What if company websites could act as a live radar for technological change? This study uses large language models to turn corporate web language into a measurable picture of manufacturing’s shifting technology landscape.

Technology mining enables the systematic extraction and analysis of trends regarding scientific and technological developments from diverse Science, Technology, and Innovation data sources. Traditionally, indicators such as patents, surveys, and R&D statistics have been widely used to monitor technological change and innovation.

Technological advances are reshaping the global economic landscape, with the manufacturing sector experiencing significant changes. At the core of this shift are technologies such as cloud computing, the Internet of Things, the Internet of Services, big data analytics, and artificial intelligence.

These technologies enable manufacturers to transcend conventional automation by integrating intelligent systems that drive operational efficiency, flexibility, and continuous innovation. Traditional data collection methodologies, such as surveys, interviews, and patents, are costly, resource intensive, and lack the ability to capture real-time technological dynamics.

These limitations underscore the need for agile and scalable approaches that leverage dynamic data sources to track the evolution of technology and emerging trends more effectively. Corporate websites offer a dynamic source of data for monitoring technological trends, providing distinct advantages over traditional sources such as patents, surveys, or scientific publications.

While conventional methods typically capture innovation at formalized or later stages, corporate websites deliver real-time insights into how firms present their technological capabilities and strategic initiatives to external audiences. Several studies have utilized corporate web content combined with keyword-based approaches to monitor the adoption of artificial intelligence and related technologies.

However, detecting and tracking all relevant technologies remains a significant challenge. Traditional keyword-based methods often rely on manually predefined lists of terms, which may not capture emerging terminology or evolving language use, leading to incomplete assessments of technological adoption and potential human errors.

This paper proposes an LLM-assisted framework for identifying and tracking technological trends in corporate web content. The methodology proceeds in three steps: firms are classified as AI adopters by combining embedding-based semantic similarity with LLM-based assessment of website content.

Candidate unigrams and bigrams are extracted from the selected text fragments and classified as technological or non-technological using an LLM. Conceptually equivalent terms are grouped, and binary term-firm matrices are constructed for each time point, 2023 and 2025.

Figure four presents the authors’ methodological workflow in three linked stages: sampling a web corpus, detecting AI adoption, and identifying technology terms. The supporting steps include gathering firm information from SABI, retrieving websites for twenty twenty-three and twenty twenty-five, cleaning and segmenting text, applying embedding-based filtering and relevant-fragment selection, and then using an LLM for AI detection.

Finally, selected fragments are preprocessed, unigrams and bigrams are extracted, and technology terms are classified and grouped. The framework is applied to corporate websites of Spanish manufacturing firms collected in the first quarter of 2023 and 2025, enabling a structured longitudinal comparison of technological discourse without relying on manually predefined keyword lists.

The dataset was constructed using information from the Sistema de Análisis de Balances Ibéricos, a commercial database extensively utilized in economic and innovation research. Companies were selected when they had employee information available in the SABI database and when their NACE code was included in the SABI data.

A total of 20,691 Spanish manufacturing firms were identified. The sample includes firms of varying sizes, including micro-enterprises, small and medium-sized enterprises, and large companies, ensuring comprehensive coverage of the sector. Each company corpus was segmented into 1,000-byte fragments and embedded into a high-dimensional semantic space using OpenAI’s text-embedding-3-small model.

Cosine similarity was calculated against a reference embedding of the phrase Artificial Intelligence to assess the semantic relevance of the fragments. For each company, the ten most semantically relevant, non-overlapping fragments were selected to maximize contextual diversity and reduce redundancy.

This embedding-based filtering step reduces the volume of text evaluated by the LLM, lowering computational costs while preserving contextual diversity. The subsequent LLM assessment performs a deeper semantic interpretation to determine whether the fragments provide actual evidence of AI use.

