A Marketer’s Field Guide to Machine Learning

How to match the right machine learning algorithm to the right marketing problem, from linear regression to large language models.

A field guide to machine learning for marketers, organized from simpler, more explainable models to larger, less explainable ones.

Why a Field Guide

The top of my LinkedIn profile says, “Energetic & experienced executive with a passion for BETTER MARKETING. I use data science & machine learning to improve marketing.”

Although my colleagues at Georgia Tech know what machine learning is, many of my marketing and business colleagues don’t, and they ask me about it all the time. That’s the motivation for this article: to explore what machine learning is and how it can help marketing. It’s one of my longer pieces and I’ve been working on it for a while. Hopefully it’s a resource you can keep coming back to.

Figure 1. The Sasquatch Field Guide. Unlike Sasquatch, the right machine learning model for your marketing problem actually exists. This field guide will help you find it. Yes, I really do own the Sasquatch Field Guide and that hat. Please don’t judge me.

If you work in marketing, you’re probably already using machine learning even if you don’t know it. The ad platforms use it. CRM systems use it. Email platforms use it. Recommendation engines on e-commerce websites use it. Lookalike audiences are built on it. Lead scoring systems run on it. Your analytics dashboard is using it too.

The question is no longer whether to use machine learning in marketing. You already are. The question is which algorithm fits which job, and how to know when you are reaching for the wrong tool.

This article is a field guide. The goal is to help you recognize the major machine learning algorithms when you encounter them in marketing work, understand roughly how each one operates, and know what kind of marketing problem each is suited to. For any algorithm you want to learn in depth, I have written individual articles on most of them, and I link to those throughout.

Last year I wrote Marketing Model Selection: There’s No Need to Throw AI at Everything (Domaleski, 25 May 2025). The argument in that piece was that the right model for a marketing problem is usually not the most sophisticated one. It is the one that answers the question with the least friction. This article extends that argument by giving you the menu to choose from if the simple model isn’t enough.

What “Better Marketing” Actually Means

I should ground this article in what I mean by better marketing, because the whole article is about using machine learning for better marketing.

In A Better Marketing Manifesto (Domaleski, 8 Sept. 2024), I laid out ten principles that I think define better marketing. Know your audience. Create value. Tell a story. Be authentic. Maintain consistency. Engage and listen. Ground your strategy in data. Respect privacy. Adapt and innovate. And above all, focus on doing good. Better marketing is also measurable, which is why I followed the manifesto with Stop Confusing Marketing with Sales, They’re Not the Same Thing (Domaleski, 17 Nov. 2024). Better marketing claims credit honestly. It produces decisions, not dashboards.

Figure 2. A Better Marketing Manifesto, originally published September 2024. Machine learning should serve these principles, not replace them. Source/Country Fried Creative.

Machine learning is a means to those ends. It is not an end itself. A logistic regression that helps you score leads accurately is better marketing. A neural network that nobody on the team can interpret or defend is not. Of the ten principles in the manifesto, “ground your strategy in data” is the one most directly served by machine learning, but every other principle still has to hold. A model that violates privacy, that produces inauthentic content, or that drifts away from your brand story is not better marketing. It is just more marketing.

Throughout this article, the question that should guide every algorithm choice is whether the model helps you make a better marketing decision while staying true to the principles above.

If it does not, you do not need it.

A Quick Refresher on What Machine Learning Actually Is

Machine learning (ML) is a set of methods for letting a computer find patterns in data and use those patterns to make predictions or decisions. That is the working definition.

It helps to place machine learning inside the bigger picture. Artificial intelligence is the broad umbrella, any technique that gets a computer to do something we would call intelligent. Machine learning is the subset of AI that learns patterns from data rather than following hand-written rules. Deep learning is a further subset of machine learning that uses neural networks with many layers. And generative AI, including the large language models everyone is talking about, sits inside deep learning. When people say “AI” today they usually mean that innermost layer, generative AI like ChatGPT. This field guide spends most of its time in the wider machine learning category, where the simpler and more explainable methods live.

Machine learning overlaps with traditional statistics, but the emphasis is different. Classical statistics is largely concerned with explanation and hypothesis testing. Machine learning is largely concerned with prediction. The two share most of the same mathematical foundations, and the line between them is fuzzy, but the goals diverge often enough that it is useful to think of machine learning as the prediction-focused sibling. One thing is worth mentioning up front: most machine learning models learn correlations, not causes. That distinction may seem small now, but it turns out to be one of the most important ideas in this guide, and I come back to it near the end.

There are three broad flavors of machine learning you should know about.

Figure 3: The three broad flavors of machine learning. Most marketing analytics work falls into supervised or unsupervised learning. Reinforcement learning shows up mostly inside ad platforms.

Supervised learning uses labeled examples. You give the model a set of inputs along with the correct outputs, and the model learns to predict the output on new inputs it has not seen. Predicting whether a lead will convert is supervised learning. So is forecasting next month’s revenue.

Unsupervised learning has no labels. The model is given a pile of data and asked to find structure in it on its own. Customer segmentation without predefined groups is unsupervised learning. So is most anomaly detection.

Reinforcement learning is the third flavor. The model learns by doing, receiving feedback in the form of rewards or penalties, and adjusting its behavior over time. It shows up in marketing mostly inside the ad platforms, where bid optimization runs on reinforcement principles. For day-to-day marketing analytics work, you will encounter supervised and unsupervised methods far more often.

Every algorithm in the guide is a different answer to the same underlying question. What kind of pattern am I trying to find in my marketing data, and what is the cleanest way to find it?

The Map of Machine Learning Algorithms

The algorithms in this guide are roughly ordered by complexity. We start with linear regression, the foundation that almost every other method generalizes from in one way or another. We move through logistic regression, decision trees and random forests, k-means clustering, time series models, recommendation systems, gradient boosting, and natural language processing. We end with neural networks and large language models, the most capable and the least interpretable. A short final section covers a few other methods worth recognizing.

Figure 4: The ladder of machine learning algorithms covered in this article. Climbing higher gives you more capability and less interpretability. The point is not to climb to the top, but to stop at the rung that solves your problem.

