Before the Citation: replication data and code (machine-written queries and engine-specific retrieval layers in generative search)

Replication data and code for the article "Before the Citation: Machine-Written Queries and Engine-Specific Retrieval Layers in Generative Search". Korean consumer questions were sent through the vendors' APIs in September 2026 to Claude (claude-sonnet-5), ChatGPT (gpt-5.6-sol) and Gemini (gemini-3.6-flash). For each call we recorded the search queries the engine executed and the pages it retrieved or cited, and compared them with the top 20 results of Brave Search, Google and Korean Bing. reproduce.py recomputes every value in Tables 2, 3 and 6 from the four CSV files. Search-result URLs are shared as hashed normalized URLs (32-bit FNV-1a) to respect providers' terms on redistribution; questions and executed queries are shared in full, except that the domain in any site: operator is replaced by its hash. See README.md.