My research suggests that Retrieval Augmented Generation (RAG) in comparative entity prompts (e.g. "best smartphones") for "high-frequency" (i.e. a topic that appears often in the training corpus such as public datasets, licensed content) domains (of knowledge) also with strong conceptual space (i.e. structure of concepts, categories, and relations) in Chagpt is multi-purpose:
Sentiment filtering (most significant role) - the information obtained using RAG helps the Chatgpt model actively remove entities with persistent negative associations, ensuring the final list looks credible and user-trustworthy as well as boosting high-sentiment exemplars e.g. it may be possible that the latest Iphone / Samsung S series mobile phone model has got bad reviews
Recency validation
Regional balancing (ensuring information is locally relevant)
Trend confirmation
Noise suppression
Thus it may be considered from what I see that RAG may be of less influence in response generation for these types of prompts than others relating to lower-frequency domains (of knowledge) with weaker conceptual spaces e.g. "best llm visibility checker software" (an emerging new product / area of knowledge).