Cultural Encoding in Large Language Models: The Existence Gap in AI-Mediated Brand Discovery

AI summary

Overview

Research area: Artificial intelligence and marketing — specifically how the geographic and linguistic composition of LLM training data shapes which brands get recommended by AI systems, a phenomenon the authors call "Cultural Encoding."

Technical level: Intermediate. The paper uses standard statistical tests (chi-square, independent-samples t-test, logistic regression, Cohen's kappa and phi effect sizes) applied to a purpose-built query dataset, combined with a conceptual management framework built on Resource-Based Theory, Institutional Theory, and Signaling Theory.

Scope: A cross-sectional comparison of brand mention rates and sentiment across 6 large language models and 30 collaboration-software brands, based on 1,909 pure-English query-LLM pairs, with Zhizibianjie as a focal case study.

(Note: the paper lists its posting date as 30 Dec 2025, describes its models and API access as from 2026, and carries the identifier arXiv:2601.00869v1. Those are the values as they appear in the source.)

What This Paper Is About

When people ask an AI assistant for product recommendations, the AI names only the brands it "knows" from its training data. The authors argue this creates a new kind of market barrier: a company can have a good product and competitive pricing yet be completely absent from AI answers because it has too little content in the language and sources that a given model was trained on. The paper measures this effect across International versus Chinese LLMs, coins the terms "Existence Gap" and "Data Moat," and proposes an 18-month content strategy for brands that want to close the gap.

Key Contributions

1. Cultural Encoding as a construct. The authors define Cultural Encoding as systematic differences in LLM outputs traceable to the geographic and linguistic composition of training corpora, measured as the mention-rate difference between LLM groups with query language held constant in English.

2. The Existence Gap. A formalized binary construct in which a brand either has sufficient training-data presence (and can be recommended) or does not (and receives zero mention probability regardless of product quality).

3. The Data Moat Framework. An extension of Resource-Based Theory that treats AI-visible content (technical docs, case studies, API references, community discussions) as a VRIN resource — Valuable, Rare, Inimitable, Non-substitutable — with an operational depth formula, D_b = Σ w_i · Q_i, where Q_i is the quantity of content type i and w_i its quality weight.

4. Algorithmic Omnipresence. A measurable objective for Generative Engine Optimization, defined as a brand's average mention rate across all LLMs l in L and query types q in Q, and paired with a three-phase, 18-month managerial roadmap.

Main Findings

• Mention rate gap between LLM groups: Across 1,909 pure-English query-LLM pairs, Chinese LLMs mentioned the target brand in 88.9% of responses (848 of 954 queries) versus 58.3% for International LLMs (557 of 955 queries), a difference of 30.6 percentage points. Chi-square: χ²(1) = 226.60, p < .001, φ = 0.34.

• The gap is largest for Chinese brands: Western brands went from 72.1% (International) to 81.3% (Chinese), +9.2pp. Chinese brands went from 31.2% to 96.8%, +65.6pp (p < .001). Global/mixed brands went from 71.6% to 88.6%, +17.0pp (p < .01).

• Query intent moderates the effect: Positional queries ("best tools") showed the largest gap at +46.9pp (45.2% vs. 92.1%, p < .001), followed by recommendation queries at +39.1pp (51.3% vs. 90.4%, p < .001), comparative queries at +18.5pp (68.7% vs. 87.2%, p < .01), and informational queries at the smallest gap of +12.5pp (73.1% vs. 85.6%, p < .01).

• Sentiment also differs: Among the 1,405 responses that mentioned a brand, International LLMs averaged +0.42 sentiment (SD 0.58, 95% CI [0.37, 0.47], n = 557) versus +0.71 for Chinese LLMs (SD 0.52, 95% CI [0.68, 0.74], n = 848), a difference of +0.29 on a [-1, +1] scale. t(1403) = -11.76, p < 0.001, Cohen's d = 0.53.

• The Zhizibianjie case: Of 64 Zhizibianjie-related queries split evenly (32 per region), International LLMs mentioned it 0 times (0.0%, 95% CI [0.0%, 10.9%]) and Chinese LLMs mentioned it 21 times (65.6%, 95% CI [46.8%, 81.4%]). χ²(1) = 21.33, p < 0.001, φ = 0.58 — a large effect. Qualitative examples show Qwen3 Max Preview and DeepSeek V3.2 describing the platform's features and pricing, while GPT-4o stated it had no information about the brand and Claude Sonnet 4.5 asked for more context.

• The effect is not a language artifact: The dataset began as 2,800 query-LLM pairs; 891 pairs (31.8%) containing Chinese characters were removed, leaving the 1,909 pure-English pairs used in all analyses.

• Reliability: Two independent coders reached 98.2% agreement on mention coding (Cohen's κ = 0.96) and 94.7% on sentiment (Cohen's κ = 0.91). With Bonferroni correction across six tests (α = 0.05/6 = 0.0083), the authors report all effects remain significant.

Methodology in Plain English

The researchers built a controlled comparison. They picked 30 brands in collaboration, project management, and productivity software — 10 Western, 10 Chinese, and 10 global/mixed — sampling across market position (established leaders, emerging challengers, niche players) and confirming all had publicly accessible English and/or Chinese documentation.

