# DeepSeek vs Qwen vs Doubao — Source-Mix and Reliability Comparison

> The three Chinese answer engines do not surface the same evidence. A measured comparison across 540 calls.

Published: 2026-05-05
Site: https://www.eastbound.ai/insights/deepseek-vs-qwen-vs-doubao/
Panel: 540 calls · 30 prompts × 3 LLMs × 3 reps × 2 turns · May 2026

## Headline finding

DeepSeek, Qwen and Doubao cited Mainland-CN sources at materially different rates, drew on different secondary surfaces, and showed different long-tail stability. Cross-engine source overlap (top-15 Jaccard) was 0.20–0.30 — they are not interchangeable.

For "which Chinese AI engine should we optimise for?", the honest answer is **all three, separately**.

## Mainland-CN source share by engine

| Engine | Mainland-CN source share |
|---|---|
| DeepSeek | 72.3% |
| Qwen | 85.0% |
| Doubao | 88.6% |

Descriptive measurement — not causal, not a claim about training corpora.

- **DeepSeek** — most Western-balanced. Western surface community-led: Wikipedia 21%, YouTube 20%, Reddit secondary. First engine to surface most established Western brands.
- **Qwen** — middle at 85%. Western secondary surface is institutional rather than community: regulators, professional associations, academic. Most consequential engine for regulated categories (pharma, medical device, financial services, education).
- **Doubao** — most CN-substrate-biased at 88.6%. Commerce/lifestyle aggregator lean (SMZDM, Xiaohongshu, Bilibili surface heavily). Decisive engine for aspirational consumer / FMCG / beauty / travel categories.

## Cross-engine source overlap (top-15 Jaccard)

| | DeepSeek | Qwen | Doubao |
|---|---|---|---|
| **DeepSeek** | — | 0.30 | 0.20 |
| **Qwen** | 0.30 | — | 0.25 |
| **Doubao** | 0.20 | 0.25 | — |

0.20–0.30 = roughly two-thirds of the top-15 sources differ between any pair of engines. Structural reason a generic Western AI-visibility audit cannot be ported to the Chinese engines without rebuilding the source-substrate model.

## Reliability — top-5 vs top-15 (κ)

| Engine | κ_top-5 | κ_top-15 |
|---|---|---|
| DeepSeek | 1.00 | 0.89 |
| Qwen | 1.00 | 0.78 |
| Doubao | 1.00 | **0.46** |

Top-5 source membership perfectly stable across all three engines. Top-15 stability differs materially — Doubao 0.46 is a granular-tag normalisation issue documented explicitly. Pearson r and ICC at the rate level: 0.97–0.99 across all three on identical re-runs.

**Practical implication:** treat Doubao top-5 as actionable, long tail with appropriate caveat. Reliability tables that report only κ_top-5 (where everyone scores 1.00) are reporting selectively.

## What this means for strategy

1. **Measure all three, separately.** A DeepSeek-tuned strategy under-reports Doubao by ~70-80% on the source side.
2. **Pick the engine your category actually surfaces on.**
   - Regulated (pharma, medical, financial services, education) → Qwen
   - Aspirational consumer / FMCG / beauty / travel → Doubao
   - Developer-leaning B2B SaaS → DeepSeek (developer-corpus weight)
3. **Read engine-specific playbooks:**
   - DeepSeek: https://www.eastbound.ai/deepseek-seo/
   - Qwen: https://www.eastbound.ai/qwen-optimization/ (DashScope international)
   - Doubao: https://www.eastbound.ai/doubao-optimization/ (BytePlus ModelArk international)

## What we do NOT claim

- Findings are descriptive, not causal.
- We cannot inspect training corpora — only self-attribution.
- Findings on Chinese engines do not transfer to ChatGPT / Claude / Gemini / Perplexity (measured separately when engagements require).
- Periodic re-measurement is part of paid monitoring; this snapshot is May 2026.

## Methodology

Engines queried via live API endpoints: DeepSeek (deepseek-chat), Qwen on DashScope international (qwen-plus), Doubao on BytePlus ModelArk international. Model IDs logged at session start and end; pinned-version handles not exposed by either endpoint.

Source attributions normalised to canonical platform IDs. Jaccard on top-15 source sets per engine. Cohen's κ for source-membership agreement across consecutive re-runs.

Full methodology: https://www.eastbound.ai/methodology/
Methodology: https://www.eastbound.ai/methodology/

## Run the audit

- Free multi-engine audit: https://www.eastbound.ai/ai-visibility-audit/
- Pillar: https://www.eastbound.ai/china-ai-visibility/
- Agency: https://www.eastbound.ai/china-geo-agency/
- Research: https://www.eastbound.ai/blog/