The Coalition for Innovative Media Measurement this week published a paper arguing that marketing mix modeling has moved from a niche analytical exercise into something closer to market infrastructure, and that the shift carries risks the industry has not yet addressed. Titled "Models as Masters? Marketing Mix Modeling and AI-Driven Media Decision Making," the report sets out six practical steps meant to shore up the data and documentation that private MMM systems rely on, without asking any advertiser to give up proprietary models or disclose sensitive business results.

The paper was co-authored by Chris Williams, an independent consultant and former vice president of digital at the Association of Canadian Advertisers, and Jon Watts, CIMM's managing director. It runs to 52 pages, including two appendices covering technical model components and a bibliography of prior industry and academic research. According to CIMM, the analysis draws on discussions with senior executives from companies including Meta, Google, Warner Bros Discovery, the Association of National Advertisers, and several independent measurement providers, though the paper states plainly that its interpretations and recommendations belong to the authors and CIMM rather than to the individuals consulted.

The timing matters. MMM has quietly become one of the most consequential measurement techniques in modern advertising, not because it is new, but because it has re-emerged at the exact moment the industry is trying to automate budget decisions with AI. CIMM's argument is that this convergence -- an aging statistical method meeting a new generation of automated buying systems -- creates a structural problem that has gone largely undiscussed: if the data feeding these models is inconsistent, and if AI systems act on that data without human interpretation, errors that used to be caught by an analyst can now be baked directly into how billions of dollars get allocated.

What the paper argues

According to the report, MMM was historically a retrospective tool. Analysts used it to explain what had already happened -- which channels seemed to drive sales, and by how much. That interpretive layer meant a human being was always positioned between the model's output and any decision based on it. CIMM's paper argues that this arrangement no longer describes how MMM is actually used.

Instead, the paper states that MMM increasingly generates forward-looking outputs: forecasts, budget allocations, response curves, and scenario plans that feed directly into planning systems, financial decision-making processes, and programmatic optimization workflows. Because those outputs are estimates rather than observed facts, and because different data inputs or modeling choices can produce materially different pictures of which channel is working, the paper contends that this matters more than it once did. When a model is used by an analyst, it can be challenged and cross-checked against other evidence. When the same model is wired into an automated system, it can influence real allocation decisions repeatedly and at scale, with far less opportunity for anyone to catch a mistake before it compounds.

The report identifies three interdependent sources of what it calls fragility in current MMM practice: inconsistent data across channels, interpretive instability in how models are built, and structural asymmetry in who controls the data feeding those models. On the data side, the paper points out that impressions, clicks, and gross rating points are not equivalent measures of exposure, yet MMM has to reconcile all of them into a single framework. Television and premium video present a particular case, according to the paper, because final delivery data for TV can arrive materially later than digital platform data, creating a structural disadvantage for slower channels even when they are delivering real value. As the report frames it, channels with faster and more structured data may simply appear more actionable within a model refresh cycle, whether or not that reflects their actual contribution to sales.

The AI amplification problem

The paper's central concern is not MMM in isolation but what happens when MMM becomes an input into automated decision systems. According to the report, in a human-mediated environment, imperfect data and uncertain model estimates are typically moderated by expert interpretation: an analyst contextualizes the output, challenges the assumptions, and brings in outside knowledge before anyone acts on it. In AI-driven systems, the paper argues, that layer of interpretation is reduced or removed altogether, because model outputs get operationalized directly through APIs into execution systems.

CIMM's report describes three mechanisms through which this amplifies existing weaknesses. The first is scale and speed: automated systems act continuously, so small biases or errors in a model's output can produce large cumulative effects once they are applied across millions of impressions, bids, or budget decisions. The second is what the paper calls feedback loops and recursive bias -- if a model over-allocates budget to a channel because of biased or incomplete inputs, the resulting performance data reflects that allocation, which can then reinforce the original bias over time and entrench a distorted pattern rather than correct it. The third is objective function misalignment: AI systems optimize toward whatever objective the available data supports, and where that data is incomplete -- inside a single platform or walled garden, for instance -- the system may end up prioritizing locally observable outcomes rather than genuine cross-media effectiveness.

