In this article, a new unsupervised feature extraction method for aspect-based sentiment analysis is proposed. This method improves the performance of frequency based feature extraction by using an online search engine. Although frequency based feature extraction methods produce good precision and recall values on formal texts, they are not very successful on informal texts. Our proposed algorithm takes the features of items suggested by frequency based feature extraction methods, then, eliminates the features which do not co-occur with the item, whose features are sought, on the Web. Since the proposed method constructs the candidate feature set of the item from the Web, it is domain-independent. The results of experiments reveal that for informal Turkish texts, much higher performance than frequency based method is achieved.