The so-called real-time web (RTW) is a web of opinions, comments, and personal viewpoints, providing personalized commentary in real-time. Facebook is undoubtedly the king of the RTW. It boasts 1+ billion users that spend about 23 billion minutes per day, generating in the region of 60 million posts per day. This RTW data is far from structured (in contrast to data such as ratings, product features, etc.), but it is very useful to consider for reputation monitoring. For enterprises in particular, being present and aware of what is discussed on social media about their products and services has become a must. It allows them to establish a communication channel with their customers, market their products, build brand equity and boost clientele faithfulness. However, as social media are two-way channels, they require effort and care to manage this communication. Dissatisfied customers can protest out loud, easily influence many other customers and damage the brand's image. In order to avoid these risks, several social media monitoring tools have been implemented, enabling enterprises to have access to real customers’ opinions, complaints and questions at real time in a highly scalable manner. Characteristic examples include the SM2 tool of Alterian, Brandwatch, Converseon, Cymfony Maestro, My BuzzMetrics, Radian6, Sysomos and others. Even though very interesting and somehow effectively performing on open social media (twitter, blogs, microblogs etc), most aforementioned tools have very limited performance on rule-stringent media (where tough access policies are imposed). Furthermore usually extensive human guidance is needed and raw-tag manipulations are necessary. In order to bypass these problems, in this paper we propose an intelligent wrapper system that automatically segments closed social media web pages into structural tokens, extracts and associates opinions to each token. A key step towards retrieving the data of interest is to discover the sections contained in a web page and identify the ones holding the interesting information. To do that, our method is based on a clustering and statistics origin. The proposed system can operate without human intervention and training. Initial experiments are presented over Facebook content, which indicate the promising performance of the proposed scheme.