How ChatGPT Selects Sources: A Deep Dive into OpenAI's Recommendation Engine

ChatGPT, built on OpenAI's GPT-4 and subsequent models, is the most widely used AI assistant. Understanding how it selects brands and sources it mentions is essential for any AI visibility strategy.

ChatGPT's Response Generation

ChatGPT generates responses through two mechanisms. In default mode, it draws on training data patterns. In browsing mode, it uses retrieval-augmented generation to search the web via Bing, retrieve pages, and incorporate information with inline citations.

How ChatGPT Selects Brands in Default Mode

Brand selection is influenced by frequency of mention in training data, context associations, sentiment patterns, and authority signals from training data. Brands discussed more often across authoritative sources are more likely to surface.

How ChatGPT Selects Sources in Browsing Mode

When browsing, ChatGPT relies on Bing's search ranking to surface candidate pages, then evaluates content quality, relevance, and recency. Pages with clear structure and direct answers are preferred.

Optimizing for ChatGPT

For default mode: Build broad web authority. Ensure your brand appears in authoritative publications, industry reports, and well-referenced articles. Maintain consistent entity information.

For browsing mode: Optimize for Bing SEO. Create well-structured content answering common questions. Ensure content is current and comprehensive.

Across both modes: Implement structured data, maintain llms.txt, and publish comprehensive content. Citerna tracks your ChatGPT appearance across both modes.

ChatGPT-Specific Considerations

ChatGPT tends to provide balanced, comparative answers listing multiple options. Being included in the consideration set is the key objective. Memory features mean individual users may receive personalized responses, complicating measurement. Citerna uses fresh sessions for each query.

Common Mistakes

Over-focusing on a single keyword. Publishing thin content with no unique value. Blocking GPTBot in robots.txt while expecting browsing mode citations.