The contents of online news documents are almost the same that will lead to the redundancy of news or called yellow journalism. Yellow journalism can make it difficult for readers to distinguish documents containing fact or opinionated information. Therefore, it is necessary to extend more research about multi-document summarization so that readers can easily understand the intent of online news documents. Latent Dirichlet Allocation (LDA) - Significance Sentences is one of the methods for summarization, which performs better than the term frequency algorithm. However, document summarization using the method is only able to summarize multiple documents as a whole without grouping by topic. Subsequently, it can give an unorganized summary result. Therefore, this research proposes a novel summarization method which combines K-Means Clustering and LDA - Significance Sentences, so it can generate document summaries based on the topic. We implemented two scenarios of the experiment. The first experimental results the best alpha value is 0.001 with the ROUGE-1 value of 0.5545 and the best summarization level is 30% with the ROUGE-1 value of 0.6118. While the second experiment results, the best obtain of ROUGE-1 value is 0.61991 for the first cluster which is consists of documents 1, 2, 3, 4, and 6 and 0.6139 for the second cluster which is consists of 5, 7, 8. Multi-document summarization using the proposed method has good performance when the K-Means method can cluster the document according to the topic correctly, which is highly dependent on the accuracy of determining the initial centroid.