Triangulation Collapse: How LLM-Mediated Search Removes the User's Last Line of Defense in High-Stakes Queries

For decades, search engines functioned as triangulation tools. A user searching for information about a worrying symptom, an employment dispute, or a financial decision would encounter multiple sources, compare them, disagree with some, and arrive at a conclusion of their own making. That comparison process (imperfect, slow, occasionally frustrating) was also the primary error-correction mechanism available to non-expert users navigating high-stakes information domains. The rapid adoption of LLM-mediated search, which synthesizes ranked results into single authoritative-sounding responses, is systematically eliminating this behavior. I call this elimination Triangulation Collapse (TC). In this paper, I characterize TC as a distinct failure mode in AI-augmented search systems, give the engagement-accuracy decoupling mechanism that drives it a formal model with three proven propositions, document its consequences across four high-stakes domains (health, legal, financial, and hiring), and propose SARI (Stakes-Aware Re-ranking and Inference), a framework for re-introducing accuracy-oriented signals into LLM search pipelines for queries where the cost of error is high. The core argument is straightforward: LLM search systems have inherited the engagement-optimization logic of their underlying ranking infrastructure, and for high-stakes queries, engagement and accuracy are not the same thing. The consequences of that mismatch are not visible in any standard quality metric, and that invisibility is, itself, a provable consequence of how single-answer synthesis is presented, not merely an observation.