# YouTube AI Visibility — 20% DeepSeek mention rate, despite the geo-block

**Canonical:** https://www.eastbound.ai/insights/youtube-ai-visibility/
**Updated:** 2026-05-07
**Panel:** DeepSeek, n=1,620 responses, luxury handbag category, May 2026

In our 1,620-response luxury-handbag prompt panel against DeepSeek (May 2026), YouTube surfaced in **20%** of responses — within one percentage point of Wikipedia (21%). YouTube is geo-blocked from typical Mainland-CN consumer access, which makes the result counterintuitive: the engine cites a platform many of the answer's intended consumers cannot directly visit.

## The geo-block paradox

YouTube has been blocked in Mainland China since 2009. A consumer searching DeepSeek from Shanghai cannot click a YouTube link in the answer and watch the video. By any obvious commercial logic, the engine should not surface YouTube — the citation is not actionable for the user.

But the engine does surface YouTube, at 20% of responses. Three explanations are consistent:

- **Training-time exposure.** DeepSeek's training corpus included Western web data fetched before deployment. YouTube channel descriptions, video titles, transcripts where available, comment-thread summaries — all part of that corpus. The engine learned that YouTube is a common artefact of brand-discovery and product-review information landscapes, and surfaces it at retrieval time.
- **Cross-language re-publication.** YouTube content is mirrored, transcribed and discussed on Mainland-CN platforms — Bilibili clips, Weibo discussion threads, Zhihu posts that link to or describe YouTube videos. The engine may cite the original YouTube URL because it is the canonical anchor for content re-published across the substrate it can access.
- **Citation hygiene.** Engines tend to credit primary sources rather than the platforms that re-host them.

**What this is not.** Not evidence that YouTube content is "absorbed during training" in some special way — we cannot inspect training corpora. Not evidence that the engine "trusts YouTube" — the study measures what surfaces, not how the engine weights sources internally. Both inferences would overclaim. The defensible claim is the simpler one: YouTube URLs surface as citation anchors at meaningful frequency, despite the geo-block.

## What this means for global brands

YouTube content is part of the AI-citation surface for Mainland-CN consumers even though it is not part of the consumer's direct browsing surface. A brand that ships substantive YouTube content — product reviews, channel-owned explainers, third-party tutorials, executive interviews — is contributing to the source pool DeepSeek (and likely Qwen and Doubao) draw from.

Three caveats:

- **The citation is not actionable for the consumer.** The brand cannot expect Mainland-CN viewers to click through. Contribution is to the answer's evidence base, not to direct viewer traffic.
- **Channel-owned has different weight from third-party.** The 20% is aggregate; we have not yet decomposed into "brand-owned channel" vs "third-party reviewer". Anecdotally, third-party content surfaces more often.
- **Subtitle and metadata quality matter more than production value.** AI engines cannot watch a video; they parse titles, descriptions, transcripts, and comment-thread structure. A well-described video with a thorough caption and accurate auto-transcript is more citable than a beautiful video with a one-line description.

## What to actually ship on YouTube for Mainland-CN AI surfacing

The optimisation targets are textual, not video-craft:

1. **Long-form, descriptive titles.** Not "Bag Review" — "Hermès Birkin 25 review: leather wear, hardware patina, resale considerations after 5 years".
2. **Substantive descriptions.** 300+ words covering what the video actually contains, named entities, products compared, links to related references.
3. **Accurate captions and auto-transcripts.** Edit YouTube's auto-transcript for accuracy. Misspellings of brand names propagate.
4. **Cross-link to and from durable web pages.** YouTube descriptions linking to a brand's encyclopedia-style content on its own site build the entity graph. Inverse links — from blog posts to YouTube videos — reinforce it.
5. **Pursue third-party coverage, not just self-published.** A video on a recognised reviewer's channel is more citable than the same content on your own channel.

## Caveats and uncertainties

- 20% is from one engine, one category, n=1,620, May 2026.
- Reliability statistics: top-5 κ = 1.00; top-15 κ = 0.46. YouTube consistently in top-5 — high-confidence headline. Specific channels in long-tail positions are less stable.
- Per-engine decomposition (Qwen, Doubao) queued, not yet measured.
- We are not measuring sales, conversion, or attributable revenue from YouTube-mediated AI citations.

## Related reading

- [Wikipedia AI visibility (DS 21%)](https://www.eastbound.ai/insights/wikipedia-ai-visibility/)
- [Zhihu AI visibility](https://www.eastbound.ai/insights/zhihu-ai-visibility/)
- [SMZDM AI visibility](https://www.eastbound.ai/insights/smzdm-ai-visibility/)
- [Xiaohongshu AI visibility](https://www.eastbound.ai/insights/xiaohongshu-ai-visibility/)
- [DeepSeek vs Qwen vs Doubao comparison](https://www.eastbound.ai/insights/deepseek-vs-qwen-vs-doubao/)
- [China AI visibility pillar](https://www.eastbound.ai/china-ai-visibility/)
- [Methodology](https://www.eastbound.ai/methodology/)