Large language models can be sensitive to the position of information inside a prompt, potentially favoring items because of ordering rather than content. This study evaluates positional bias across multiple LLMs by asking them to generate descriptions for sets of topics whose order is systematically permuted. The analysis compares structural and coherence-related properties of the resulting outputs and examines whether models reorder or omit topics. Results show that prompt position can influence response structure and coherence, although the magnitude and form of the effect vary across models and topics. The findings demonstrate that order effects should be considered when LLMs are used for ranking, judging, or other tasks where equivalent items should receive position-independent treatment.