Personalized news recommendation is crucial for engaging users on news websites. In this domain, there are numerous specific content-based methods that employ the profiles of news articles. Among these, embedding-based approaches (Embedding-Based News Recommendation; EBNR) are widely recognized for their efficacy. However, their performance for news publishers remains unclear, as prior EBNR research has primarily focused on news aggregators. News publishers and aggregators are generally distinct in the following two points: (1) news publishers’ ability to leverage richer metadata, and (2) the narrower topic range of some news publishers that focus on specific affairs compared to news aggregators. This paper aims to provide practical insights into how these differences affect news recommendation. Through an online experiment with the Nikkei, Japan’s leading financial news company, we reported that basic EBNR is not superior to simple content-based recommendation systems that directly utilize metadata as news profiles, and further analysis using open datasets shows that the narrowness of the topics covered by each news website affects the diversity of embeddings. Overall, our findings suggest that characteristics of news publishers impact EBNR performance, highlighting the need for future research on methods to mitigate these effects.