Large language models give brands different lives: Chinese LLMs mention brands in English queries about 30 percentage points more often than international models, creating an 'Existence Gap' that can make firms invisible to AI-mediated discovery.

Main Finding

Large Language Models encode cultural and geographic patterns from their training corpora that produce systematic differences in brand visibility. Using 1,909 pure-English query–LLM pairs over 6 LLMs and 30 brands, the authors find Chinese LLMs mention brands far more often than International (Western) LLMs (88.9% vs. 58.3%; χ2 = 226.60, p < .001, ϕ = 0.34). They introduce the “Existence Gap”: brands absent or underrepresented in a model’s training data effectively do not “exist” in that model’s recommendations. A focal case (Zhizibianjie) shows 65.6% mention rate in Chinese LLMs versus 0% in International LLMs (χ2 = 21.33, p < .001, ϕ = 0.58), illustrating how training-data geography—not query language—drives cross-LLM visibility gaps.

Key Points

• Cultural Encoding: systematic output differences reflecting the linguistic/geographic composition of LLM training corpora. These operate even when queries are identical and in English.
• Existence Gap: a binary-like phenomenon where brands lacking sufficient presence in a model’s training data get little-to-no mention regardless of product quality or relevance.
• Data Moat (theoretical contribution): an extension of Resource-Based Theory framing AI-visible content (technical docs, case studies, community signals, APIs, etc.) as a VRIN resource that confers durable visibility advantages in AI-mediated discovery.
• Algorithmic Omnipresence: proposed strategic objective / metric—comprehensive brand visibility across multiple LLM ecosystems and query types.
• Managerial playbook: an 18-month roadmap (semantic coverage, technical depth, cultural localization) to build Data Moats and reduce Existence Gaps; introduces Generative Engine Optimization (GEO) as the analogue to SEO for LLM-mediated discovery.
• Hypotheses tested (summary):
  • H1: Chinese LLMs have higher brand mention rates than International LLMs even for pure-English queries (supported).
  • H2: Chinese LLMs will express more positive sentiment when mentioning brands that are well-represented in their training data (proposed/tested).
  • H3: Query intent moderates the Cultural Encoding effect (stronger for positional queries, weaker for comparative queries).
• The paper integrates Institutional Theory (linguistic boundary barriers as a de facto regulatory/normative mechanism) and Market Signaling (Data Moat signals quality to LLMs).

Data & Methods

• Dataset: 1,909 pure-English query–LLM pairs; 30 brands drawn by stratified sampling (10 Western, 10 Chinese, 10 global/mixed), balanced across market position (leader/emerging/niche) and documentation availability.
• LLMs compared (accessed via official APIs Nov–Dec 2026):
  • International: GPT-4o Search Preview, Claude Sonnet 4.5, Gemini Pro Latest
  • Chinese: Qwen3 Max Preview, DeepSeek V3.2, Doubao 1.5 Thinking Pro
• Queries: identical pure-English queries across models, covering informational, comparative, and positional intents; Zhizibianjie tested with 32 related queries as focal case.
• Design: quasi-experimental cross-LLM comparison controlling for query language and inference parameters to isolate training-data geography effects.
• Operational constructs (formalized in paper):
  • Cultural Encoding CE = P(Mention | Chinese LLM) − P(Mention | International LLM)
  • Data Moat depth Db = Σ wi · Qi (weighted sum of content types)
  • Existence indicator ELLM_b (1 if Db ≥ threshold τ, else 0)
  • Algorithmic Omnipresence AOb = average mention indicator across LLMs × query types
• Statistical findings (selected):
  • Overall mention rates: Chinese LLMs 88.9% vs International 58.3% (χ2 = 226.60, p < .001, ϕ = 0.34).
  • Zhizibianjie: 65.6% (Chinese LLMs) vs 0% (International) over 32 queries (χ2 = 21.33, p < .001, ϕ = 0.58).
• Additional analyses: sentiment comparisons and query-type moderation (posited and partially evaluated).

Implications for AI Economics

• New non-price market barrier: Training-data geography creates an algorithmic, invisible market-access barrier (Existence Gap). This functions like an entry barrier distinct from traditional costs, regulation, or distribution networks.
• Strategic asset redefinition: AI-visible content (Data Moats) becomes a rivalrous strategic asset for market positioning. Firms with richer multilingual, technical, and community content gain durable advantages in discovery—potentially increasing concentration and first-mover benefits in AI-mediated markets.
• Changes to competitive dynamics and investment priorities:
  • Firms must invest in Generative Engine Optimization (GEO) and cross-lingual content/localization to capture AI-mediated demand.
  • Returns to investments in documentation, open-source ecosystem signals, and bilingual content may rise relative to traditional marketing spend when consumers rely on LLM answers.
• Consumer welfare and market structure:
  • LLM-driven answers can narrow the set of considered vendors compared to traditional search—raising risks of under-exposure for quality entrants and potential welfare losses from reduced choice.
  • Geographic biases in training corpora can skew global competition and trade patterns; domestic incumbents may enjoy protected “algorithmic home markets.”
• Policy and measurement implications:
  • Antitrust/regulatory agencies should consider training-data composition and model-induced visibility when assessing market power and exclusionary effects.
  • Transparency and diversity requirements for training data (or targeted inclusion of cross-cultural sources) could mitigate Existence Gaps.
  • Empirical measurement of Algorithmic Omnipresence and Data Moat depth should become part of market analysis in digital platform regulation.
• Research directions for AI economics:
  • Incorporate algorithmic visibility as an endogenous variable in models of entry, pricing, and consumer search costs.
  • Quantify welfare effects of Existence Gaps (consumer surplus losses) and model optimal investment in GEO vs. traditional marketing.
  • Study dynamic effects: how Data Moats evolve, imitation limits, and whether equilibria lead to entrenched winners across language ecosystems.

Summary judgment: The paper documents a robust, practically significant phenomenon—Cultural Encoding and the Existence Gap—with clear theoretical framing (Data Moat as a VRIN resource) and actionable managerial prescriptions. For economists, it highlights that training-data composition is an essential determinant of market structure and competitive advantage in the age of generative AI.