How it works
ChatGPT Search is the web-search-enabled variant of ChatGPT, launched 2024. When a query requires fresh information, the model invokes its web_search tool, retrieves ~5–8 candidate pages via Bing's index + OpenAI's own crawler, then synthesizes an answer with inline citations. The retrieval is hybrid (dense embeddings + BM25 + authority). Citations are presented as numbered footnotes [1][2] that expand to source cards. Not every ChatGPT query triggers retrieval: the model decides per turn based on query freshness needs.
Signals it weights heavily
- JSON-LD schema
- Author E-E-A-T
- Citation density
- Freshness for time-sensitive queries
- Open Graph metadata for source cards
How to optimize for ChatGPT Search
- Allow GPTBot + OAI-SearchBot in robots.txt explicitly (not just defaults).
- Add og:image + og:title. Source cards in ChatGPT render visually.
- BLUF-format your top pages: ChatGPT prefers direct-answer leads.
- Strong Author + Date markup signals freshness for time-sensitive queries.