The ordering in Figure 4 is intentional. Each step up the ladder gives you more flexibility in the patterns you can capture. Each step also costs you something in interpretability, in data requirements, in compute, or in your ability to explain the result to the person who has to act on it. The point of the field guide is not to push you up the ladder. The point is to help you stop at the rung that solves your problem.

The sections that follow walk through each algorithm one at a time. For each, I cover what it is in plain language, how it works mechanically, and one or more marketing examples where it fits. Where I have written a deep dive on the topic, I link to it.

Linear Regression

What it is. Linear regression is the workhorse of marketing analytics. It fits a straight line, or a flat plane in higher dimensions, through your data in a way that minimizes the squared distance between the line and the actual data points. The line is your model. Once you have it, you can plug in a new input and read off a predicted output. Linear regression predates the term “machine learning” by more than 150 years, and it remains one of the most widely studied and used algorithms for prediction and correlation analysis today.

How it works. The math has been around since Gauss in the early 1800s. You probably remember the equation for a line y = mx + b from middle school. That’s linear regression at its simplest. You start with an outcome you want to predict, like revenue. You have one or more inputs you think drive that outcome, like marketing spend, seasonality, or price. The algorithm finds the coefficients that, when multiplied by the inputs and summed, produce the predicted value closest to the actual observed value across your data. The “closest” part is defined by the sum of squared errors. The coefficients themselves tell you the marginal effect of each input. A one-dollar increase in spend on this channel produces an estimated X dollars in revenue, holding everything else constant.

Figure 5: Linear regression fits a line through the data by minimizing the sum of squared residuals, the vertical distances between each observed point and the predicted line.

Marketing examples. Sales forecasting is the classic use case. Customer lifetime value estimation is another. The whole family of Marketing Mix Models, which I covered at length in An Introduction to Marketing Mix Modeling (Domaleski, 8 Mar. 2026), are essentially constrained linear regressions with adstock and saturation transformations layered on top. I have also written a head-to-head comparison of linear regression against XGBoost on a sales prediction task in Linear Regression vs. XGBoost: Which Predicts Sales Better (Domaleski, 1 June 2025), and the short version is that linear regression holds its own surprisingly well.

If you want a deeper walk through the algorithm itself, I covered it in Linear Regression: The Classic Machine Learning Algorithm You Need to Know (Domaleski, 1 Dec. 2024).

Logistic Regression

What it is. Logistic regression is the sibling of linear regression for yes-or-no outcomes. Instead of predicting a continuous number, it predicts a probability between zero and one. That probability can then be turned into a binary decision by setting a threshold, usually fifty percent.

How it works. The trick is the logistic (sigmoid) function. It takes any number and squeezes it into a value between zero and one, which gives you a probability. The model starts the same way as linear regression. First, you take a weighted sum of your inputs. Next, you run that sum through the sigmoid function and out comes a probability. Figure 6 shows the S-curve, with the dashed line marking the classification threshold. The coefficients are found using maximum likelihood, which picks the values that best fit your data. What you read off the model is “given these inputs, how likely is this lead to convert?” If the probability is above your threshold, you predict conversion. If it is below, you predict no conversion.

Figure 6: Logistic regression predicts a probability between zero and one. The dashed line marks the classification threshold. Points above the line get predicted as conversions, points below as non-conversions.

Marketing examples. Lead scoring is the obvious one. Will this lead become a customer? Churn prediction is another. Will this customer leave in the next thirty days? Conversion likelihood for a given website visitor is a third. Logistic regression also shows up under the hood in many ad platforms for click-through prediction, although those are usually layered into much larger ensemble models.

I went through logistic regression in detail in What Happens Next? Predicting Website Conversions with Logistic Regression (Domaleski, 20 July 2025), and the question of how to evaluate a logistic model is the subject of Confused? Don’t Worry, Your Marketing Model Is Too (Domaleski, 13 July 2025).

Decision Trees and Random Forests

What it is. A decision tree is a model that makes predictions by asking a series of yes-or-no questions about the input. Each question splits the data into two groups. The tree keeps splitting until the groups are homogeneous enough to assign a prediction. A random forest is many such trees, each trained on a slightly different sample of the data and each allowed to ask only a random subset of the possible questions at any split. The forest makes its prediction by taking a vote or an average across the trees.

How it works. Each split is chosen to maximize the purity of the resulting groups. For classification, purity is measured with metrics like Gini impurity or entropy. For regression, it is measured by variance reduction. The tree keeps splitting until it hits a stopping criterion. That could be a maximum depth or a minimum number of observations per leaf. A single tree is interpretable and easy to draw, but it tends to overfit. A random forest fixes that. It builds many trees, each grown on a slightly different sample of the data, and each tree only gets access to a random subset of features at each split. Figure 7 shows the difference between a single tree and a forest of voting trees. The diversity of the trees reduces variance and produces a more reliable prediction. The trade-off is you give up the simple flowchart you can show your team.

Figure 7: A single decision tree (left) is interpretable but tends to overfit. A random forest (right) combines many trees voting together to produce a more reliable prediction, at the cost of being harder to explain.

Marketing examples. Lead scoring with interpretable rules is a perfect fit for a single decision tree. Customer segmentation that needs to be explained to a non-technical audience is another. Random forests work well when you have many possible inputs and want a robust prediction without spending weeks on feature engineering. Attribution modeling, especially when you are trying to identify which channel combinations drive conversion, has been done successfully with both decision trees and random forests.

I covered the foundations in Smarter Marketing with Decision Trees (Domaleski, 9 Mar. 2025).

K-Means and Clustering

What it is. K-means is one of the most common unsupervised clustering algorithms. You give it a dataset and tell it how many clusters to find. It groups similar points together. Points in the same cluster are similar to each other. Points in different clusters are different. There are no labels involved. The algorithm finds the groupings on its own.

How it works. The algorithm starts by placing K (number) of centroids at random locations. Every data point gets assigned to its nearest centroid. The centroid of each cluster then moves to the mean of its assigned points. Points get reassigned to the new nearest centroid. Centroids update again. The process repeats until the cluster assignments stop changing. You end up with K clusters, each defined by its centroid. Figure 8 shows what this looks like in two dimensions.

K-means has some well-known limitations. You have to specify K in advance, which is often the question you wanted the algorithm to answer for you. The clusters depend on the random initial placement of the centroids, so results can vary run to run unless you fix the random seed. It assumes clusters are roughly spherical and similar in size, which is not always true in real customer data.

