The 60% Threshold: Why Publishing More Content Lowers Your Brand Score in China's AI Search (2026)

There is a number circulating in Chinese GEO research this month that should stop every brand team currently paying an agency by the article: when low-authority sources make up more than 60% of the material carrying your brand, your overall citation priority in Chinese AI models drops by roughly 18%.

Not plateaus. Drops.

That finding, reported in tracking of 12,000-plus Chinese enterprises, sits alongside a second one from the same body of work: about 72% of corporate press releases have never been cited by any major Chinese model at all, and low-quality releases are cited less than 5% of the time. Put those two numbers together and the conclusion is uncomfortable but unavoidable — the dominant GEO service model in China, which prices work by content volume across as many placements as possible, is not merely inefficient. Past a threshold, it is actively subtracting from the brand score it was hired to raise.

For international brands running China AI visibility programs, this reframes the budget question entirely. The issue is no longer "are we publishing enough?" It is "are we publishing so much, in so many low-tier places, that we have diluted ourselves out of the answer?"

The weights finally got published — and they explain the penalty

Until recently, brand teams optimizing for Doubao, DeepSeek, Qwen, Kimi, and Wenxin were working against a black box. That changed in September 2026, when an industry report on China's AI search ranking algorithms and platform governance consolidated estimated ranking-factor weights derived from black-box testing across 120 cross-border brands and more than 3,000 queries.

The weights are estimates, not disclosed algorithms — the researchers say so explicitly, and any brand team using them should hold them loosely. But the shape of the distribution is the important part, and it is consistent with what practitioners have been observing in score movements all year.

Read the top three rows together: entity authority, citation quality, and E-E-A-T account for an estimated 58% of the outcome. Now read the bottom three: recency, technical accessibility, and multimodal richness total about 12%.

Nearly every GEO retainer sold in China is priced against that bottom 12%. Publishing cadence maps to the 5% recency factor. Site technical fixes map to 4%. Adding images and video to existing pages maps to 3%. These are real factors and they are not worthless — but none of them independently moves a brand score, because none of them can compensate for a weak entity.

The 58% at the top is slow, expensive, and cannot be produced at volume. That is precisely why so little of the market sells it.

Why volume doesn't just fail — it reverses

The mechanism behind the 18% penalty is not mysterious once you understand how Chinese models tier their sources.

Testing across Doubao, DeepSeek, Kimi, and Qwen shows a consistent four-level source hierarchy. Tier one is authoritative official media and government sites — Xinhua, People's Daily, ministry portals — which get top priority and resist downweighting. Tier two is academic journals, industry white papers, and licensed institutions, with DeepSeek and Kimi weighting these noticeably more than the others. Tier three is mainstream commercial media — 36Kr, Huxiu, Caixin — where competition is fiercest. Tier four is self-media and UGC: WeChat official accounts, Zhihu posts, personal blogs, which carry low weight and are frequently filtered or folded out of answers entirely.

A volume-based GEO program produces tier-three and tier-four material almost exclusively, because that is the only tier you can buy at scale and on a monthly schedule. Every batch shifts the brand's source mix further down the hierarchy. Cross 60%, and the model's assessment of the brand as a source entity degrades — which is exactly the 25% factor at the top of the table.

There is a second, harsher filter running underneath. Chinese models apply a credibility check that zeroes out weight entirely on content showing 软文 characteristics — promotional copy dressed as objective coverage. Not a downgrade. Zero. Semantic coverage does not rescue it. Mass-produced brand content optimized to include brand keywords is close to a worked example of what that filter is built to catch, which is why 72% of corporate releases have never been cited once.

So the volume program produces three effects at the same time: it fails to earn citations, it dilutes the source mix toward the penalty threshold, and a share of its output is pattern-matched as marketing and discarded. The brand pays for all three.

The regulator just named the tactic

In April 2026, the Cyberspace Administration of China launched a four-month "Qinglang" campaign against AI application abuses, structured in two phases of seven problem categories each — fourteen in total, spanning model filing compliance, training-data poisoning, and generated-content labeling.

Buried in the data-poisoning category is a detail that the GEO services industry has largely not discussed with its clients. The CAC defines data poisoning across three methods: tampering with training corpora, fabricating authoritative data, and using GEO techniques for malicious marketing. Tutorials and tools sold on e-commerce platforms for these methods were named as enforcement targets.

The enforcement is not theoretical. Shanghai's cyberspace authority reported that since late April, 17 key platforms cleared or intercepted more than 4.87 million items — de-labeling tools, poisoned data, AI-generated rumors, low-quality content — and removed over 18,000 violating accounts and 14,000 agents. Regulators have also flagged paid-citation contamination of AI answers as a specific target, requiring platforms to cross-verify sources and label citation links.

The practical reading for an international brand: a tactic your agency describes as "authority building" or "source matrix construction" may sit inside a category a Chinese regulator has enumerated as poisoning. Brands with China entities carry the compliance exposure, not the offshore vendor that shipped the content.

What this means for brand marketers

Audit your source mix before you commission another word. Pull every placement carrying your brand from the last twelve months and tier it against the four-level hierarchy. If tier three and tier four exceed 60% of the total, your next batch of content makes your score worse, not better. This is a one-afternoon exercise and it is the highest-value thing most China GEO programs could do this quarter.

Reprice the retainer against the weight table. If the deliverable is a monthly article count, you are buying the 5% recency factor at the price of the 25% authority factor. Ask any vendor to show, specifically, which of the top three factors their work moves and what evidence they have that it moved. "Mentions increased" is not that evidence — mention counts rise from volume even as citation priority falls.

Build fewer, higher-tier assets. One industry white paper co-published with a licensed institution, one dataset that trade media cite independently, one properly structured Baike entity with consistent facts across sources — these move entity authority and citation quality, the 43% at the top. They cannot be produced twelve times a year, and a program that promises they can is selling something else.

Strip promotional register from everything. The softwen filter zeroes weight regardless of how complete the content is. Content that reads as objective, carries real data, names its sources, and is traceable to a single verifiable origin clears the filter. Content built around brand keyword density does not.

Ask the compliance question explicitly. Get your vendor's methodology in writing and check it against the CAC's fourteen categories — particularly whether any part of it involves fabricated authority signals or coordinated multi-site placement designed to simulate consensus. The Doubao ad system moves to full rollout through Juliang Engine in Q4 2026, which means paid brand placement will soon have a legitimate, labeled channel. The window where gray-zone organic tactics look like a reasonable substitute is closing.

The blunt version: in China's AI search, brand visibility behaves less like search ranking and more like credit. You do not improve it by borrowing more often from weaker sources. The weight table simply put a number on what the models were already doing.

Related: see how individual brands score across Doubao, Kimi, DeepSeek, Qwen, Wenxin, and Hunyuan on the hubGEO brand index.