand.

For a mid-market firm the exposure is sharper than for a large advertiser. A multinational can absorb a misjudged channel across dozens of markets; a business with a single marketing lead and a three-person demand team cannot. The cost of error is also asymmetric. Cutting a channel that quietly feeds the pipeline shows up in revenue two or three quarters later, by which time the cause is forgotten. Over-investing in a channel that merely captures existing demand produces good-looking dashboards and flat growth. Both mistakes are predictable consequences of measuring the last visible step.

What to do: make the shortcomings of the current reports explicit to the leadership team before proposing a new method. List the three biggest budget lines, the evidence currently used to justify each, and what that evidence cannot see. That inventory is the case for measurement investment, and it defines what the first model must answer.

The evidence of confidence is consistent. In the EMARKETER and TransUnion Marketing Measurement Confidence survey, fielded in July 2025, 27.6% of US brand and agency marketers found MMM the most reliable measurement methodology. Multi-touch attribution came second at 19.4% and unified, holistic measurement solutions third at 18.9%. In a separate EMARKETER and Snap survey from July 2024, 30.1% of US marketers said MMM was best at identifying drivers of business value, ahead of web analytics at 20.2%, incrementality lift testing at 19.9% and third-party multi-touch attribution at 11.7%.

Intent is rising. The 2025 TransUnion survey found that nearly half, 46.9%, of US brand and agency marketers plan to invest in MMM over the next year. The 2024 Snap survey reported that 53.5% of US marketers already use MMM, and that 61.4% of marketers spending US$500,000 or more a year on digital advertising want better or faster MMM. These samples are American and skew towards larger advertisers, so they should not be read as a measure of European mid-market practice. They do show the direction in which the market's confidence is moving.

Two cautions apply. First, a plurality of 27.6% is not a majority: marketers disagree about what is most reliable, which is itself a reason to combine methods rather than choose one. Second, using MMM and benefiting from it are different things. A model that is commissioned annually, presented at a leadership offsite and then ignored has not changed any decisions. The useful measure of maturity is not whether a firm has a model but how many budget choices it has altered because of one.

European readers should note a gap in the public evidence. Most of the large, recent surveys of MMM adoption are American, and the Gartner CMO survey is the only one cited here that includes UK and European respondents. We have not found a comparable, recent European survey of mid-market B2B measurement practice, and we would treat any claim of a precise European adoption rate with caution. What the American data show is the sequence of change, from click-based attribution towards aggregate and experimental methods, which regulatory and platform changes in Europe have, if anything, reinforced.

What to do: define the decisions the model must inform before choosing a tool. For most mid-market B2B teams there are three: the split between paid search, paid social and events; the split between brand and demand capture; and the level of total spend. If the model cannot move those three, it is not worth building.

Google made Meridian, its open-source marketing mix model, available to everyone after testing with hundreds of brands, and MarTech reported that more than 20 measurement partners are certified to help companies implement it. Meta's Robyn offers a similar open route. For a mid-market firm, the economic change is real: modelling that once depended on a proprietary licence can be run with published code on a modest cloud budget by someone competent in statistical modelling, using methods that are published and inspectable.

The constraint has moved from software to skills. A Harvard Business Review study sponsored by Google, reported by EMARKETER, found that 45% of organisations cite lack of expertise as a barrier to acting on MMM insights. Funnel and Ravn Research found in April 2024 that only 26% of in-house marketers conduct MMM in-house. Both findings imply that many teams buy modelling from outside, and that the buyer needs enough knowledge to judge the work. A model delivered as a black box transfers none of its learning to the team that must act on it.

The same research points at data quality as the lever. Over half, 51%, of MMM leaders prioritise improving data quality, against 36% of MMM laggards. That is a useful signal for a mid-market firm: the organisations getting the most from MMM invest in the plumbing before they invest in model sophistication. A simple model on clean, consistent data usually beats a sophisticated model on spend records that nobody has reconciled with finance.

