--- title: "From Meridian to NNN: How Transformers Are Redefining Marketing Mix Modeling" date: 2025-04-21 summary: "Google's NNN brings Transformer self-attention to Marketing Mix Modeling — moving past Meridian's parametric assumptions to learn temporal dependencies, cross-channel synergies, and creative nuance directly from aggregated data. ~22% lift in predictive accuracy on early benchmarks." standfirst: "Google's NNN replaces Meridian's fixed Adstock and Hill curves with Transformer self-attention, learning temporal lags, cross-channel synergies, and creative nuance straight from aggregated data rather than assuming them. Early benchmarks show roughly 22% better predictive accuracy on about 5x less data. For advertisers in a post-cookie world, it turns MMM from a backward-looking ROI estimate into a privacy-safe simulator for budget and scenario planning." canonical: https://nofluffadvisory.com/writing/from-meridian-to-nnn-how-transformers-are-redefining-marketing-mix-modeling/ --- In the fast-evolving world of digital advertising, understanding the true impact of marketing efforts has never been more critical — or more challenging. **Marketing Mix Modeling (MMM)** has long been a cornerstone for advertisers seeking to quantify how channels like TV, digital, and social drive sales. In 2024, Google's **Meridian** MMM brought Bayesian rigor to traditional regression-based models, offering robust ROI estimates. Now, in April 2025, Google's **NNN (Next-Generation Neural Networks)** takes MMM into a new era — leveraging Transformer-based neural networks to deliver unprecedented attribution accuracy and actionable insights. For advertising professionals, NNN signals a transformative shift in how we measure and optimize media in a privacy-first, post-cookie world. ## The evolution from Meridian to NNN Meridian MMM, launched in 2024, was a leap forward for traditional MMM. It used scalar inputs like weekly spend or impressions, applying parametric functions — **Adstock** for lagged effects and **Hill** for saturation — to model channel impact. Bayesian Markov Chain Monte Carlo (MCMC) provided confidence intervals, making Meridian a reliable tool for estimating ROI. However, its limitations were clear: it struggled with granular creative nuances, cross-channel synergies, and long-term effects, often assuming uniform decay patterns across campaigns. Enter NNN — Google's Transformer-based MMM model introduced in 2025. Unlike Meridian, NNN uses high-dimensional **[embeddings](/writing/embeddings-the-next-frontier-in-advertising/)** to represent marketing activities, blending quantitative data (spend, impressions) with qualitative factors like ad-creative attributes or search-query types. Its Transformer architecture, powered by self-attention and multi-head attention, dynamically learns temporal dependencies and channel interactions, offering a richer, more accurate picture of marketing impact. > Early benchmarks show NNN outperforming traditional MMM by **~22% in predictive accuracy** — a game-changer for advertisers seeking precision. ### Meridian vs NNN at a glance | Aspect | Meridian MMM | NNN Neural MMM | |-------------------------|-------------------------------|---------------------------------------------| | **Inputs** | Scalar spend / impressions | Embedded qualitative and quantitative inputs | | **Temporal dynamics** | Parametric (Adstock) | Learned via self-attention | | **Attribution** | Regression coefficients | Non-linear attribution via causal Transformer | | **Handling interactions** | Limited (manual terms) | Dynamic via multi-head attention | | **Creative sensitivity** | None | Embedded creative impact | ## Why Transformers matter for MMM The power of NNN lies in its Transformer mechanisms, which address longstanding MMM challenges: - **Self-attention for temporal dynamics.** Traditional MMM relies on fixed decay curves to model carryover effects. NNN's self-attention allows it to "attend" to any past time period, learning complex lag patterns — like a TV campaign from three months ago driving today's sales — without rigid assumptions. This flexibility captures both short-term bursts and long-term brand effects, critical for holistic measurement. - **Multi-head attention for interactions.** NNN's multi-head attention processes multiple influence patterns simultaneously. One head might focus on immediate digital ad impacts, another on TV-driven search uplifts. This enables NNN to quantify synergies — such as social media amplifying display-ad performance — which Meridian often missed. - **Rich embeddings for creative nuance.