A classification model that estimates the probability that an input belongs to a category by applying a sigmoid function to a linear combination of input features, producing an output constrained between 0 and 1. Logistic regression is the standard baseline model for binary classification tasks in marketing, from conversion prediction to churn modeling to lead scoring.
Also known as logit model, binary regression, sigmoid regression
Logistic regression is a generalization of linear regression for binary outcomes, where the goal is to predict the probability that an example belongs to one of two categories, such as converted versus not-converted, churned versus retained, or clicked versus not-clicked. Rather than predicting the output directly as a linear function of the inputs, logistic regression applies a sigmoid function to the linear combination of inputs, squashing the output to the range between 0 and 1 so it can be interpreted as a probability. The model outputs a number such as 0.73, meaning the model estimates a 73% probability that the input belongs to the positive class.
The parameters of logistic regression are estimated by maximum likelihood: the model finds the coefficients that maximize the probability of observing the actual outcomes in the training data under the model. This is equivalent to minimizing the cross-entropy loss, which penalizes confident incorrect predictions more heavily than uncertain incorrect predictions. The resulting coefficient estimates have an intuitive interpretation in terms of log-odds: each unit increase in an input variable changes the log-odds of the positive outcome by the corresponding coefficient, and exponentiating the coefficient gives the odds ratio, a measure of how much more likely the positive outcome is per unit increase in that variable.
A classification threshold converts the continuous probability output to a binary prediction. The default threshold of 0.5 predicts the positive class when the estimated probability exceeds 50%, but this threshold can be adjusted based on the relative costs of false positives and false negatives. In lead scoring, where the cost of a missed high-intent lead is much higher than the cost of pursuing a low-intent lead, a lower threshold that captures more true positives at the cost of more false positives may be appropriate. The ROC curve and precision-recall curve visualize model performance across all possible thresholds, enabling principled threshold selection based on business cost considerations.
A working ad agency building propensity models for audience targeting, lead scoring for client sales teams, or churn risk models for subscription clients should always build a logistic regression baseline before building more complex models. Logistic regression’s coefficient estimates are interpretable in a way that tree-based models and neural networks are not: each coefficient directly quantifies the effect of that feature on the log-odds of the outcome, enabling the model to serve both as a predictive tool and as an explanatory analysis of which variables drive the outcome. Complex models often outperform logistic regression modestly in predictive accuracy while losing this interpretability entirely.
Conversion propensity scoring for lookalike audience targeting is built on logistic regression foundations. A model that predicts which website visitors are likely to convert is a logistic regression or its near equivalents: inputs are behavioral features such as pages visited, time on site, and content categories engaged with; the output is a conversion probability; and the training labels are historical conversion outcomes. The predicted probabilities are used to rank visitors by conversion likelihood, enabling the top-scoring visitors to be targeted with retargeting ads or personalized onsite experiences. Logistic regression’s probability calibration is an advantage here: the predicted probabilities have a meaningful absolute interpretation, enabling comparisons across campaigns and over time.
Churn risk scoring for client retention programs uses logistic regression on behavioral signals. A subscription client wanting to identify customers at risk of cancellation needs a model that takes behavioral signals such as declining engagement, reduced feature usage, and support contact patterns as inputs and outputs a churn probability. Logistic regression trained on historical churn outcomes produces a score that the client’s customer success team can use to prioritize outreach. The coefficients identify which behaviors are most predictive of churn, providing not just a score but an explanation of why specific customers are flagged as high-risk, enabling targeted intervention strategies rather than generic retention outreach.
Lead scoring models for B2B clients combine demographic and behavioral signals in logistic regression. A B2B client wants to prioritize sales follow-up by scoring inbound leads on their likelihood to convert to customers. Logistic regression takes features such as company size, industry, job title, website pages visited, and content downloaded as inputs and predicts the probability of conversion based on historical CRM data. The resulting scores rank leads for sales prioritization, and the model coefficients identify which signals are most predictive, enabling the marketing team to optimize lead generation campaigns to attract the characteristics associated with high conversion probability.
An agency is building a lookalike audience model for an e-commerce client to identify new potential customers from a publisher’s first-party audience data. The seed audience is the client’s 18,000 existing customers; the universe is the publisher’s 4.2 million registered users who have not yet made a purchase from the client. The agency trains a logistic regression model using client customer status as the binary outcome label and the publisher’s behavioral feature set of 85 variables as inputs. After applying L2 regularization to handle the correlated feature structure and scaling all features to unit variance, the model achieves an AUC of 0.74 on a held-out validation set. The coefficients reveal that the three most predictive features are engagement with home improvement content, purchase frequency on the publisher’s platform, and household income estimated from behavioral signals. These three features alone produce an AUC of 0.71, suggesting that a simplified three-feature scoring rule would perform nearly as well as the full model. The agency delivers both the full model for automated scoring and the three-feature interpretation to the client, which informs the client’s understanding of who their best potential customers are and enables more targeted creative messaging to the lookalike audience at the top of the predicted probability distribution.
The generative AI foundations module covers the classification models underlying marketing AI applications including logistic regression, its extensions, and the evaluation metrics used to assess and improve model quality.