The paper is careful to frame this as a risk that emerges from weak governance rather than an argument against automation itself. It states that properly governed AI can improve the MMM workflow by accelerating data integration, identifying anomalies, and helping decision-makers interpret uncertainty. The risk, according to the authors, is automation without sufficient controls over data quality, model assumptions, validation evidence, and human accountability.

Agentic buying raises new questions

A section of the report addresses the emerging protocols for agentic advertising, including AAMP and AdCP, which the paper describes as early-stage but strategically important because they define how AI agents discover inventory, exchange data, and execute buying instructions across advertising systems. CIMM frames the central unresolved question as what evidence buying agents should actually use when evaluating media value. If agentic systems default to deterministic conversion feeds or platform-specific APIs, the paper warns, they risk reproducing exactly the limitations that MMM was designed to correct in the first place: short attribution windows, weak treatment of offline outcomes, and limited visibility into cross-channel effects.

MMM can help close that gap, according to the report, because it incorporates cross-media exposure and long-term effects that closed-loop attribution typically misses. But the paper cautions that MMM-derived signals are probabilistic and model-dependent, often slower-moving than platform conversion data, and therefore need to be translated carefully into machine-readable form without losing the context and uncertainty that make them meaningful. The standards question, as CIMM frames it, is not whether agentic systems should use an MMM signal, but how future protocols distinguish between different classes of evidence -- deterministic conversions, experimental results, and MMM-derived response curves each carry a different evidentiary weight, and treating them interchangeably risks having probabilistic estimates mistaken for deterministic instructions.

The six recommendations

CIMM's paper sets out six practical steps, framed explicitly around shared inputs and evidence rather than any attempt to standardize advertiser-specific models or force disclosure of proprietary results.

The first calls for an MMM-ready media data specification covering television, streaming, connected TV, digital video, and related formats. According to the paper, the specification would define minimum fields for campaign identifiers, publisher and platform identifiers, media type, audience definition, reach, frequency, spend, timing, delivery status, and known limitations, with the goal of making media data usable across MMM workflows regardless of where it originates.

The second recommendation is a standard MMM model factsheet, a disclosure template that MMM providers, agencies, and internal analytics teams would use to document data sources, transformations, priors, adstock assumptions, validation methods, uncertainty ranges, and known limitations, without requiring disclosure of proprietary code or client-specific results.

Third, the paper proposes an experimentation and validation playbook covering geo tests, lift tests, calibration studies, sensitivity analysis, and out-of-sample validation, intended to specify when evidence is strong enough to inform model calibration versus when results should be treated as merely directional.

The fourth step calls for an independently overseen premium video evidence library synthesizing evidence on carryover effects, long-term brand effects, reach and frequency dynamics, co-viewing, and cross-channel synergies. According to the paper, the library should not create fixed priors that all models must adopt, but instead provide evidence ranges, confidence levels, and applicability notes.

Fifth, the paper recommends translating these materials into a contractual adoption toolkit -- sample RFP language, data provision requirements, audit rights, change-control expectations, and AI safeguards that could be embedded directly into commercial relationships between advertisers, agencies, media owners, and platforms.

The sixth step calls for engagement with agentic advertising protocol developers, so that MMM-informed evidence can be represented in machine-readable form without being treated as deterministic truth by automated buying systems.

Stakeholder positions vary sharply

The paper devotes a substantial section to how different parties in the ecosystem are positioned by these dynamics, and the differences are stark. Advertisers, according to the report, are the primary beneficiaries of MMM's evolution, gaining more comprehensive outcome-based views of performance, but they also carry new responsibility for reviewing supplier contracts around data rights, auditability, and portability if they switch MMM providers.

Media sellers -- particularly broadcasters and publishers -- face what the paper calls a structural asymmetry: they are expected to respond to model outputs without visibility into the underlying data, assumptions, or calibration that produced them. The report argues that the most credible response for sellers is not to build competing MMM systems or push for seller-controlled priors, but to contribute to independently governed evidence resources and transparent documentation practices instead.

Platforms occupy what the paper describes as a structurally complex position, operating simultaneously as media suppliers, data providers, and in some cases measurement providers themselves. Their access to granular, real-time data creates clear advantages in how models represent their channels, which the paper says does not necessarily imply intentional bias but does underscore the importance of independent validation.

