Generative AI for Scalable and Explainable E-Commerce Product Title Evaluation: A Prompt-Driven Framework

Priyadarshini Balachandran

Abstract: Product titles are a decisive factor in e-commerce, directly impacting search visibility, customer engagement, and sales conversion. However, practices such as keyword stuffing, excessively long titles, and promotional claims often degrade discoverability and marketplace quality. This paper introduces a generative AI?based framework for automated product title evaluation that classifies titles as "Good" or "Bad" with interpretable reasoning. Unlike traditional machine learning pipelines requiring labeled datasets and feature engineering, our approach leverages large language models (LLMs) with few-shot prompting to deliver transparent, audit-ready outputs. The system architecture comprises of four modular components, Normalizer & Precheck, LLM Judge, LLM Provider Layer, and Policy Aggregator, ensuring scalability, explainability, and adaptability to policy changes. Experiments conducted on a benchmark dataset of 1,000 manually annotated product titles achieved 91% accuracy, 88% precision, and 85% recall, demonstrating strong performance in detecting violations such as keyword stuffing, category mismatch, and irrelevant descriptors. The framework reduces development overhead, supports rapid policy iteration, and provides structured feedback for sellers, thereby enhancing marketplace compliance and buyer trust. Future work will explore multilingual extensions, continuous human-in-the-loop refinement, and integration with complementary metadata such as images and seller reputation to further strengthen title quality evaluation.

Keywords: Search Relevance Optimization, Content Quality Evaluation, Generative AI Evaluation, Large Language Models, AI Orchestration Evaluation

How to Cite?: Priyadarshini Balachandran, "Generative AI for Scalable and Explainable E-Commerce Product Title Evaluation: A Prompt-Driven Framework", Volume 14 Issue 9, September 2025, International Journal of Science and Research (IJSR), Pages: 171-177, https://www.ijsr.net/getabstract.php?paperid=SR25906035244, DOI: https://dx.doi.org/10.21275/SR25906035244