The patent for Dynamic and precise website categorization using a large language model (Patent Number: 12699724) was recently granted to Palo Alto Networks, Inc.. Invented by Zeyu You and Wei Wang, this technology introduces a sophisticated method for classifying web content. The system trains a concept tagging model using a large language model to analyze website text and identify core concept terms. It then applies an embedding model to these terms, generating multi-dimensional embeddings to match the website to the most granular and accurate category within a vast taxonomy.
This invention is highly innovative because it overcomes the limitations of traditional, static keyword matching and manual domain classification. By integrating large language models and advanced embedding structures, the system possesses a semantic understanding of web page text. This allows for dynamic, highly accurate categorization even when dealing with novel, ambiguous, or rapidly changing websites. This context-aware classification represents a significant advancement in network security, ensuring that filtering policies adapt to the nuances of modern web traffic in real time.
September 2026 Patent of the Month
This technology earned the title of “Patent of the Month” for the ai-software-crypto-cloud industry in September 2026 due to its critical role in modernizing digital infrastructure security. As cloud environments, artificial intelligence tools, and decentralized crypto platforms generate increasingly complex and dynamic web traffic, traditional categorization methods often fall behind. Palo Alto Networks’ LLM-driven approach provides the real-time adaptability required to maintain zero-trust security postures across these advanced ecosystems, making it a foundational technology for managing and securing modern enterprise internet traffic.
Eligibility for US R&D Tax Credits
The practical applications of this patent provide a strong framework for claiming Research and Development (R&D) tax credits in the United States. To implement or build upon this technology, software engineering teams would need to undertake significant technical experimentation to integrate custom large language models, refine embedding parameters, and minimize processing latency for live network environments. Under the IRS four-part test, resolving these engineering uncertainties through iterative development, modeling, and testing qualifies as valid research. Organizations that invest in developing or heavily customizing such AI-driven categorization systems can likely claim R&D tax credits to offset the costs of domestic engineering wages, cloud computing infrastructure, and third-party technical contractors.