Figure 8: K-means groups customers into clusters based on similarity. Each cluster has a centroid at its center, and every customer belongs to the cluster whose centroid is closest. Here, customers cluster naturally into four segments based on purchase frequency and spend.

Marketing examples. Customer segmentation is the canonical use case. RFM segmentation, which groups customers by Recency, Frequency, and Monetary value, can be done by hand with quartiles or much more flexibly with k-means. Persona discovery, where you want to find natural groupings in customer behavior without imposing a predefined framework, is another strong fit. Survey response clustering, market basket grouping, and content audience segmentation all use k-means or its close relatives.

Time Series Models

What it is. Time series models are designed for data that depends on its own past. You already know what a time series looks like even if you have never called it that. Think of a temperature chart that goes up in summer and down in winter, year after year. That is a time series. Sales, traffic, demand, and most other marketing outcomes work the same way. They go up and down with seasons, with the day of the week, with the calendar, and with their own recent history. A time series model captures those patterns and projects them forward.

How it works. The classical time series workhorse is ARIMA, which stands for AutoRegressive Integrated Moving Average. ARIMA was formalized by Box and Jenkins in 1970, and the underlying ideas go back to the 1920s, making it one of the older methods we now consider part of the machine learning toolkit. The AR part says the current value depends on its own recent values. The MA part says the current value depends on recent prediction errors. The I part removes trends so the series is easier to model. ARIMA is tuned by picking the right values for the three parameters that govern those components.

Meta’s Prophet, which I have used extensively, takes a different approach. It treats the time series as a sum of pieces:

y(t) = trend(t) + seasonality(t) + holidays(t) + noise(t)

The trend can bend as needed. The seasonal piece captures repeating patterns. The holiday piece captures known events. Noise picks up whatever is left. Prophet handles missing data, outliers, and irregular trends without requiring the analyst to be a time series expert. Figure 9 shows what a time series looks like.

Figure 9: A time series breaks into trend, seasonality, and noise. A forecast extends those patterns into the future, with a confidence band that widens as uncertainty grows.

Marketing examples. Website traffic forecasting is a good example of a marketing time series, which I demonstrated in How We Forecasted Website Traffic for a Local Newspaper Using Meta’s Prophet (Domaleski, 24 Nov. 2024). Demand forecasting for inventory and budget planning is another. Economic forecasting using ARIMA models, which I walked through in How to Create a Big-Picture Economic Forecast Using Time-Series (ARIMA) Models in Python (Domaleski, 11 Jan. 2026), is a third. The family of naive forecasting methods, which I covered in Simplicity is All You Need: The Art (and Science) of Naive Forecasting (Domaleski, 22 June 2025), is the baseline every other time series model can be measured against.

Recommendation Systems

What it is. A recommendation system is a model that predicts what a user might like based on the preferences of similar users or the similarities between items. The two main flavors in classical recommendation systems are user-based collaborative filtering (UBCF) and item-based collaborative filtering (IBCF). Both work without any explicit understanding of what the items are. They learn purely from patterns of interaction.

How it works. User-based collaborative filtering looks for users similar to the target user. Similarity is measured with something like cosine similarity or Pearson correlation, based on the items both users have rated. Then it recommends items those similar users liked. Item-based collaborative filtering flips the idea. It looks for items similar to ones the target user has already liked. Item similarity is measured by how often pairs of items are rated similarly across the user base. Then it recommends those similar items.

Both methods rely on the user-item interaction matrix. Figure 10 shows what that matrix looks like and how the two approaches read it differently. User-based methods adapt quickly to new user preferences. Item-based methods scale better when the catalog is large and stable. That is why most production systems tend to be item-based.

Figure 10: The user-item matrix is the foundation of collaborative filtering. The goal is to predict the missing rating, in this case, how User B would rate Item 4. User-based methods look across users (green arrows) to find similar people. Item-based methods look across items (purple arrows) to find similar products.

Modern recommendation systems often combine collaborative filtering with content-based features and deep learning. The underlying idea is still the same. Find a way to estimate what is in the blank cell of the user-item matrix.

Marketing examples. Product recommendations on e-commerce sites are the textbook use case. Content recommendations on media platforms are another. Cross-sell engines that suggest complementary products to existing customers run on collaborative filtering at their core. Email personalization, where you tailor the products featured in a campaign based on what similar customers have purchased, is another use case where the math is the same.

I went through the comparison between user-based and item-based methods in detail in UBCF vs. IBCF: Comparing Marketing Recommendation System Algorithms in R (Domaleski, 6 Apr. 2025).

Gradient Boosting and XGBoost

What it is. Gradient boosting is an ensemble method. It builds many small decision trees in sequence. Each tree corrects the errors made by the trees before it. The most popular implementation is XGBoost, short for Extreme Gradient Boosting. It is fast and accurate. It is also the current darling of data science competitions on platforms like Kaggle, where XGBoost and its cousins consistently take the top spots in prediction competitions. When you just need the most accurate prediction and you do not need to explain how you got there, gradient boosting is usually the model to beat. If random forests are a democracy of trees voting independently, gradient boosting is an assembly line of trees. Each one picks up where the last one fell short.

How it works. The algorithm starts with a simple initial prediction. Often that is just the average of the target variable. It then computes the residuals, which are the differences between the predicted and actual values. The next tree is trained not to predict the original target, but to predict those residuals. The output of that tree, scaled by a learning rate, is added to the initial prediction. New residuals are computed. Another tree is trained on those. This continues for hundreds or thousands of iterations, with each tree shrinking the remaining error. Figure 11 below shows the assembly line in action. The “gradient” in gradient boosting comes from gradient descent. Each new tree is one step in the direction that reduces error the most.

Figure 11: Gradient boosting builds trees in sequence. Each tree is trained to correct the errors of the trees before it. The final prediction is the combined output of all the trees.

Marketing examples. High-stakes lead scoring, where accuracy matters more than interpretability, is where gradient boosting shows up often. Demand forecasting with many predictors is another strong fit. Customer churn prediction at scale, where you have rich behavioral data and need the best possible accuracy, often ends up running on XGBoost or one of its cousins like LightGBM or CatBoost.