Joint Industry Committees face what the report frames as a strategic question about their own relevance: rather than serving solely as currencies, JIC outputs increasingly function as inputs into modeling frameworks, and the paper asks whether these bodies evolve to support MMM through better data provision, build their own governed MMM solutions, or risk marginalization as modeling becomes the dominant mechanism of evaluation.

Context from the wider measurement debate

CIMM's paper lands amid a broader industry reckoning with how MMM tools are built and who controls them. A PPC Land analysis published on April 2, 2026, examined Meta's Robyn and the structural dynamic that emerges when a platform builds the tool guiding its own budget allocation -- a dynamic CIMM's paper addresses directly when it warns that data access advantages can become modeling advantages regardless of intent.

That concern has already surfaced in practical terms. When Amazon moved its Marketing Mix Modeling API to general availability in May 2026, the shift gave advertisers programmatic access to the retail giant's aggregated advertising and retail signals across 14 countries for the first time, illustrating exactly the kind of platform-controlled data pipeline the CIMM paper flags as a source of representation risk. Similarly, an IAB white paper published April 7, 2026, argued that standard MMM is structurally misaligned with retail media measurement, causing systematic undervaluation of commerce channels -- a finding that echoes CIMM's broader argument that channels less fully represented in default data pipelines risk being under-valued regardless of their true contribution.

The paper's concerns about television specifically connect to reporting on Barb's Ads Hub reaching 600 users in June 2026, where PPC Land noted that the difficulty of capturing linear TV's contribution inside standard MMM frameworks has been a recurring source of frustration for marketing teams, with linear's effectiveness routinely underestimated in models that rely heavily on digital signals.

On the open-source side, Google's Meridian and Meta's Robyn have continued to expand their footprint through 2025 and 2026. Google opened Meridian to all marketers globally in January 2025 after testing with hundreds of brands, then updated the tool in September 2025 to incorporate non-media variables such as pricing and promotions alongside channel-level contribution priors. The growing reach of these tools is precisely what CIMM's paper points to when it argues that default priors and embedded assumptions in widely adopted open-source frameworks can reconfigure data asymmetry rather than eliminate it, since channels more fully represented in a framework's built-in data integrations may be modeled with more confidence regardless of their true incremental contribution.

The agentic buying protocols discussed in CIMM's sixth recommendation have moved quickly since their initial launch. AdCP launched on October 15, 2025, founded by Scope3, Yahoo, PubMatic, Swivel, Triton Digital, and Optable, and built on Anthropic's Model Context Protocol. The IAB Tech Lab formally named its own umbrella initiative AAMP on February 26, 2026, providing a three-tier agent hierarchy for buyers and sellers. Most recently, IAB Tech Lab released AAMP version 2.3, adding a pricing provenance field designed to stop AI buying agents from fabricating bid prices when real market data is unavailable -- a concrete illustration of the kind of evidentiary discipline CIMM's paper argues the industry still needs to build around MMM signals specifically.

Adoption of MMM itself has continued climbing. Research covered by PPC Land in November 2025 found that 46.9 percent of marketers planned to increase investment in marketing mix modeling over the following 12 months, representing the highest investment priority among measurement methodologies surveyed, while 52 percent reported using incrementality testing and experiments alongside it. That same coverage noted an IAB Australia vendor landscape from September 2025 profiling twelve separate MMM providers, underscoring how crowded and fragmented the provider ecosystem CIMM's paper is trying to bring shared standards to has already become.

Why this matters for the industry

For advertisers, agencies, and media owners, the practical stakes are straightforward even if the underlying statistics are not. Budget allocation decisions worth billions of dollars increasingly rest on model outputs that few people outside a data science team can fully interrogate. CIMM's paper does not claim MMM is broken or unreliable; it claims that the conditions under which MMM outputs are produced, documented, and validated have not kept pace with how much weight those outputs now carry. As the paper puts it, the future of MMM will be determined less by advances in modeling technique than by the strength of the data, standards, and institutions that support it.