Open-source does not mean uncontested. Different tools make different modelling choices: some use Bayesian estimation that expresses uncertainty directly, others use regularised regression with automated parameter search. Neither is inherently better for a mid-market firm, but they differ in how easily a non-statistician can interrogate the result. The selection criterion we would apply is explainability: can the marketing lead see why the model assigns a channel its return, and what would change that conclusion? A tool whose outputs cannot be explained to the finance director will not survive the first budget round.

What to do: decide at the outset who owns the model after the first build, and make that ownership a condition of any external engagement. Require the code, the assumptions and the data pipeline as deliverables, and require a working session in which your team re-runs the model without help. Open-source software is only an advantage if the knowledge stays inside the business.

The common objection is that mid-market firms lack the volume for a statistically valid model. The objection is partly right. A business with a few hundred wins a year, a long sales cycle and a handful of channels cannot support a granular, campaign-level model. It can, however, support a model at the level where budget decisions are actually made: monthly or weekly spend by channel, against a clean outcome series, over a reasonable history. The honest framing is that the model will produce wide ranges, and that wide ranges are still more useful than a single number that is wrong.

The inputs are less exotic than they sound. In our experience of how these models are specified, the essentials are weekly spend and impressions by channel, a consistent outcome measure such as qualified pipeline or booked meetings, and a record of everything else that moved demand: price changes, product launches, events, seasonality, sales-team changes and major external shocks. Our working rule of thumb, which is a judgement rather than a published standard, is that two to three years of weekly history gives the model enough variation to separate channels. Shorter histories need stronger priors.

Outcome choice is the decision that most affects usefulness in B2B. Revenue lags spend by months, so a model fitted to closed-won revenue will describe decisions taken a year ago. Qualified pipeline created, or sales-accepted opportunities, responds faster and is closer to what marketing controls. The cost of this choice is that pipeline quality must be stable, which requires the sales and marketing teams to agree a definition and keep it. That alignment is often a bigger obstacle than any statistical issue.

Seasonality and sales-team activity deserve special attention in B2B. Many mid-market firms have strong quarter-end effects, because deals close when budgets reset, and marketing spend often rises at the same moments. A model that does not control for these patterns will attribute the seasonal lift to whichever channel was heaviest at the time. The remedy is not exotic: include explicit seasonal terms and a record of sales headcount and territory changes, and test whether results are stable when those controls are added. A result that moves sharply with each control is telling you it is not yet reliable.

What to do: before touching software, assemble a one-page data inventory. For each channel, record where spend is held, how far back it goes and who can verify it. For the outcome series, record the definition and any changes. Gaps found at this stage cost days to fix; gaps found after modelling cost the credibility of the result.

A model fitted to observational data can explain the past in several equally plausible ways, particularly when channels move together. Experiments break the tie. EMARKETER and TransUnion data show that 52% of US brand and agency marketers already use incrementality testing, and that 36.2% plan to invest more in it over the next year. For a mid-market B2B firm, the practical forms are geographic holdouts, scheduled pauses of a channel in a region, and audience splits within a platform.

The barriers are known. Skai and the Path to Purchase Institute, surveying consumer goods brands, found that concerns about accuracy or reliability (44%), applying incrementality across ad types, targeting methods and retailers (43%) and limited tools or technology (41%) were the top obstacles. B2B marketers face the same problems with smaller samples. A test with too few conversions will produce a result that is statistically indistinguishable from zero and politically indistinguishable from failure, so power needs to be assessed before launch, not after.

The best use of experiments is to test the claims that matter most to the budget. Meridian's documentation describes converting experiment results into priors for the model, and recommends geo experiments to calibrate it. That closes the loop: the model identifies where uncertainty is greatest and money is at stake, the experiment resolves it, and the result tightens the next model. Over two or three cycles the combination becomes far more reliable than either component would have been alone.

The Gartner finding that paid media accounts for 30.6% of budgets gives a sense of scale. A typical mid-market team can run only a few experiments a year, so choosing them well matters. The highest-value tests are those where the budget is large, the model is uncertain and the decision is reversible. Testing a small channel that nobody would cut anyway wastes the opportunity. Testing the largest channel with a design that cannot detect a plausible effect wastes it just as thoroughly. A short pre-registration of the hypothesis, the duration and the minimum detectable lift keeps the discipline.