The trade-off, which I explored in Linear Regression vs. XGBoost: Which Predicts Sales Better (Domaleski, 1 June 2025), is that you give up the clean coefficient story. You cannot point to a number and say “a one-dollar increase in this channel produces this much revenue.” You get accuracy and you lose explainability. For some marketing decisions that is the right trade. For others it is not.

Natural Language Processing

What it is. Natural language processing is the family of techniques for turning unstructured text into something a machine can work with. Reviews, social posts, support tickets, survey responses, sales call transcripts, and emails are all unstructured text. NLP is how you extract sentiment, topics, categories, and patterns from them at scale. It is also one of the machine learning techniques I use the most in my marketing agency work, especially for sentiment analysis. I wrote about a simple approach in Basic Sentiment Analysis Using R with VADER (Domaleski, 20 Oct. 2024).

How it works. The classical NLP pipeline starts with tokenization. That is the process of breaking text into words or sub-word units. Stop words like “the” and “and” are often removed because they carry little information. Tokens are then converted into numerical representations. The simple methods are bag-of-words and TF-IDF. The more sophisticated approach uses dense vector representations called embeddings. From there, you can do many things. Sentiment analysis uses lexicons or trained classifiers to score the emotional tone of the text. Topic modeling, using methods like Latent Dirichlet Allocation (LDA), finds the underlying themes in a body of text. Text classification assigns labels using supervised learning. Figure 12 shows the basic pipeline from raw text to structured output. Modern NLP often uses pretrained transformer models, which I will get to in the LLM section, but the classical methods are still useful and often sufficient.

Figure 12: The classical NLP pipeline turns raw text into something a model can use. Each stage transforms the text a little more. Splitting it, cleaning it, converting it to numbers, until a model can extract structured insights like sentiment, topics, or labels.

Marketing examples. Sentiment analysis of online reviews, which I covered across a series of articles starting with Sentiment Analysis of Online Reviews using R (Domaleski, 22 Sept. 2024), is one of the most accessible marketing applications. Social listening on platforms like Reddit, which I walked through in Step-by-Step Guide to Social Listening on Reddit using Python (Domaleski, 15 Dec. 2024), is another. Support ticket triage, lead qualification from form text, and competitive intelligence from public review sites are other examples. For a broader overview I wrote A Marketer’s Guide to NLP: How Machines Actually Process and Understand Language (Domaleski, 19 Oct. 2025).

Neural Networks

What it is. A neural network is a model made up of layers of simple processing units called neurons. The approach was originally inspired by how neurons in the human brain work. The earliest version was the perceptron, introduced by Frank Rosenblatt in 1958. It had a single layer and could only learn very simple patterns. Modern neural networks stack many layers, which is why they are often called deep learning models. They take an input, pass it through the layers, and produce an output.

Neural networks are what first got me interested in AI back in the early 1990s. They are still one of my favorite models to work with. Who would have guessed they would end up powering today’s large language models (LLMs), which I will cover next.

How it works. The neurons in a network are arranged in layers, which are connected by weighted links. Each neuron takes the outputs from the previous layer and multiplies them by its own weights. It sums the weighted values and adds a bias term. The result is passed through an activation function such as Sigmoid or ReLU. We covered Sigmoid earlier in this article. That output becomes the input to the neurons in the next layer. Figure 13 shows the basic structure which consists of an input layer, one or more hidden layers, and an output layer.

The network learns through a process called backpropagation. It compares its prediction to the correct answer. The difference is called a loss which is propagated backward through the network using the chain rule of calculus. The weights are updated in the direction that reduces the loss. This process repeats many times. The network gradually learns to produce better predictions.

Figure 13: A feedforward neural network. Inputs flow left to right through hidden layers, with each neuron applying a weighted sum and an activation function. The output layer produces the final prediction. Training adjusts the weights so the outputs get closer to the correct answers.

Deep neural networks, which have many hidden layers, can learn very complex patterns. They are the engine behind image recognition, speech recognition, and most of what people mean today when they say “AI.” The cost is that they need large amounts of training data, significant compute, and their reasoning is extremely hard to explain.

Marketing examples. Image recognition for visual content moderation or for analyzing user-generated photos is one. Voice analysis from call center recordings is another. Complex pattern recognition in user behavior where the relationships are too non-linear for simpler models is a third. In most small business marketing contexts, neural networks are overkill. In large-scale digital marketing, especially anything involving images or audio or very large structured datasets, they are increasingly the default.

I wrote about neural networks in three previous articles. Marketing with Neural Networks: What They Are and How to Use Them (Domaleski, 13 Apr. 2025) is the conceptual overview. Neural Networks Explained from First Principles in R (Domaleski, 20 Apr. 2025) walks through the math in code. Building a Neural Network from First Principles in Excel (Domaleski, 17 May 2026) shows the same network built in a spreadsheet so you can watch every calculation.

Large Language Models

What it is. Large language models are a type of deep learning model, meaning neural networks with many layers. LLMs use a specific layered architecture called a transformer. They are trained on enormous amounts of text (aka “corpus”). They learn to predict the next token in a sequence based on the tokens that came before. In simple terms, an LLM is a giant word prediction machine. That sounds modest but at scale, they can generate text, summarize documents, classify content, answer questions, follow instructions, and increasingly take actions through agentic workflows.

LLMs are also different from the other machine learning models in another important way. With traditional machine learning, you are running a model defined by measurable statistics and knowable algorithms. Even the harder-to-read ones, like a random forest or XGBoost, can be audited. You can pull feature importances, trace a path through a tree, or inspect the weights. An LLM is different in kind. You feed it a prompt and it gives you an output, and there is no comparable way to trace why. Take sentiment analysis. I could use a traditional NLP model and understand exactly how it scored the text. Or I could ask ChatGPT to rate the sentiment, and it would. The LLM answer is usually less precise. What it gives up in precision, it makes up for in speed, ease of use, and flexibility.

How it works. The transformer architecture was introduced in a 2017 paper called Attention Is All You Need by researchers at Google. It replaced the older recurrent neural network approach with a mechanism called self-attention. Self-attention lets the model weigh how relevant every token in the input is to every other token. This captures long-range patterns that older architectures missed. The model is trained on huge amounts of text through self-supervised learning. The training signal comes from predicting the next token. After the base training, models are often fine-tuned with human feedback to better match what users actually want.

