Context
A 2026 research preprint tests how search-augmented LLMs use current evidence across 225 changed facts and reports that snippet-only answers correlate with more stale answers and unsupported citations than answers citing opened pages.
What changed
The authors trace whether assistants open retrieved source pages before answering and quantify stale answers and citation faithfulness by reading depth, using deployed and frozen-retrieval experiments.
Why it matters
Source retrieval and even citation do not guarantee that an AI answer read or faithfully used the underlying evidence. This adds a measurable distinction between discovery, source opening, evidence use and answer support.
Visibility OS interpretation
Visibility OS interpretation: measure the source-to-answer funnel as retrieval, page reading, evidence incorporation and citation faithfulness; do not treat citation counts as proof of answer influence.
What remains uncertain
Do the associations persist across other proprietary engines and natural free-form tasks? Would randomized page-opening interventions reduce failures? Can mechanistic claims be confirmed causally?