Doubao reportedly weights week-old content 3.2x over 30-day-old content, and decays untouched pages ~70% per month. Your China brand score isn't a standing — it's a snapshot of a decaying curve.

The Half-Life Problem: Why Chinese AI Models Forget Your Brand Content Faster Than You Publish It (2026)

Most international brands treat their China content as an asset. Chinese AI models treat it as a perishable good.

Practitioner data circulating through China's GEO service industry in mid-2026 puts a number on it: content published within the last seven days is crawled and weighted roughly 3.2x more heavily than content older than 30 days, and material that sits untouched for more than three months decays at approximately 70% per month in retrieval weight. Separate teardowns of Doubao's citation behavior report that content published within 1–2 weeks carries the highest weight of all, and that pages carrying live or recent data are cited 40%+ more often than otherwise-comparable static pages.

These figures come from Chinese GEO agencies reverse-engineering platform behavior, not from ByteDance disclosures — treat them as directional rather than exact. But the direction is consistent across every practitioner source we reviewed, and it lands on a fact most brand teams have not internalized: in Chinese AI search, your brand's visibility has a half-life.

That changes what a brand score actually measures. It is not a standing. It is a snapshot of a decaying curve.

The Asset Class Chinese AI Discounts Hardest

Consider what a typical international brand's China content stack actually looks like:

A brand story page written at market entry and revised every 18 months
A product catalog updated when the SKU lineup changes
A press release cadence of six to twelve items a year
A Baike entry edited rarely, by design
Agency-produced campaign content that goes live, runs, and stops

Every one of these is optimized for permanence. That was the correct instinct for a decade of Baidu SEO, where domain authority and backlink accumulation rewarded assets that aged well. It is close to the worst possible instinct for a retrieval layer that applies a 70%-per-month decay to anything untouched for a quarter.

The practical consequence is that a brand can be materially present in China — real stores, real distribution, real media coverage in 2024 — and still be functionally invisible in a Doubao answer generated in September 2026, because nothing in its owned or earned footprint has moved recently enough to clear the freshness threshold.

Industry timing benchmarks reinforce how fast this cycle runs. New content typically enters a crawl window within 24–48 hours, first citations appear in 5–7 days, and consistent "authoritative source" status generally takes 2–3 months of sustained publishing. Practitioners report that sites maintaining 3–5 posts per week are meaningfully more likely to be selected as citation sources. That is a publishing cadence closer to a newsroom than to a brand marketing calendar.

Decay Rates Are Not Uniform — And That's the Useful Part

The more actionable finding is that freshness sensitivity varies sharply by model.

Model	Freshness sensitivity	Practical implication
Doubao	Highest — trend-responsive, tightly coupled to Douyin/Toutiao recency logic	Visibility rises and falls with recent content velocity
Qwen	High — frequent knowledge refresh, commerce-linked signals	Catalog and pricing staleness shows up quickly
Yuanbao	High — WeChat ecosystem freshness	Recent official-account output matters disproportionately
DeepSeek	Lower — less time-sensitive to new information	Scores hold longer on stable entity knowledge
Wenxin / ERNIE	Mixed — index-driven	Recency matters less than source authority
Hunyuan	Moderate	Ecosystem-linked, less trend-reactive

This spread creates a diagnostic that brand teams can use immediately, and that a single-model score cannot produce.

If your Doubao score falls while your DeepSeek score holds, your visibility was being carried by fresh third-party content, not by stable entity knowledge. The model that forgets first tells you what your position was made of. Recency-weighted models are effectively measuring your current content velocity; recency-insensitive models are measuring what the ecosystem has durably concluded about you. A brand strong on both has genuinely established itself. A brand strong only on Doubao is renting its position from whatever campaign just ran.

The inverse divergence matters too. A brand that scores well on DeepSeek but poorly on Doubao is usually one with legacy authority and no current momentum — often a market-entry-era brand that has since gone quiet in Chinese-language channels.