The size of these models is part of what makes them work. Modern frontier models contain enormous numbers of parameters, often in the hundreds of billions. That scale lets them generalize across tasks they were never explicitly trained on. This is what people mean when they talk about “emergent capabilities.” Figure 14 shows the prediction process in action. The model generates one token at a time. Each new token becomes part of the input for the next prediction.

Figure 14: An LLM generates text one token at a time. Each prediction depends on everything that came before it, including tokens the model just produced. A full ChatGPT response is the result of thousands of these iterative predictions strung together.

Marketing examples. Content drafting is the obvious one, with the caveat that humans should review the output. Customer support automation through chatbots is another. Classification of text at scale, where you might once have built a logistic regression classifier or a topic model, can now be done with a well-prompted LLM. Emerging agentic workflows let the LLM take actions like searching the web, summarizing results, and producing recommendations. These systems are still new and they need real guardrails before you turn them loose on anything customer-facing. I walked through building one in Agentic AI with R: Building an Automated Website Auditor from Scratch (Domaleski, 27 Apr. 2025), and a more accessible version in Getting Started with Agentic AI: Building A Virtual Assistant with Google Gemini (Domaleski, 1 Mar. 2026).

A word of caution on LLMs in marketing work. They are not reliable sources of factual truth. They hallucinate. They generate confident-sounding text that is wrong. For any marketing use case where accuracy matters, you need guardrails. That includes product descriptions with specs, claims about your services, customer support responses with policy details, and anything customer-facing.

The guardrails come in two forms. The first is retrieval-augmented generation, where the model pulls from a verified source instead of making things up. The second is human review. Often you need both. The fluency of these models makes their errors hard to spot. That is exactly why they are dangerous when nobody is checking. I have written about how I use and do not use LLMs in my own agency work in I Own a Marketing Agency. Here’s How We Use (and Don’t Use) AI Like ChatGPT in 2025 (Domaleski, 2 Feb. 2025), and about the limitations of these models in AI is Messy, Flawed, and Useful (Domaleski, 22 Feb. 2026).

Other Useful Machine Learning Methods

The algorithms above cover the bulk of what you will encounter in marketing analytics and data science, but four others deserve at least a mention. Figure 15 shows each one visually.

Naive Bayes is a simple probabilistic classifier based on Bayes’ theorem. It assumes the features are independent of each other. That assumption is often unrealistic. Despite that, it works surprisingly well on text classification problems. Spam filtering is the classic example. Sentiment classification, topic categorization, and lead qualification from form text are marketing-relevant variants. Naive Bayes trains fast. It needs relatively little data. It gives you a probability output you can interpret. It is often a strong baseline before you reach for anything more complex.

K-Nearest Neighbors (KNN) is a supervised classifier that does no real training in advance. When you give it a new data point, it looks at the K labeled training points closest to it and predicts the new point’s label based on what those K neighbors are. Despite sharing a letter with K-means, KNN is fundamentally different. KNN is supervised. K-means is unsupervised. In marketing, KNN works for lead scoring, churn classification, and basic product recommendation when you have rich labeled data. Its main drawback is speed at prediction time. It has to compare every new point against every training point. You will see KNN more often in academic settings than in production marketing systems, but it is a useful baseline to keep in mind.

Figure 15: Four other ML methods worth recognizing in marketing analytics. Naive Bayes assigns probabilities based on Bayes’ theorem. KNN classifies a new point by looking at its nearest labeled neighbors. SVM finds the widest margin between two classes. Anomaly detection flags points that deviate from the normal pattern.

Support Vector Machines (SVMs) are classifiers that find the best boundary between two classes. That boundary is called a hyperplane. The SVM finds the version with the widest possible margin between the classes. They were one of the dominant classification methods before tree ensembles and neural networks took over. SVMs still hold up well on smaller datasets with many features, when you do not have enough samples to train a deep model. You will see them less often in modern marketing pipelines than you used to, but they are worth knowing about when you encounter them in older systems or in academic literature.

Anomaly detection is a family of methods rather than a single algorithm. The goal is to identify observations that deviate substantially from the normal pattern in your data. Some methods are statistical, like using z-scores or interquartile ranges. Others use unsupervised learning, like isolation forests or one-class SVMs. In marketing, anomaly detection helps catch spend spikes from misconfigured campaigns, bot traffic that inflates engagement numbers, fraud in conversion data, and tracking failures where a key event suddenly stops firing. It is one of the more underused tools in the marketing analytics toolkit, and one of the most operationally valuable.

Machine Learning Model Trade-offs

The machine learning algorithms are only half the picture. The other half is the set of trade-offs that come with using them. These are some of the common issues that get people in trouble with machine learning models. Consider these four things before you deploy any model in a marketing context.

Black box versus explainability. A random forest you can explain to your CFO beats a neural network nobody on the team understands. Some marketing decisions require justification. That includes anything around budget allocation, customer targeting, or compliance. If you cannot explain why a model recommended what it did, you may not be able to use it for those decisions, regardless of how accurate it is.

Bias. Models inherit the biases of the data they were trained on. In marketing, this shows up in many ways. Audience targeting can systematically exclude certain groups. Lead scoring can mirror historical patterns of who got attention rather than who deserved it. Feedback loops amplify whatever the model started with. Bias is not always intentional. But it is always present in some form. It is the responsibility of the analyst to check for it.

Figure 16: Four trade-offs every marketer needs to weigh before deploying a model. Interpretability tells you whether you can explain the prediction. Bias tells you whether your training data systematically excluded certain groups. Overfitting warns that a model fitting the training data too perfectly may not generalize. Data requirements remind you to match model complexity to the size of your dataset.

Overfitting. A model that fits last quarter’s data perfectly will often fail next quarter. The model has memorized the noise as well as the signal, and the noise does not repeat. I explored this problem in Why Many Marketing Models Fail: The Problem of Imbalanced Data (Domaleski, 14 Dec. 2025). The fix is honest evaluation on data the model has not seen. The standard way to do that is cross-validation, where you repeatedly hold out part of the data, train on the rest, and check how well the model does on the part it never saw. It also takes a willingness to choose a simpler model that generalizes well over a complex one that fits the training data exactly.

