Five webpage categories could track what companies are really doing online

Company websites change constantly, but not every change means the same thing. A new product page signals routine commercial activity; a rewritten data-protection policy may hint at internal restructuring. A team of researchers in Spain has now built a framework that sorts company webpages into five functional categories and tested how reliably machine learning models and large language models can tell them apart, offering what they describe as a first step toward monitoring business behavior directly from the web.

The study, published as an open-access regular paper in the International Journal of Data Science and Analytics, was conducted by Joan Manuel Valenzuela Rubilar, Josep Domenech and Ana Pont-Sanjuán. Their central idea is that the pages making up a corporate website differ in how long they live and how often their content changes, and that these patterns carry information about the firm behind the site. Routine activity, they argue, tends to show up in Post, Product and Catalogue pages, while structural transformations are more often visible in Corporate and Transactional pages.

The five categories are Corporate, Post, Product, Catalogue and Transactional. Corporate pages hold information about the organization itself, such as company profiles or policy documents, and their content is usually stable, so changes there may indicate significant internal shifts. Post pages carry news articles and blog entries and accumulate user-generated material like comments. Product pages present goods and services, and their creation or removal tracks the product life cycle. Catalogue pages aggregate links to products and posts, with content that changes regularly through curated updates. Transactional pages support purchases, forms, reservations and account access, and changes there can reflect new operational processes such as new payment methods.

To test whether these categories can be distinguished automatically, the researchers compiled a dataset of 1440 manually labeled company webpages in 17 languages, 68.5 percent of them in Spanish, balanced with 288 records per class. They started from 999 Spanish firms drawn from the SABI business database, stratified into large, medium-sized and small companies according to the European Union definition of small and medium-sized enterprises. From a random subset of one-third of these firms, crawling conducted between 19 and 24 January 2026 collected 31,171 pages from 327 sites, capped at 300 pages per website with a crawl depth of three levels.

Labeling was carried out over four weeks by one expert annotator, who visually inspected pages in a browser while excluding global navigation elements such as menus and footers. Seventeen pages with unclear classes were excluded, and minority classes were augmented with newly annotated pages to keep the dataset balanced without duplication. To estimate reliability, a second expert independently labeled a stratified subsample of 145 pages, reaching what the authors report as almost perfect agreement under the Landis and Koch benchmark, with Cohen’s kappa of 0.819 and raw agreement of 85.5 percent.

The classification experiments compared seven approaches. Four conventional machine learning classifiers were trained on more than 57,000 features extracted from the HTML, text and URLs of the labeled pages: logistic regression, random forest, LightGBM and a multilayer perceptron. Two few-shot large language models, Gemma 4 31B from Google DeepMind and Ministral 3 14B from Mistral AI, were prompted with a fixed set of ten hand-picked demonstration examples, two per category, and given 3000 characters of cleaned plain text per page. Because the conventional models and the language models differed in both model family and input, the team added a controlled baseline: a TF-IDF text representation fed into the same LightGBM algorithm, using the identical 3000-character extracts the language models received.

The results show that all approaches could separate the five categories, with Post and Transactional classes the most readily identifiable. Among the conventional models, LightGBM achieved the highest macro-averaged area under the ROC curve at 0.94, followed by random forest at 0.91, logistic regression at 0.87 and the multilayer perceptron at 0.85. LightGBM also led on accuracy and F1 at 0.75. Only Gemma 4 31B beat the best conventional models on hard classifications, reaching accuracy and F1 of 0.80, while Ministral 3 14B, at 0.69 accuracy, underperformed LightGBM and random forest. The text-only TF-IDF baseline scored lowest, with accuracy and F1 of 0.64.

The controlled baseline proved decisive for interpretation. Restricting the classifier to plain text reduced accuracy by 0.11, from 0.75 to 0.64, while switching from the machine learning classifier to Gemma 4 under the same text input raised accuracy by 0.16, from 0.64 to 0.80. The two effects partly offset, yielding Gemma 4’s net 0.05 advantage over the full-feature LightGBM. This indicates that the language model’s advantage is not merely a consequence of reading plain text but also of the model family itself. Per-class results reinforced the point: the Catalogue class was hardest for every conventional model, with LightGBM recall of 0.65, yet Gemma 4 reached 0.88 on it, while Ministral excelled on Corporate and Post but struggled on Product.

The authors are careful about several caveats. The language models’ AUC values were computed from one-hot hard labels rather than token-level probabilities, which the cloud API does not expose, so they reflect a fixed combination of recall and specificity rather than genuine ranking and cannot be compared directly with the machine learning AUCs. Specificity values above 0.90 are inflated by construction, since in each one-versus-rest task the negative class contains four of the five categories. The outer cross-validation loop also splits at the page level, meaning 70.3 percent of test pages share a domain and class with a training page, which may inflate the absolute metrics of the conventional models, although the prompt-based language models, which perform no parameter adaptation, are unaffected. A website-level split is left for follow-up work.

The study also has clear limits on scope. It covers Spanish companies, a pragmatic choice driven by institutional access to the SABI database, and evaluates pages at a single point in time without analyzing changes over time. The researchers stress that their results support the taxonomy only as a classification scheme: they show the categories are separable, not that webpage changes actually reflect business activity, a link that remains an open question. Future work plans to compare snapshots of company webpages over time, classify the changes detected, and test whether they connect to firms’ regular activity, potentially allowing near-real-time assessment of company health for market analysis if validated.

Subject of Research: A webpage classification framework distinguishing five categories of company webpages and comparing conventional machine learning with few-shot large language models

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

Article References: Valenzuela Rubilar, J. M., Domenech, J., & Pont-Sanjuán, A. (2026). A framework for classifying company webpages in business research: evaluating conventional ML against few-shot LLMs. International Journal of Data Science and Analytics, 22(1), Article 366. https://doi.org/10.1007/s41060-026-01337-w

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01337-w

Keywords: webpage classification, machine learning, large language models, few-shot learning, company websites, business research, LightGBM, TF-IDF, text classification, corporate web monitoring, dataset annotation, cross-validation