Marketing mix modeling spent two decades being characterised as either a static legacy technique or a forward-looking analytics discipline depending on who was selling it. Through 2023-2024 the practice underwent something close to a rehabilitation, driven by the cookie-deprecation reality, the maturity of available tooling, and the increasing recognition that platform-attributed conversions weren’t telling the whole story. The May 2026 picture is one of MMM as established discipline rather than recently-rediscovered technique, and the practice has matured in important ways.

The headline observation is that MMM in 2026 is genuinely different from MMM as practised a decade ago. The analytical methodologies have advanced. The integration with other measurement approaches (incrementality testing in particular) has produced more credible answers. The data infrastructure required has shifted from monthly aggregated spend data to weekly or daily granularity with proper audience and creative variables included. The output isn’t a quarterly slide deck; it’s an ongoing analytical capability that supports marketing decisions in close to real time.

This isn’t true everywhere. The practice quality varies enormously across advertisers, and the gap between the better and worse implementations is wider than ever. The advertisers running mature MMM practices are getting genuinely useful answers. The advertisers running MMM as a periodic compliance exercise are getting answers that look reasonable on the surface but don’t survive scrutiny.

What’s working in mature MMM practice

Bayesian frameworks have largely displaced the older regression-based approaches. The integration of prior beliefs (from prior tests, prior models, industry knowledge) with new data produces more stable and more credible estimates than the regression approaches that dominated the 2010s practice. The major platforms supporting Bayesian MMM at scale — Robyn, Meridian, the various commercial vendor products — have all converged on broadly similar mathematical approaches, with the differentiation being in data integration, ease of use, and ongoing operational support.

The integration of MMM with controlled experimentation produces substantially better results than either practice in isolation. MMM provides the breadth of decision support across the full channel mix; experiments provide the causal grounding for specific assumptions. Advertisers running both disciplines, with feedback loops between them, produce more credible models than advertisers running either in isolation.

The use of MMM for budget allocation rather than just attribution has matured. The shift from “explaining what happened” to “informing what to do next” has changed how the work is consumed by marketing decision-makers. The output of a mature MMM practice is now typically a recommended budget allocation across channels and time, with associated uncertainty bounds, rather than a retrospective decomposition of past results. This is more useful and more demanding to produce well.

The treatment of long-term effects has improved. The recognition that brand-building investments produce returns over multi-quarter horizons rather than within campaign measurement windows has been incorporated into the better commercial models. Adstock parameters, decay rates, and saturation curves are now standard model components in ways they weren’t a decade ago. The implications for brand investment justification are real: advertisers running MMM that captures long-term effects can defend brand budgets in ways that advertisers running short-term-only models can’t.

What’s still hard

Granularity. Aggregate-level MMM produces useful answers about what’s working at the channel and overall budget level. Granular MMM that addresses creative effectiveness, audience-level efficiency, and campaign-level returns is harder to produce credibly. The data requirements are higher and the modeling complexity is higher, and the answers are often less stable than the aggregate-level models.

External effects. The macro-environment effects on consumer behaviour — economic conditions, weather, competitor activity, news cycles, holidays — need to be in the model in some form. The variable selection and operationalisation for these effects is genuinely hard. The models that capture them well outperform the models that don’t, but capturing them well requires both data infrastructure and analytical judgement that not all teams have.

Integration with operational marketing systems. The ideal state is MMM output flowing directly into media planning and bidding decisions. The reality for most advertisers is MMM output flowing into quarterly planning conversations, with operational systems still optimising on shorter-term signals. The integration work to close this gap is substantial and is incomplete in most organisations.

Talent. The skill set required to produce credible MMM combines statistical modeling depth, marketing domain expertise, data engineering capability, and the communication skills to translate technical output into decision support. People with all four are scarce, and the workforce-development pipeline for this combination has been slow to mature.

What advertisers are paying for

The commercial MMM market has stratified clearly:

At the high end, the large analytics consultancies provide bespoke MMM services with substantial human analytical input. The annual cost is meaningful — six figures and up for serious advertisers. The output quality is generally high but the dependence on specific analyst relationships is also real.

In the middle, software-as-a-service MMM platforms (Meridian, Robyn-as-a-service implementations, several specialist vendors) provide structured tooling that produces models with less customisation but at lower ongoing cost. The output quality depends substantially on the quality of the data fed into the models and the practitioner running them.

At the entry level, internal capability development using open-source tooling like Robyn has become viable for advertisers willing to invest in internal capability. The startup cost is real but the ongoing cost is lower and the institutional knowledge stays in-house. The quality of output depends on the strength of the internal team.

The choice between these models depends on the size of the advertiser, the complexity of the marketing programme, the available internal capability, and the tolerance for vendor dependence. There isn’t a single right answer.

What I’d avoid

Three patterns that produce bad MMM outcomes:

Building MMM models on inadequate data. The data requirements for credible MMM are substantial — multiple years of weekly or daily data, granular spend and impression data by channel, control variables for external effects, and outcome data with proper attribution to time periods. Advertisers trying to produce models on inadequate data routinely produce models that don’t survive validation.

Treating model output as deterministic. MMM produces estimates with uncertainty. The decisions that flow from the models should incorporate this uncertainty. Marketing teams that treat the model output as “the answer” rather than as “the best estimate with associated uncertainty” make worse decisions than teams that work with the uncertainty appropriately.

Failing to validate against external evidence. MMM models that produce answers wildly inconsistent with controlled experiments, with platform-reported behaviour, or with historical performance need to be questioned. The triangulation between MMM and other evidence sources is part of producing credible answers, and the MMM that doesn’t get triangulated is often less reliable than the practitioner believes.

What I’d watch

Three things over the next 12 months.

The integration of AI assistance into MMM workflows. The use of language model agents for model interpretation, decision recommendation, and stakeholder communication is advancing. The capability is genuine. The risk of over-confidence in AI-generated analysis is real. The discipline around how AI is integrated into the work matters.

The continued evolution of platform-provided MMM (Meta’s Robyn, Google’s Meridian). These first-party tools provide accessible MMM capability but raise questions about the platforms scoring their own homework. The practitioner discipline in using these tools, alongside other measurement evidence, will matter for the credibility of the answers they produce.

The regulatory environment around third-party data use. The Australian Privacy Act reforms and the broader international consent framework changes continue to evolve. The implications for the data inputs that MMM relies on are real. Models built on data that may not be available in equivalent form in three years’ time may produce results that are less projectable than they appear.

The honest summary for mid-2026: MMM is a serious discipline and a useful one for advertisers willing to invest in doing it well. The renaissance of the past few years has produced genuine capability uplift across the industry. The work continues to mature. The gap between the best practitioners and the average practitioners is wide, and closing that gap is the productive direction for the next phase of the discipline.