That framing matters because it reorients the conversation away from a familiar and largely unproductive argument over which vendor's algorithm is best, and toward a question the industry has more control over: whether the inputs feeding every vendor's model are complete, consistent, and documented well enough to trust. The paper is explicit that this agenda will not resolve on its own. Incentives across the ecosystem are not aligned -- platforms, media owners, advertisers, agencies, and measurement providers do not all benefit equally from more transparency or standardization -- and the costs of improvement, such as data standardization and audit capability, fall unevenly across large platforms and smaller vendors alike. No single institution currently has the mandate or funding to coordinate the full MMM ecosystem, which is why the paper explicitly proposes voluntary alignment and staged, testable steps rather than a comprehensive governance regime.

For premium video and television specifically, the paper's argument carries a particular urgency. If TV and video inputs remain fragmented, delayed, or inconsistently classified, the report warns, their contribution may become harder for both MMM systems and the buying agents built on top of them to detect, validate, and act upon -- not because the medium lacks value, but because the data trail documenting that value cannot keep pace with how quickly automated systems now need to act on it.

Timeline

  • January 29, 2025 - Google opens its Meridian marketing mix model to all marketers and data scientists globally following testing with hundreds of brands.
  • March 13, 2025 - ANA Aquila LLC announces progress on its cross-media measurement initiative, including a calibration panel launched with Kantar Media in 1,000-plus homes.
  • July 15, 2025 - Prescient AI launches a marketing mix model built from scratch, positioning itself against open-source frameworks built on Meta's Robyn and Google's Meridian.
  • September 30, 2025 - Google updates Meridian to include non-media variables such as pricing and promotions alongside channel-level contribution priors.
  • October 15, 2025 - The Ad Context Protocol launches, founded by Scope3, Yahoo, PubMatic, Swivel, Triton Digital, and Optable, built on Anthropic's Model Context Protocol.
  • October 21, 2025 - Aquila partners with Samba TV to add streaming viewership data to its independent cross-media measurement platform.
  • November 5, 2025 - Industry research finds 46.9 percent of marketers plan to increase marketing mix modeling investment over the following year, the highest priority among measurement methods surveyed.
  • February 26, 2026 - IAB Tech Lab formally names its Agentic Advertising Management Protocols, providing a three-tier agent hierarchy for buyers and sellers.
  • April 2, 2026 - A PPC Land analysis examines the structural dynamic when a platform builds the MMM tool that guides its own budget allocation, using Meta's Robyn as a case study.
  • April 7, 2026 - The IAB publishes a white paper arguing that standard MMM is structurally misaligned with retail media measurement, causing systematic undervaluation of commerce channels.
  • May 2026 - Amazon moves its Marketing Mix Modeling API to general availability across 14 countries.
  • June 8, 2026 - Coverage of Barb's Ads Hub notes that linear TV's contribution is routinely underestimated inside standard MMM frameworks that rely on digital signals.
  • July 2026 - CIMM publishes "Models as Masters? Marketing Mix Modeling and AI-Driven Media Decision Making," co-authored by Chris Williams and Jon Watts.
  • July 31, 2026 - DBC Agency circulates CIMM's press release announcing the paper and its six recommendations to press contacts including PPC Land.

Summary

Who: The Coalition for Innovative Media Measurement, an industry group whose members span the media and advertising ecosystem, published the paper. It was co-authored by independent consultant Chris Williams and CIMM managing director Jon Watts, with input from executives at companies including Meta, Google, and Warner Bros Discovery.

What: A 52-page report arguing that marketing mix modeling has evolved from a retrospective analytical technique into embedded decision infrastructure, and warning that inconsistent data, opaque assumptions, and weak validation could be amplified when MMM outputs feed AI-driven and agentic buying systems. The paper proposes six practical steps -- a media data specification, a model factsheet, a validation playbook, a premium video evidence library, a contractual adoption toolkit, and engagement with agentic protocol developers -- aimed at improving shared inputs without standardizing proprietary models.

When: The paper was published in July 2026, with the accompanying press release circulated to media contacts on July 31, 2026.

Where: CIMM is based in the United States, and the paper addresses the global marketing measurement ecosystem, with particular emphasis on the television and video marketplace.

Why: As AI and agentic buying systems increasingly act on MMM outputs with less human interpretation between the model and the decision, the paper argues that weaknesses once confined to analysis can now be operationalized directly into how billions of dollars in media capital are allocated – making the quality, documentation, and validation of MMM's underlying inputs a matter of shared market importance rather than a narrow technical concern.