A framework for classifying company webpages in business research: evaluating conventional ML against few-shot LLMs

Abstract Company websites serve as the official online representation of companies and reflect various aspects of a business. As companies evolve and carry out their operations, their website pages undergo changes, which can be used to track their economic activity. However, not all changes are equally connected to the company’s business operations. This paper proposes a webpage classification framework to analyze changes on these webpages and associate them with changes in the company. Five categories of company webpages were proposed as representative of a company’s activities: Corporate, Post, Product, Catalogue and Transactional. A dataset of 1440 manually labeled webpages was compiled to evaluate how effectively the defined classes can be distinguished using four conventional machine learning classifiers (logistic regression, random forest, LightGBM, and a multilayer perceptron) and two few-shot large language models (LLMs; Gemma4 31B and Ministral 3 14B). To isolate input modality from model family, a term frequency–inverse document frequency (TF-IDF) + LightGBM (LGBM) baseline was trained on the same text used by the LLMs. The results show that all approaches can separate the five categories, with the Post and Transactional classes being the most readily identifiable. The LLMs still outperform the TF-IDF + LGBM baseline under identical text input, indicating that the advantage is not solely attributable to input modality. This classification provides a first step toward monitoring the business behavior of companies based on webpage categories .