Explanations are essential for building trustworthy recommender systems, yet selecting an appropriate explanation method is challenging. Numerous methods exist, but guidance on which works best in which scenario is scarce. Most generation‑based explanations are abstract and need further formatting to become user‑friendly, leading to an apparently endless pool of options. Running user studies for every option is usually infeasible, and automated metrics either assess only the abstract output or require a ground‑truth reference that is often unavailable.\ \ Recent work suggests that large language models (LLMs) can act as "judges" for explanation evaluation, but their reliability has not been thoroughly examined. This paper first explores LLMs' ability to generate explanation prototypes given varying information about the recommender system and the user. We create 18 distinct explanation prototypes and evaluate them with 14 LLMs of different sizes under two temperature settings, then compare the scores to human ratings from a user study.\ \ Results show that LLMs exhibit human‑like rating patterns and achieve moderate rank correlation with human raters, but absolute agreement is low and varies considerably with model size and evaluation construct. From these findings we derive four practical recommendations: keep explanation‑generation prompts concise; prefer larger models for evaluation; pre‑test evaluation constructs; and audit explanations for factual accuracy, as neither humans nor LLMs reliably detect non‑factual content.\ \ Review