Chapter 18 AI Models in Marketing Analytics

18.1 Introduction

Marketing analytics has used statistical and machine-learning models for decades. Response models, customer segmentation, propensity scores, and marketing mix models were standard well before large language models existed. What has changed is the range of tasks that models can now take on. Predictive models score audiences. Causal models estimate what spend actually caused. Generative models write copy, summarize research, and turn questions into SQL. Agents chain these steps together and call governed tools.

This chapter maps where each kind of model fits in day-to-day marketing analytics work, especially in an agency setting that serves many clients, markets, and data sources. It ties the earlier chapters to the domain: prompting and verification (Chapters 8 and 14), MCP (Chapter 16), and context and harness engineering (Chapter 17).

Two principles run through the chapter:

Use the right model family for the question. A language model is not a substitute for a causal design, and a propensity model does not estimate incrementality. Many costly mistakes in marketing analytics come from answering a causal question with a predictive tool.
The model chooses and explains; governed engines compute. Generative AI is most useful as an interface to validated analyses, not as the source of the numbers.

Measurement theory (marketing mix modeling, attribution, experimentation, and segmentation) is covered in depth in the companion Marketing Analytics book. This chapter focuses on how AI models are used across that work.

18.2 A map of AI model families in marketing

Four families of models solve different kinds of problems: predictive (classification, regression, ranking), causal (experiments, quasi-experiments, econometric models), generative (language, image, and multimodal models), and agentic systems that combine models with tools. For each, this section will cover the typical marketing questions it answers, the data it needs, and the question it cannot answer on its own.

18.3 Audience and customer modeling

This section covers segmentation, lookalike and propensity modeling, customer lifetime value, and churn. It will also discuss where foundation models and embeddings add value, for example by turning unstructured behavioral or text data into features. The recurring risks are leakage from post-outcome features, seed-audience bias, and treating a high propensity score as proof of incrementality. Chapter 19 treats lookalike modeling as positive-unlabeled learning and covers two-tower embeddings and typed decision models in depth.

18.4 Measurement, attribution, and incrementality

This section explains how AI fits around measurement without replacing causal identification. Topics include marketing mix models (for example the open-source Meridian and Robyn projects), multi-touch attribution and its limits, geo and holdout experiments, and how language models can help explain results to clients while the estimates themselves come from validated models.

18.5 Media planning, reach, and audience overlap

This section covers budget allocation and scenario planning, reach and frequency estimation, and audience overlap across platforms and segments. It discusses optimization on top of response curves and how an assistant can expose these analyses as governed tools rather than ad hoc calculations.

18.6 Basket analysis and recommendations

This section covers market basket analysis (support, confidence, and lift), product affinity, and recommendation systems. It also covers embedding-based approaches and how to explain affinity results in plain language without overstating causality.

18.7 Generative AI for content and creative

This section covers copy and creative variation, localization across markets, brief generation, and creative pre-testing. It addresses brand safety, factual claims in ads, intellectual-property and disclosure considerations, and why creative variants still need a proper experiment to show which one works.

18.9 AI-assisted analytics workflows

This section covers analytics copilots and text-to-SQL assistants (such as Databricks Genie), automated reporting and insight narration, and Streamlit apps built on foundation models. It applies the metadata-first, numbers-from- the-engine design from Chapter 17.

18.10 Agents in marketing operations

This section covers multi-step agents that monitor campaign performance, detect anomalies, prepare client reports, or trigger reusable jobs such as lookalike builds. It addresses permissions, human approval for actions that spend or publish, and auditability.

18.11 Data foundations and privacy

This section covers first-party data, identity resolution, data clean rooms, consent, and regional privacy regulation. It explains why aggregation, minimum cell sizes, and per-user authorization matter more once client data can flow into model context and logs.

18.12 Evaluation and governance

This section covers how to evaluate each model family: offline metrics for predictive models, validity checks for causal estimates, human-validated benchmarks for generative tasks, and repeated-run reliability (pass^k) for assistants and agents. It also covers fairness in targeting, monitoring for drift, and documenting model limits for clients.

18.13 Interview takeaways

This section will summarize how to explain these choices in an interview: which model family fits a stated business question, how to defend a measurement approach, and how to describe the safe use of generative AI on client data.

18.14 Key takeaways

Marketing analytics uses four model families: predictive, causal, generative, and agentic. Match the family to the question.
Predictive accuracy is not incrementality; causal questions need causal designs.
Generative AI is strongest as an interface to validated analyses and as a tool for unstructured text and creative work.
Governed metadata, deterministic tools, privacy controls, and evaluation turn AI prototypes into dependable client-facing products.