What to do: choose one experiment per quarter, aimed at the channel where the model's range of returns is widest and the spend is largest. Agree the decision rule in advance, for example that a lift below a stated threshold will reduce the budget, so that the result changes behaviour rather than prompting a debate about whether it counts.

A frequent source of disappointment is asking a model to do a job it was not built for. MMM works on aggregated time series. It can estimate the contribution of paid search, events or content in total, and it can describe diminishing returns as spend rises. It cannot tell a sales director which accounts a campaign influenced, nor whether a particular webinar outperformed another. For that layer, B2B firms need CRM data, buyer-reported attribution at the point of enquiry, and sales feedback.

Combining layers is therefore the sensible architecture, not a compromise. The model sets the envelope: roughly how much to put into each broad channel. Self-reported attribution, asking each new enquiry how they first heard of the firm, captures the dark-funnel activity that tracking misses. CRM analysis shows which segments and accounts convert. None is complete alone, and when they disagree the disagreement is a prompt for investigation rather than a reason to dismiss one source.

European firms should also consider the data environment. Privacy rules, consent requirements and browser restrictions have reduced the reach of user-level tracking, which is part of the reason aggregate methods have returned to favour. An MMM uses no personal data about individual users, only aggregated spend and outcome series, so it is comparatively robust to those restrictions. Even so, the input data should be handled under the firm's normal governance, and any customer-level extracts used for validation should be reviewed with the data protection lead.

Communicating uncertainty is a skill in its own right, and it determines whether the model is used. Present ranges beside point estimates, show the decision the range implies, and say which result would change the recommendation. Finance directors in particular respond well to this framing, because it resembles forecasting practice. Sales leaders respond to evidence from their own pipeline. Marketing leaders who bring both groups into the interpretation, before the numbers are final, find that the model becomes a shared reference point rather than a marketing document that others are invited to doubt.

What to do: write a one-page measurement charter that assigns each question to a method. Budget allocation across channels goes to the model. Incremental value of a specific tactic goes to an experiment. Account and segment performance goes to CRM. Early buyer sources go to self-reported attribution. When someone asks the model a question outside its remit, the charter names the right tool.

The size of the first commitment should reflect the uncertainty. A mid-market firm does not need a permanent measurement team to find out whether MMM is useful. It needs a time-boxed pilot with clear success criteria. The pilot should run a first model on existing data, report the ranges candidly, and identify the single decision most likely to be changed. If the model cannot change a decision at all, the firm has learned that cheaply.

Three build options exist, and each suits a different context. An in-house build using open-source tools gives control and transferable skills but needs a capable analyst. A specialist partner certified for the chosen tool gives speed and method discipline but requires the knowledge-transfer safeguards described above. A commercial platform reduces effort but increases cost and may limit transparency. Gartner's finding that 39% of CMOs plan to cut agency budgets suggests many will favour the first or second options, with outside help concentrated in the initial build.

Governance matters as much as method. A model that nobody is accountable for will drift as data sources change. We recommend a named owner in marketing, a finance counterpart who validates spend data, and a quarterly cadence for refreshing the model and reviewing experiment results. This takes a few days per quarter once established, and it is what turns a project into a capability.

The decision to scale deserves as much care as the decision to start. If the pilot works, the next investments are typically in automation of the data pipeline, a regular refresh cadence and a library of past experiments that new models can draw on. If the pilot shows that the data are too thin, the right response is to fix the recording of spend and outcomes for six months and try again, not to abandon measurement. In both cases the work of the first quarter has value, because it establishes how much the organisation actually knows.

What to do: set success criteria for day 90 in advance. Examples are a reconciled spend dataset, a model with documented assumptions, one experiment designed and under way, and at least one budget reallocation decision taken on the evidence. Then decide whether to scale, change approach or stop, and record the reasoning.

Karter McKinley helps mid-market firms through our AI and technology implementation service: auditing the measurement data, standing up an open-source model with documented assumptions, and designing the first experiments. We hand over code and method so your team owns the capability rather than renting it.