Dimensional Collapse in AI-Mediated Search: Large Language Models as Metameric Observers of Brand Advertising

Large language models (LLMs) increasingly mediate consumer brand evaluation in search and agentic commerce. This study demonstrates that LLMs systematically collapse brand perception to Experiential and Economic dimensions, leaving brands differentiated on narrative, ideological, cultural, or temporal dimensions metameric by inference, not by measured inter-brand distance. Using PRISM-B across 21,350 API calls to 24 models spanning seven training traditions and 15 languages, a Dimensional Collapse Index (DCI; uniform baseline = .250) is computed. Mean DCI equals .294 for global brands (p < .001) and .356 cross-culturally (d = 3.373). Cross-model convergence is near-perfect (cosine = .977). Local brands collapse 20% more severely than global brands (d = .92, in five FMCG pairs tested). Native-language prompting is null in home markets (p = .716) but reduces collapse up to .095 in foreign-market frames (p = .002). Geopolitical context independently modulates weights (p < .001). Structured Brand Function specifications remove two-thirds of above-baseline excess (.103 to .034). Collapse is stable across model versions. For search advertising, these findings reframe strategy from platform selection to dimensional defensibility: advertisers must encode soft brand dimensions in verifiable, machine-readable form or risk erasure by LLM intermediaries. Patterns reflect the text-conditioned LLM observers tested here. Includes zharnikov-2026v-ai-search-metamerism.yaml (Paper Spec v0.1.0) – a machine-readable specification of the paper’s claims, assumptions, and dependencies. The paper’s full machine-first bundle (the SPINE claim/dependency graph and the ONTOLOGY term module) lives in the public repository; see github.com/spectralbranding/paper-spec for the standard. This PDF is generated programmatically from that machine-first source under a research-as-repository model.