Data requirements. Neural networks and LLMs need more data than most small businesses have. Linear regression can work on a few dozen observations. A deep neural network typically cannot. Matching the complexity of the model to the size and quality of your data is one of the most important judgment calls in any marketing analytics project. I wrote about how to do meaningful analytics with limited data in How to Be Data-Driven in Marketing Even If Your Small Business Doesn’t Have a Lot of Data (Domaleski, 3 Aug. 2025).

Garbage in, garbage out. There is one more thing that matters more than any of the four trade-offs above, and more than the choice of algorithm itself. It is the quality of the inputs you give the model. The process of turning raw data into useful inputs is called feature engineering, and in practice it is often the hardest part. A thoughtfully engineered logistic regression will frequently beat a poorly prepared neural network. Good data usually beats a fancy model. Before you reach up the ladder for something more powerful, ask whether better inputs would get you there with the model you already have.

Evaluation Metrics

A model is only as useful as the metric you judge it by, and the right metric is not always obvious. Different problems demand different evaluation metrics. Picking the wrong one can make a bad model look good or a good model look bad.

For binary classification, like lead scoring or churn prediction, accuracy is rarely the right metric. That may sound counterintuitive, but most marketing datasets are imbalanced. Only a small percentage of leads convert. Only a small percentage of customers churn. A model that predicts “no” for everyone will have high accuracy but no value. Some relevant metrics for classification are:

• Precision. Of the leads you predicted would convert, how many actually did?
• Recall. Of all the leads that actually converted, how many did your model catch?
• F1 score. A balance of precision and recall in a single number.
• AUC (Area Under the ROC Curve). A measure of how well your model separates the two classes across all possible thresholds.
• Calibration. When your model says “70% likely to convert,” does it actually convert 70% of the time?

I covered some of these in detail in Confused? Don’t Worry, Your Marketing Model Is Too (Domaleski, 13 July 2025).

For regression problems, like revenue forecasting or customer lifetime value, the metrics tell you how far off your predictions are. The common ones:

• RMSE (Root Mean Squared Error). The square root of the average squared error. Penalizes large errors more heavily.
• MAE (Mean Absolute Error). The average error, in the same units as what you are predicting.
• MAPE (Mean Absolute Percentage Error). The same idea expressed as a percentage.
• R-squared. The proportion of the variation in the outcome that the model explains. Higher is better.

Figure 17: Different machine learning problems demand different evaluation metrics. Classification problems use precision, recall, and AUC. Regression problems compare predictions to actual values using error measures like RMSE. Time series problems also care about capturing turning points. Attribution measures lift, the difference between treatment and control.

For time series problems, like traffic forecasting or sales forecasting, you use regression-style metrics. But you should also care about whether the model captures the turning points. A model that predicts a flat line might score well on RMSE while missing every peak and valley that matters for the business.

For attribution and marketing measurement, the metrics shift entirely. The question is no longer “how accurate is the prediction?” but “did the marketing actually cause the outcome?” The key metric here is lift, the difference between treatment group performance and control group performance. Incremental ROI and iROAS (incremental Return on Ad Spend) are closely related. They measure the return on marketing spend after adjusting for what would have happened anyway. The next section goes deeper on this because prediction metrics and causal metrics answer fundamentally different questions.

Picking the right metric is part of picking the right model. Use the metrics that match the question you are actually trying to answer.

When Prediction Is Not Enough: Causality and Lift

There is a category of marketing question that machine learning, by itself, cannot answer. That is whether your marketing actually caused the outcome you observed.

Machine learning is exceptional at finding correlation. Given enough data, a model can learn that customers who saw a particular ad are more likely to convert. What the model cannot tell you is whether the ad caused those conversions. The customers who saw the ad might have converted anyway. That distinction is the difference between prediction and causation. It is also the difference between describing what happened and knowing whether your marketing earned its budget.

This matters enormously in attribution. Most attribution models are correlational. That includes last-click, first-click, and the various data-driven multi-touch attribution methods you find in ad platforms. They assign credit based on observed patterns in conversion paths. They tell you what touched the customer before they converted. They do not tell you which of those touches mattered. A customer who would have converted without ever seeing your ad still shows up in the data as a converted customer who saw your ad. The model gives that ad credit it did not earn. Figure 18 shows the difference between what prediction can tell you and what causality can tell you.

Figure 18: Prediction tells you what is correlated with conversion. It cannot tell you which of those customers would have converted without your marketing. Causal measurement, like a controlled lift study, compares treatment and control groups directly. The difference between them is the lift, the part of the outcome your marketing actually caused.

The fix is causal measurement. That usually means running an experiment. Geo holdouts withhold marketing in some regions and run it in others, giving you a clean comparison. Conversion lift studies randomly assign some users to a control group that does not see your ad. Marketing Mix Models with proper validation can approximate causal estimates at the channel level. Incrementality testing is the discipline that ties these methods together. I covered the full landscape in Measuring Lift: A Marketer’s Guide to Incrementality (Domaleski, 5 Apr. 2026), and I walked through the modern attribution frameworks in my four-part MMM series, ending with The Best of Both Worlds: Combining MMM and MTA with Unified Marketing Measurement (Domaleski, 29 Mar. 2026).

The takeaway is straightforward. Use machine learning to predict, segment, and personalize. Use experiments and causal methods to measure lift. They answer different questions, and you need both. A marketing organization that runs only on prediction models will systematically overcredit its marketing. A marketing organization that runs only on experiments will miss the day-to-day patterns that prediction models surface. The two work together. Neither replaces the other.

Three Questions to Consider When Choosing The Right Machine Learning Model

There are three questions to ask before reaching for any machine learning model in this field guide.

First, what decision will this model inform? If you cannot point to a specific decision the model will help you make, you do not need the model. A surprising number of marketing analytics projects start without a clear decision in mind. Someone wants to “use AI” or “do machine learning.” That is not a decision. A decision is which leads to call first, which customers to send a retention offer to, or how much to spend on a campaign next month.