** NNN encodes qualitative factors like ad-creative tone or keyword intent into embeddings. This allows the model to differentiate, say, an emotional video ad from a comedic one, or branded search queries from generic ones. For advertisers, this means granular insights into what drives performance *within* a channel. These mechanisms make NNN a dynamic, data-driven tool that captures the complexity of modern marketing ecosystems — from cross-channel customer journeys to creative effectiveness. ![NNN Transformer architecture — multi-channel input tensor passes through a target channel mask, then n× stacked blocks of factored & causal NNN self-attention plus channel-wise MLPs, into channel-specific prediction heads producing the multi-channel output tensor.](/essays/from-meridian-to-nnn-architecture.png) ## Taming complexity with regularization Deep-learning models like NNN thrive on large datasets, but MMM often works with limited aggregate data — think a few years of weekly sales and media metrics. To prevent overfitting, NNN employs **L1 regularization,** which prunes less-impactful features, creating a sparser, more generalizable model. This not only ensures NNN performs well with smaller datasets (reportedly requiring **~5× less data** than traditional MMM) but also enhances interpretability by highlighting key drivers. For advertisers, this means reliable insights even for new campaigns or niche markets, without needing years of historical data. ## Actionable insights and scenario planning NNN isn't just a measurement tool — it's a marketing **simulator** that empowers strategic decision-making. Its interpretability features make complex outputs accessible: - **Attention visualization.** By analyzing attention weights, marketers can see which past campaigns or channels drive current performance. High attention on last quarter's TV spend might reveal strong carryover effects, guiding budget allocation. - **Creative-effectiveness analysis.** NNN lets you swap creative embeddings while holding spend constant — simulating how different ads impact sales. This helps identify high-performing creatives or keywords without costly A/B tests. - **Scenario simulation.** NNN supports *"what-if"* planning, like projecting sales if digital spend doubles and TV is cut. A "marketing pause" scenario — setting all media to zero — reveals baseline sales and decay rates, showing how long past investments sustain performance. These simulations enable bold, data-driven strategies. ## Navigating a privacy-first future NNN arrives at a pivotal moment. With cookies phasing out and privacy regulations like GDPR and CCPA tightening, user-level tracking is fading. MMM, with its aggregated approach, is regaining prominence as a privacy-safe solution. NNN enhances this by incorporating first-party data (e.g., CRM segments) via embeddings, blending it with media metrics to deliver granular insights without personal identifiers. This aligns perfectly with the industry's shift toward holistic, privacy-compliant measurement. Retail media is converging on the same discipline — [incrementality as the IAB/MRC retail media guidelines define it](/standards/retail-commerce-media-measurement/) is the causal bar MMM outputs increasingly get held to. Moreover, NNN's potential for continuous learning — updating with weekly data — could blur the line between periodic MMM projects and real-time optimization. This agility is crucial as marketing cycles accelerate and walled gardens limit cross-channel visibility. ## What this means for advertisers For advertising professionals, NNN is a call to evolve. It offers: - **Holistic optimization.** By modeling cross-channel synergies and creative impacts, NNN enables true budget optimization across online and offline media, even in fragmented ecosystems. - **Confidence in experimentation.** With accurate attribution, marketers can test new creatives or channel mixes — knowing NNN will capture the effects reliably. - **Collaborative analytics.** NNN's complexity demands teamwork. Data scientists tune the model, while strategists translate outputs into campaigns. This collaboration will define next-gen analytics teams. ## The future of marketing measurement NNN marks a new chapter where AI and marketing science converge. It builds on Meridian's econometric roots — preserving MMM's practicality while unlocking deep learning's flexibility. By capturing long-term effects, creative nuances, and channel interactions within one framework, NNN offers advertisers a clearer, more actionable view of their media mix. As we navigate a privacy-first landscape, NNN's ability to deliver precise, interpretable insights from aggregated data positions it as a cornerstone for future measurement. The journey from Meridian to NNN shows that innovation can honor tradition while pushing boundaries — ensuring advertisers can optimize with confidence, no matter how the digital ecosystem evolves. --- ## References - Google. *"NNN: Next-Generation Neural Networks for Marketing Mix Modeling."* April 2025. - Google. *"Meridian: Open Source Marketing Mix Modeling."* 2024. - Vaswani et al. *"Attention Is All You Need."* 2017. - Digital Advertising Industry Reports. *"The Role of MMM in the Post-Cookie World."* 2024. - Evgeny Popov. *"Understanding Regression-Based Attribution (RBA) and Meridian MMM."* Medium, 2024. - Evgeny Popov. *"Revolutionizing AI: How the Transformer Model is Redefining Data Processing."* Medium, 2024.