Primary-Source Retrieval and Citation Visibility: An Exploratory Comparison

This work was produced through an AI-assisted workflow directed and reviewed by the author. Ten queries for specific primary documents, drawn from live research work, went through five retrieval paths on 2 September 2026 and three on 20 October 2026. A hit is the expected document's address in a path's top ten. AI Act query: no matching Official Journal address was returned. No path on either date returned an address matching the specified Official Journal document, Regulation (EU) 2024/1689 (0 of 5 paths on 2 September and 0 of 3 on 20 October, at any rank). Other official pages were returned, including Commission pages on Article 50, and their contents were not inspected. DSA query: one path on each date. The Official Journal text was in the top ten of DuckDuckGo (rank 10 of its 10 results) on 2 September, and Codex (rank 2 of its 38 results) on 20 October. Chapter query: one path on each date. The current edition's quartz chapter was in the top ten of Brave (rank 3 of its 20 results) on 2 September, and Claude (rank 7 of its 9 results) on 20 October. Provenance: one path returns provider tokens. The Gemini path returned a provider redirect token in place of every source address on both dates (88 of 88 rows on 2 September, 59 of 59 on 20 October), with a bare domain as each title. Its citations cannot be read without a request to the provider. The key changes the order of the paths. On 20 October key v2 gives Codex 7, Claude 4 and Gemini 3 hits of 10, and key v1 gives Claude 6, Codex 5 and Gemini 4. Codex's key v2 count includes q10, whose address held the act at the time state dated 18 June 2022 in its latest archive copy, earlier than the state dated 31 December 2025 that the unsuffixed address showed (limitation 9). The research design, method, sourcing decisions and analytical judgements are the author's. The author directed and reviewed an AI-assisted workflow using models from Anthropic, OpenAI and Google. These makers are also subjects of the study; Brave Search is both a measured path and a declared supplier of Anthropic. These relationships can bias judgement. The Verification and Reproduction sections state what was checked and what remains unverified; numerical reproduction does not independently validate the scoring keys or establish what a source held on the run day. AI errors may survive review. No warranty is offered beyond the terms of the CC BY 4.0 licence.