Second, do you have the data and the time to support the approach? A neural network without enough data is worse than a linear regression with enough data. A model nobody has time to build, validate, and maintain is worse than no model at all. There’s a hidden cost in every machine learning project. Someone has to keep the data clean, retrain the model when things change, and make sure the outputs reach the people who use them. If you do not have time for that, you do not have a project.

Third, will you be able to explain the result to the person who has to act on it? If the answer is no, you are probably using the wrong model or you need to do more work translating the output into something actionable. The person acting on the model output is usually not a data scientist. It is a CFO, a sales manager, a client, a campaign manager, or a board. They need to understand the recommendation well enough to defend it, push back on it, or own the outcome. If they cannot, the model will not get used.

Figure 19 lays out these three questions as a decision flowchart.

Figure 19: Three questions to ask before reaching for any machine learning model. If the answer to any of them is no, you are likely choosing the wrong tool. The right model is one that informs a real decision, fits your data and time, and produces results you can explain.

If you cannot answer all three questions, the model is probably the wrong choice. It does not matter how clever the math is or how impressive the algorithm sounds. I made a similar point in Marketing Model Selection: There’s No Need to Throw AI at Everything (Domaleski, 25 May 2025).

The right model is the simplest one that informs a real decision, fits your data and time, and produces results someone can act on.

When to Use What Machine Learning Model

Figure 20 is a compact reference for matching common marketing problems to a starting model. Treat it as a starting point, not a final answer. The right choice always depends on your specific data, the decision you are trying to inform, and the operational constraints around the model.

Figure 20: A reference table matching common marketing problems to a starting model. Start with the first model suggested. Upgrade only when the simpler model falls short. Keep the main risk in mind before deploying any model in production.

Regarding Figure 20 above, the “first model to try” column is almost all simple methods. That is intentional. The simpler model is usually the right answer. Move up the ladder of model complexity only when you can point to something specific the simpler model is missing. Pay attention to the main risk column. Every model has a way of going wrong and this is where marketing analytics projects usually trip up.

Conclusion

My LinkedIn profile says I use data science and machine learning for BETTER MARKETING. Now you know what I mean by machine learning. It’s not just ChatGPT and LLMs. It’s the full family of methods we just covered.

There is a natural temptation to reach for the most advanced tool every time. Better marketing rarely calls for that.

A few key points to take away from this field guide:

• Better marketing is about better decisions, not bigger models.
• Each algorithm is suited to a specific kind of pattern. Match the tool to the job.
• The escalation from linear regression to large language models is one of capability and cost, not correctness.
• Black box concerns, bias, overfitting, and data requirements are important parts of a successful model, not just an afterthought.
• Prediction and causation are different things. Be aware of the difference.
• LLMs trade precision for speed and ease. For many marketing problems, a simpler model is the better way to go.

On prediction vs. causation: use machine learning to predict, segment, and personalize. Use experiments and causal methods to measure whether your marketing actually drove the outcome. Both belong in a healthy marketing analytics practice. Neither replaces the other.

If you remember one thing from this article, let it be this. Start with the simplest model that solves your problem. Evaluate it against the question you are trying to answer, not against some imagined benchmark of sophistication. Escalate only when the simpler approach doesn’t do what you need it to do.

The deep-dive articles linked throughout are your next step for any model or technique you want to know more about. I wrote those first. This field guide came later to help orient you. The field guide gives you the map. The individual articles teach you how to use the tools.

Before dumping everything into an LLM and getting a response that either does not make sense or is wrong, consider another form of AI. It is called machine learning. I use machine learning for BETTER MARKETING and you can too!

Resources

Chapman, Christopher N., and Elea McDonnell Feit. R for Marketing Research and Analytics. 2nd ed., Springer, 2019.

Cunningham, Scott. Causal Inference: The Mixtape. Yale University Press, 2021, https://mixtape.scunning.com/.

Domaleski, Joe. “A Better Marketing Manifesto.” Marketing Data Science, 8 Sept. 2024, https://blog.marketingdatascience.ai/a-better-marketing-manifesto-24b6541a6cb9.

Domaleski, Joe. “A Marketer’s Guide to NLP: How Machines Actually Process and Understand Language.” Marketing Data Science, 19 Oct. 2025, https://blog.marketingdatascience.ai/a-marketers-guide-to-nlp-how-machines-actually-process-and-understand-language-3d452febb3de.

Domaleski, Joe. “Agentic AI with R: Building an Automated Website Auditor from Scratch.” Marketing Data Science, 27 Apr. 2025, https://blog.marketingdatascience.ai/agentic-ai-with-r-building-an-automated-website-auditor-from-scratch-1a42d0bd7bae.

Domaleski, Joe. “AI is Messy, Flawed, and Useful.” Marketing Data Science, 22 Feb. 2026, https://blog.marketingdatascience.ai/ai-is-messy-flawed-and-useful-6133052d9282.

Domaleski, Joe. “An Introduction to Marketing Mix Modeling (MMM): How It Works and Why It’s Making a Comeback.” Marketing Data Science, 8 Mar. 2026, https://blog.marketingdatascience.ai/an-introduction-to-marketing-mix-modeling-mmm-how-it-works-and-why-its-making-a-comeback-3791c1ccb914.

Domaleski, Joe. “Basic Sentiment Analysis Using R with VADER.” Marketing Data Science, 20 Oct. 2024, https://blog.marketingdatascience.ai/basic-sentiment-analysis-using-r-with-vader-4eecb738566f.

Domaleski, Joe. “The Best of Both Worlds: Combining MMM and MTA with Unified Marketing Measurement.” Marketing Data Science, 29 Mar. 2026, https://blog.marketingdatascience.ai/the-best-of-both-worlds-combining-mmm-and-mta-with-unified-marketing-measurement-097e36cdae5b.

Domaleski, Joe. “Building a Neural Network from First Principles in Excel.” Marketing Data Science, 17 May 2026, https://blog.marketingdatascience.ai/building-a-neural-network-from-first-principles-in-excel-0562c4e542bf.

Domaleski, Joe. “Confused? Don’t Worry, Your Marketing Model Is Too. Here’s How a Confusion Matrix Can Help.” Marketing Data Science, 13 July 2025, https://blog.marketingdatascience.ai/confused-dont-worry-your-marketing-model-is-too-here-s-how-a-confusion-matrix-can-help-63c017375de4.