Why This Compounds With Multi-Hop Retrieval

Doubao's answer generation increasingly follows a "search while thinking" loop: decompose the question into sub-questions, search the first, read and reason over the results, decide whether the evidence is sufficient, generate the next query, and repeat until a stopping condition is met.

Recency weighting does not apply once in that flow. It applies at every hop.

A brand that clears the freshness bar has multiple chances to enter the evidence set across a multi-step retrieval chain. A brand that does not gets filtered repeatedly — and the compounding is multiplicative, not additive. This is a plausible mechanism behind an effect brand teams observe and struggle to explain: visibility in Chinese AI search tends to be lumpy. Brands are either well-represented across many query formulations or nearly absent across all of them, with less of the gradual middle that Western search results produce.

Not Every Category Decays at the Same Speed

Freshness weighting interacts with query type, and the interaction decides how much a brand should actually worry.

Categories where the underlying facts move — pricing, availability, model years, promotions, openings and closings, regulatory status — inherit the full decay penalty. Automotive, consumer electronics, travel and hospitality, and anything sold through a fast-moving e-commerce calendar sit here. A hotel brand whose most recent Chinese-language content predates a renovation, or an electronics brand whose latest indexed spec sheet describes last year's model, is not merely stale; it is being actively out-ranked by whichever competitor published this month.

Categories built on comparatively stable attributes — heritage luxury, spirits, established pharmaceutical brands — decay more slowly, because the questions users ask ("is this brand prestigious," "what is it known for") map onto durable entity knowledge rather than retrieved current documents. This is part of why luxury houses can post strong scores on recency-insensitive models with almost no recent Chinese-language publishing behind them.

The trap is assuming category membership grants permanent immunity. It does not. The moment a query shifts from identity ("what is this brand") to decision ("which should I buy in 2026, and at what price"), the retrieval path shifts with it — and decision queries are precisely the ones that convert. A luxury brand can hold a strong identity score and still lose every purchase-intent answer to a competitor with fresher merchandising content.

What Actually Resets the Clock

Not all updates are equal. Based on how these systems reportedly weight signals, the interventions worth prioritizing:

Carries real weight

Genuinely new data, figures, or dated case material — not a modified timestamp on unchanged text
Explicit, visible publication and revision dates on every page
Rapid response to industry developments, ideally within 48 hours
Quarterly substantive rewrites of core brand and product pages, monthly data refreshes
Sustained Chinese-language third-party coverage rather than isolated campaign bursts

Carries little or none

Republishing identical content under a new date
Press release syndication without earned pickup
English-language corporate updates with no Chinese equivalent
Campaign content that stops the day the media buy ends

The uncomfortable budget implication: a brand that spends heavily on one large annual content push and nothing in between will show a visibility spike followed by a three-to-four-month slide — and will typically measure the spike, not the slide, because the measurement happens right after the campaign.

Takeaway for Brand Marketers

Three changes to how you run China AI visibility:

1. Measure the trend, not the point. A single brand score taken once is nearly meaningless under a 70%-per-month decay regime. Track the same queries across the same models at a fixed cadence, and read the slope. A score of 61 falling from 78 and a score of 61 rising from 44 describe opposite situations.

2. Use cross-model divergence as a diagnostic. Run recency-sensitive models (Doubao, Qwen, Yuanbao) against recency-insensitive ones (DeepSeek, Wenxin) on identical queries. The gap between them tells you whether your position is durable or rented. This is the highest-value read you can get from multi-model tracking, and it is invisible if you only test one platform.

3. Budget for cadence, not campaigns. If the freshness data is even approximately right, the marginal RMB spent on a sustained monthly Chinese-language publishing rhythm buys more AI visibility than the same RMB concentrated in a quarterly push. That is a genuinely different allocation than most China marketing plans currently reflect — and it requires holding budget back from the campaign line to fund the always-on line.

The brands that will hold position in Chinese AI search through 2027 are not the ones with the best brand story page. They are the ones that gave the models something new to read this month.