This paper suggests a content-based image retrieval scheme, applicable to large web collections. A fast and robust system is proposed, based on the popular bag-of-words model. SURF features are extracted from images and a visual vocabulary is created, through which images are efficiently represented. Furthermore, geometric constraints on the image features are taken into account, facilitating accurate retrieval. The performance of the proposed methods is evaluated on two common datasets, while an advanced web-based image retrieval application is presented, that yields a geographic position estimation about the query image, exploiting geo-tagged datasets.