Domaleski, Joe. “Getting Started with Agentic AI: Building A Virtual Assistant with Google Gemini.” Marketing Data Science, 1 Mar. 2026, https://blog.marketingdatascience.ai/getting-started-with-agentic-ai-building-a-virtual-assistant-with-google-gemini-080054a10ede.

Domaleski, Joe. “How to Be Data-Driven in Marketing Even If Your Small Business Doesn’t Have a Lot of Data.” Marketing Data Science, 3 Aug. 2025, https://blog.marketingdatascience.ai/how-to-be-data-driven-in-marketing-even-if-your-small-business-doesnt-have-a-lot-of-data-831d233d1ec6.

Domaleski, Joe. “How to Create a Big-Picture Economic Forecast Using Time-Series (ARIMA) Models in Python.” Marketing Data Science, 11 Jan. 2026, https://blog.marketingdatascience.ai/how-to-create-a-big-picture-economic-forecast-using-time-series-arima-models-in-python-08093acfa6f1.

Domaleski, Joe. “How We Forecasted Website Traffic for a Local Newspaper Using Meta’s Prophet and R.” Marketing Data Science, 24 Nov. 2024, https://blog.marketingdatascience.ai/how-we-forecasted-website-traffic-for-a-local-newspaper-using-metas-prophet-and-r-0ca046bfcb34.

Domaleski, Joe. “I Own a Marketing Agency. Here’s How We Use (and Don’t Use) AI Like ChatGPT in 2025.” Marketing Data Science, 2 Feb. 2025, https://blog.marketingdatascience.ai/i-own-a-marketing-agency-heres-how-we-use-and-don-t-use-ai-like-chatgpt-in-2025-f85f40ca55d5.

Domaleski, Joe. “Linear Regression: The Classic Machine Learning Algorithm You Need to Know.” Marketing Data Science, 1 Dec. 2024, https://blog.marketingdatascience.ai/linear-regression-the-classic-machine-learning-algorithm-you-need-to-know-1fe0b48b06a3.

Domaleski, Joe. “Linear Regression vs. XGBoost: Which Predicts Sales Better (Using R)?” Marketing Data Science, 1 June 2025, https://blog.marketingdatascience.ai/linear-regression-vs-xgboost-which-predicts-sales-better-using-r-3bdf5aaab6bc.

Domaleski, Joe. “Marketing Model Selection: There’s No Need to Throw AI at Everything.” Marketing Data Science, 25 May 2025, https://blog.marketingdatascience.ai/marketing-model-selection-theres-no-need-to-throw-ai-at-everything-8030ede44248.

Domaleski, Joe. “Marketing with Neural Networks: What They Are and How to Use Them.” Marketing Data Science, 13 Apr. 2025, https://blog.marketingdatascience.ai/marketing-with-neural-networks-what-they-are-and-how-to-use-them-3e0114771c40.

Domaleski, Joe. “Measuring Lift: A Marketer’s Guide to Incrementality.” Marketing Data Science, 5 Apr. 2026, https://blog.marketingdatascience.ai/measuring-lift-a-marketers-guide-to-incrementality-ab4f57b21cc6.

Domaleski, Joe. “Neural Networks Explained from First Principles in R.” Marketing Data Science, 20 Apr. 2025, https://blog.marketingdatascience.ai/neural-networks-explained-from-first-principles-in-r-795bc158cf37.

Domaleski, Joe. “Sentiment Analysis of Online Reviews using R.” Marketing Data Science, 22 Sept. 2024, https://blog.marketingdatascience.ai/sentiment-analysis-of-online-reviews-using-r-e2afbc9fcc68.

Domaleski, Joe. “Simplicity is All You Need: The Art (and Science) of Naive Forecasting in Marketing Analytics.” Marketing Data Science, 22 June 2025, https://blog.marketingdatascience.ai/simplicity-is-all-you-need-the-art-and-science-of-naive-forecasting-in-marketing-analytics-9c9bba859940.

Domaleski, Joe. “Smarter Marketing with Decision Trees: A Practical Introduction Using R and Python.” Marketing Data Science, 9 Mar. 2025, https://blog.marketingdatascience.ai/smarter-marketing-with-decision-trees-a-practical-introduction-using-r-and-python-11077a964f34.

Domaleski, Joe. “Step-by-Step Guide to Social Listening on Reddit using Python.” Marketing Data Science, 15 Dec. 2024, https://blog.marketingdatascience.ai/step-by-step-guide-to-social-listening-on-reddit-using-python-22a7905afcee.

Domaleski, Joe. “Stop Confusing Marketing with Sales, They’re Not the Same Thing.” Marketing Data Science, 17 Nov. 2024, https://blog.marketingdatascience.ai/stop-confusing-marketing-with-sales-theyre-not-the-same-thing-1ad4e9466c1f.

Domaleski, Joe. “UBCF vs. IBCF: Comparing Marketing Recommendation System Algorithms in R.” Marketing Data Science, 6 Apr. 2025, https://blog.marketingdatascience.ai/ubcf-vs-ibcf-comparing-marketing-recommendation-system-algorithms-in-r-38ff36bf05d3.

Domaleski, Joe. “What Happens Next? Predicting Website Conversions with Logistic Regression.” Marketing Data Science, 20 July 2025, https://blog.marketingdatascience.ai/what-happens-next-predicting-website-conversions-with-logistic-regression-d374875764e2.

Domaleski, Joe. “Why Many Marketing Models Fail: The Problem of Imbalanced Data.” Marketing Data Science, 14 Dec. 2025, https://blog.marketingdatascience.ai/why-many-marketing-models-fail-the-problem-of-imbalanced-data-ceb8123b80b3.

Hastie, Trevor, et al. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed., Springer, 2009, https://hastie.su.domains/ElemStatLearn/.

James, Gareth, et al. An Introduction to Statistical Learning: with Applications in Python. Springer, 2023, https://www.statlearning.com/.

James, Gareth, et al. An Introduction to Statistical Learning: with Applications in R. 2nd ed., Springer, 2021, https://www.statlearning.com/.

Kohavi, Ron, et al. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press, 2020.

Vaswani, Ashish, et al. “Attention Is All You Need.” Advances in Neural Information Processing Systems, vol. 30, 2017, https://papers.nips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.