Large process model collections in use today contain hundreds or even thousands of conceptual process models. Search functionalities can help in handling such large collections for purposes such as duplicate detection or reuse of models. One popular stream of search functionalities is similarity-based search which utilizes similarity measures for finding similar models in a large collection. Most of these approaches base on an underlying alignment between the activities of the compared process models. Yet, such an alignment seems to be quite difficult to achieve according to the results of the Process Model Matching contests conducted in recent years. Therefore, the Latent Semantic Analysis-based Similarity Search (LS3) technique presented in this article does not rely on such an alignment, but uses a Latent Semantic Analysis-based similarity measure for retrieving similar models. An evaluation with 138 real-life process models shows a strong performance in terms of Precision, Recall, F-Measure, R-Precision and Precision-at-k, thereby outperforming five other techniques for similarity-based search. Additionally, the run time of the LS3 query calculation is significantly faster than any of the other approaches.