Regression (ML Task)

Regression is the supervised machine learning task of predicting a continuous number rather than a category. If the answer you want is “how many” or “how much” (sessions next month, revenue per visitor, expected clicks), it is a regression problem. In MLforSEO’s task framework it sits next to classification as one of the two branches of supervised learning: classification predicts a discrete label, regression predicts a value.

How it works

A regression model learns the relationship between input variables, called features, and a numeric target, called the label. The simplest version, linear regression, fits a straight line: the prediction equals a starting value (the bias) plus each feature multiplied by a learned weight. Google’s Machine Learning Crash Course uses car weight to predict fuel efficiency, where heavier cars get a negative weight because they travel fewer miles per gallon.

More flexible models such as decision trees, random forests, gradient boosting and neural networks can capture curves and interactions between features. Predictions are judged by how far they land from the true values, using metrics like mean absolute error (MAE), root mean squared error (RMSE) and R² (the share of variation the model explains).

One naming trap: logistic regression is a classification algorithm. It outputs a probability for a class, and that probability is then turned into a label.

How it’s used in SEO and marketing

• Forecasting organic traffic, clicks or sales from historical performance data, a close relative of time series forecasting.
• Estimating expected click-through rate for a ranking position, so you can spot pages that underperform their position.
• Predicting customer lifetime value or order value from behaviour data.
• Modelling how much content length, links or page speed relate to performance (with care: relationships are not causes).
• Estimating the revenue impact of a project before you pitch it.

Example

Say you export twelve months of Search Console data for 500 product pages. For each page you have impressions, average position and title length, plus the clicks it earned. A regression model trained on most of those pages learns how the features relate to clicks, then predicts clicks for the held-out pages. Pages whose real clicks fall well below the prediction are candidates for title and snippet work.

Limitations and things to watch

• Correlation is not causation. A feature that predicts traffic well may simply travel alongside the real cause.
• Models are only as good as their training data. A core update or a new AI search feature can change the relationship the model learned.
• Complex models can overfit: they memorise the training pages and predict new ones badly. Always test on data the model has not seen.
• Extrapolating beyond the range of your data (a page with ten times more links than anything in the set) is guesswork.
• Report a range, not a single number, when sharing forecasts with stakeholders.

Learn it on MLforSEO

• The Characteristics of a Machine Learning Problem places regression in the task framework.
• What Machine Learning Classification Is, and Where It Fits in SEO explains the classification versus regression split.
• ML data characteristics mapped to SEO ML tasks (free guide).
• Introduction to Machine Learning for SEO course, Machine Learning Basics module.