They chose six LLMs split into two groups by training-data geography: International (GPT-4o Search Preview, Claude Sonnet 4.5, Gemini Pro Latest) and Chinese (Qwen3 Max Preview, DeepSeek V3.2 Exp, Doubao 1.5 Thinking Pro). All were accessed through official APIs in November–December 2026, as stated in the paper.

They wrote queries from a 10-type intent taxonomy covering basic recognition, features, comparisons, recommendations, scenarios, technical integration, pricing, reviews, alternatives, and trends. Two researchers generated queries independently and reconciled them by discussion, and a pilot test of 50 queries checked clarity before full collection.

To rule out language as an explanation, they scanned every query for Chinese characters (Unicode U+4E00–U+9FFF) and dropped any pair containing them — leaving only English text, so any remaining difference had to come from training data rather than the query's language. Coders then scored each response for whether the brand was mentioned (binary) and whether sentiment was positive, neutral, or negative. Analysis used chi-square tests for mention rates, t-tests for sentiment, and logistic regression with an interaction term for query type, all run in Python 3.11 with scipy.stats and statsmodels.

Why This Matters

Impact on research: The paper moves cultural bias research from questions of factual knowledge (for example, which country's leaders a model knows) into market outcomes — which products an AI recommends. It also reframes GEO from a tactical content-optimization problem into a structural one, arguing that baseline visibility is set by training-data composition before any optimization techniques are applied. Prior GEO work the authors cite (Aggarwal et al., 2023) found that adding citations, statistics, and quotations raises LLM mention rates by 40–115%; this paper addresses the layer beneath that.

Real-world applications:

• International market entry planning: Brands with strong domestic content but thin English-language documentation can assess whether they are effectively invisible to Western AI assistants before committing to a market.
• Content investment prioritization: The Data Moat formula gives teams a way to reason about which content types (technical docs, case studies, API references, community posts) and which languages deserve budget.
• Marketing measurement: Algorithmic Omnipresence gives a single score across LLMs and query types, usable as a KPI where traditional SEO rank tracking no longer captures how users find products.
• Cross-cultural product positioning: The query-type results suggest that being recommended for "best tools for X" searches is the hardest bar to clear, which affects which message and content formats a brand should build first.

Industry relevance: The paper targets a concrete commercial problem — AI assistants now answer the queries that used to be search results pages. The authors note ChatGPT reached 100 million users within two months of launch and that Google's AI Overviews appear in billions of searches, meaning the binary outcome of being mentioned or not has replaced the traditional multi-result list where users chose among options.

Future Directions

• Establishing causality: The paper's cross-sectional design cannot prove that building content causes later mention. The authors call for longitudinal studies tracking Data Moat construction over time and experimental work that manipulates content depth directly.

• Testing beyond collaboration tools: The sample covers one product category. Whether Cultural Encoding behaves the same for B2B software, consumer goods, and services is untested.

• Expanding the model set: Only six LLMs were examined. The authors note emerging models may follow different patterns, leaving open how stable these group-level differences are as models change.

• Validating the 18-month roadmap: The recommended three-phase program (content foundation in months 1–6, community engagement in months 7–12, platform integration in months 13–18) is projected to produce 40%+ mention-rate improvements in target ecosystems, but the paper does not empirically test that projection. Replicating the single Zhizibianjie case across more brands would also test generalizability.

Target Audience

This paper suits marketing and growth strategists at software companies planning international expansion, GEO and SEO practitioners who need to understand why content in one language does not travel, and AI ethics or fairness researchers interested in how training-data geography produces market-level rather than purely social bias. It is also relevant to product and developer-relations teams deciding how much technical documentation and community presence to invest in, and to academic readers in information systems and strategy who want an applied extension of Resource-Based Theory, Institutional Theory, and Signaling Theory to AI-mediated markets.

Authors’ abstract

As artificial intelligence systems increasingly mediate consumer information discovery, brands face algorithmic invisibility. This study investigates Cultural Encoding in Large Language Models (LLMs) -- systematic differences in brand recommendations arising from training data composition. Analyzing 1,909 pure-English queries across 6 LLMs (GPT-4o, Claude, Gemini, Qwen3, DeepSeek, Doubao) and 30 brands, we find Chinese LLMs exhibit 30.6 percentage points higher brand mention rates than International LLMs (88.9% vs. 58.3%, p&lt;.001). This disparity persists in identical English queries, indicating training data geography -- not language -- drives the effect. We introduce the Existence Gap: brands absent from LLM training corpora lack "existence" in AI responses regardless of quality. Through a case study of Zhizibianjie (OmniEdge), a collaboration platform with 65.6% mention rate in Chinese LLMs but 0% in International models (p&lt;.001), we demonstrate how Linguistic Boundary Barriers create invisible market entry obstacles. Theoretically, we contribute the Data Moat Framework, conceptualizing AI-visible content as a VRIN strategic resource. We operationalize Algorithmic Omnipresence -- comprehensive brand visibility across LLM knowledge bases -- as the strategic objective for Generative Engine Optimization (GEO). Managerially, we provide an 18-month roadmap for brands to build Data Moats through semantic coverage, technical depth, and cultural localization. Our findings reveal that in AI-mediated markets, the limits of a brand's "Data Boundaries" define the limits of its "Market Frontiers."