This paper explores the effectiveness of traditional (modular) Optical Character Recognition (OCR) pipelines compared to Large Vision Language Models (LVLMs), particularly in terms of robustness, accuracy, and resource efficiency. We propose leveraging web-generated datasets to train OCR systems, highlighting the rich diversity of layouts, styles, and linguistic variations offered by web content. Our approach shows that web scraping with smart augmentations can generate diverse OCR training datasets for training modular OCR. Acquiring website screenshots for modular OCR training has yet to be explored and requires precise word localization for training the word detection model. This differs from many LVLMs, which are mostly trained end-to-end to extract full text. Experimental evaluations demonstrate that even though we trained our OCR pipeline on design website templates for developers rather than real public websites, we achieved competitive and even superior results with the state-of-the-art LVLM-based models and superior results in noisy and distorted scenarios, while requiring fewer computational resources for training and inference. These findings underscore the long-lasting relevance of modular OCR systems in diverse and resource-constrained settings. Although LVLMs present advantages in handling diverse and generalized tasks, their use is unnecessary for the OCR task, where modular pipelines excel in terms of efficiency and performance.