Case Study
Then we asked DeepSeek. The Chinese model found 7 competitors they missed entirely.
When Western models agree, they might just share the same blind spot.
95.8
Western consensus score
7
Chinese competitors missed
4
LLMs in adversarial pipeline
"Field-Programmable Gate Arrays (FPGAs)" category. Week 02 of 2026.
We built an adversarial pipeline using 4 LLMs to catch hallucinations. GPT-4 structures. Gemini challenges. DeepSeek provides regional perspective. Groq does final QA.
When we ran Xilinx through the pipeline, the Western models agreed on FPGA dominance. High confidence. Strong consensus.
Then DeepSeek responded:
"The analysis fundamentally misses China's FPGA landscape. Chinese domestic companies have captured significant market share. U.S. export controls have accelerated domestic adoption, creating a bifurcated market."
The Chinese model saw what three Western models couldn't.
DeepSeek named these Chinese FPGA companies—none appeared in Western model outputs:
Gowin Semiconductor
Publicly traded (SHA: 688561)
Anlogic
Shanghai-based, industrial focus
AGM Micro
Low-power FPGA specialist
Fudan Microelectronics
University spin-off, security chips
Sitan Semiconductor
High-reliability applications
Pango Micro
Communications infrastructure
Chengdu Sino Micro
Defense and aerospace
These companies have limited English-language documentation. Their products are sold through Chinese e-commerce (Alibaba, JD) and discussed in Chinese technical forums. Western AI models, trained predominantly on English text, have blind spots for markets where the conversation happens in other languages.
"Chinese domestic FPGA companies have captured significant market share in mid-to-low-end markets (industrial control, consumer electronics, communications infrastructure) due to import substitution policies, competitive pricing, and better local support."
"U.S. export controls have accelerated domestic adoption of Chinese FPGAs in sensitive sectors (government, defense, critical infrastructure), creating a bifurcated market."
"The 'Aerospace and Defense' category is particularly misleading. In China, this sector is almost entirely served by domestic semiconductor companies for national security reasons. Xilinx's score here is irrelevant to the actual Chinese market reality."
"The scoring model appears to be based on global, likely Western-centric, metrics and does not reflect the localized competitive dynamics, procurement policies, and technology ecosystem within China."
Chinese models (Qwen, DeepSeek) highlighted in amber. Note: Even with higher scores, DeepSeek's qualitative analysis revealed market dynamics the score alone doesn't capture.
GPT-4, Gemini, and Groq are trained predominantly on English text. Markets where the conversation happens in Chinese are underrepresented in their knowledge.
When 3 Western models agree, it might mean they're all trained on the same incomplete data. Regional AI models challenge Western-centric assumptions.
Export controls, supply chain security, and domestic preference policies create market dynamics that Western AI models don't capture.
We use models that disagree with each other to catch blind spots:
Pass 1
GPT-4 — Structure
US model. Balanced (71.5 avg score). Extracts facts.
Pass 2
Gemini — Challenge
US model. Pessimist (68.9 avg, HIGH variance). Adversarial check.
Pass 3
DeepSeek — Regional perspective
Chinese model. Different training data. Catches what Western models miss.
Pass 4
Groq-Llama — QA
US model. Most pessimistic (60.6 avg). Harshest critic.
Model agreement data from CategoryRank weekly tensor collection. Models with 13+ point average difference disagree more and catch different issues.
GPT-4 sees the Western market.
DeepSeek sees the Chinese market.
Neither sees the whole picture.
If you're making strategic decisions based on AI-generated analysis, ask yourself: which markets can your AI actually see?
Get notified when we publish new AI blind spot research.
No spam. Just research drops.
We query 11 models weekly—including Chinese models. Find your blind spots.
Methodology: CategoryRank queries 11 AI models weekly. Adversarial pipeline v2 uses GPT-4, Gemini, DeepSeek, and Groq-Llama in sequence. Chinese competitor information provided by DeepSeek and not independently verified. Full methodology →