In both years, the selected fragments were evaluated using GPT-4o-mini, guided by a structured rubric designed to assess whether evidence of AI use was present or absent in each company’s web data. The full LLM rubric and prompt used for AI detection are reported in Appendix A, and the prompt design was iteratively refined by testing alternative formulations and retaining the version that yielded the most consistent classifications according to the evaluation metrics.

In 2023, a total of 356 companies, or 1.7 percent of the initial 20,691 firms, were classified as AI adopters. By 2025, this number increased to 467 companies, or 2.3 percent, reflecting measurable growth in AI-related engagement across the sector.

An LLM was employed to classify whether each extracted term represented a technological concept and to identify distinct terms that referred to the same underlying technology. This integrated process consolidated multilingual expressions, alternative formulations, and synonymous terms to represent a single technological entity where applicable, minimizing redundancy within the dataset.

A comparative analysis was conducted to quantify changes in term frequencies between the two years. For each technological term, absolute differences and relative growth rates were calculated, allowing the identification of temporal patterns in technological discourse.

Aggregating results across unigrams and bigrams, the model achieved a mean accuracy of zero point nine nine zero, a precision of zero point eight seven six, a recall of zero point seven five nine, and an F1-score of zero point eight zero eight. The slightly lower recall is mainly driven by unigram classifications, which are inherently more ambiguous than bigrams.

The model adopts a more conservative classification strategy, prioritizing precision and reducing false positives, which leads to lower recall but maintains high overall accuracy. Figure 2 illustrates the percentage change in web-mention frequency between 2023 and 2025 for the most referenced technologies.

Overall, most terms display an upward trend, several more than doubling over the period, while only a handful show modest declines. The far-right end of the chart is occupied by Predictive Maintenance, at one hundred percent, while Augmented Reality and Virtual Reality sit around the upper end of the scale.

At the opposite end, Mobile Device and Neural Network are the only points left of zero, with the remaining technologies clustered between roughly five and sixty-five percent. Figure 3 shows the co-occurrence network for the top technologies in 2025.

Each node represents a technology, while edges indicate that associated technologies were frequently mentioned together on firms’ websites. Edge width corresponds to the strength of association measured by Jaccard similarity, while node size indicates degree centrality and highlights technologies that frequently co-occur with multiple other technologies.

The densest central region contains automated system, computer vision, robotics, and machine learning, with multiple visible connections linking these technologies. Computer vision is also visibly connected to both robotics and machine learning, while automated system extends toward real time and Industry four point zero, forming a broader cluster of digital production terms.

The analysis of how technologies co-occur on websites provides insight into technological convergence in manufacturing. Certain technologies are frequently mentioned together, implying that firms adopt them in complementary bundles. Automated system frequently co-occurred with computer vision, robotics, real time, machine learning, and Industry 4.0, suggesting that firms often frame automation as part of a broader digital production architecture.

Internet of Things also co-occurred with real time and machine learning, reflecting the synergy between connectivity, automation, and intelligence in smart production systems. Reliance on corporate website content as a data source introduces potential biases because company websites are a form of self-presentation.

Firms might overstate involvement with fashionable technologies to signal innovativeness, or under-report certain capabilities due to secrecy or compliance reasons. The study is contextually focused on Spanish manufacturing firms, and this focus limits the generalizability of the findings.

The methodological framework proposed is scalable and can be easily applied to other contexts and languages. The paper introduced an LLM-based methodology for real-time detection and tracking of technological trends through the semantic analysis of corporate web content.

Applied to Spanish manufacturing firms’ websites from 2023 and 2025, automated technology classification illustrates how LLMs can complement traditional keyword- and survey-based approaches. Large language models enabled flexible and scalable identification of emerging and evolving technologies, providing a cost-effective alternative to conventional tech mining tools.

This approach reduces the manual effort required for identifying technological trends while enhancing the comprehensiveness of technology assessments compared to previous methodologies. The study shows that LLM-assisted analysis of corporate websites can provide a flexible, scalable, and cost-effective way to track changing technology discourse, while revealing rising AI mentions and connected technology bundles.