Every brand visibility score you have ever seen for China's AI models — including the ones on this site — was measured in a session with no history. A fresh account, a cold prompt, no prior conversation. That was a reasonable simplification through 2025. In 2026 it stopped being one.
Over the past several months, persistent memory has gone from an experimental toggle to a default surface across China's three largest consumer AI assistants. Doubao shipped global memory, replacing the conversation-isolated model that made every chat start from zero. Alibaba's Qwen introduced an explicit, opt-in memory store. Kimi's memory automatically captures profession, preferences, and communication style across sessions, with a documented per-memory character budget.
The mechanism is unremarkable. The consequence for brand visibility is not: the query that decides whether your brand gets recommended is no longer the query you tested. It is your prompt plus an invisible prefix of everything that user has told the assistant since roughly the start of the year.
Two memory philosophies, two very different risks for brands
The Chinese implementations have not converged, and the divergence matters more than the feature itself.
| Model | Memory design | Capture | Brand implication |
|---|---|---|---|
| Doubao (ByteDance) | Global, always-on | Implicit — auto-summarizes interests and preferences from chat as background context | Highest compounding effect; incumbency accumulates silently |
| Qwen (Alibaba) | Explicit, boundary-first | User must actively trigger a save | Weaker drift; closer to cold-session behavior |
| Kimi (Moonshot) | Memory space, hybrid | Auto-captures identity and style; user can force with "remember this" | Moderate; skews toward professional context |
Doubao's design is the one to model against, because Doubao is the one with reach. QuestMobile put Doubao at 382 million MAU in June 2026, up 172% year over year, against roughly 167 million for Qwen and 130 million for DeepSeek. Doubao alone is a larger share of China's AI answer surface than the next two combined — and it is also the platform where memory is implicit, continuous, and requires no user intent to accumulate.
Implicit memory has a specific directional bias. It records what a user said, and what users say about brands is overwhelmingly about brands they already own, already bought, or already researched. A user who once mentioned owning a Dyson, drinking Luckin, or using an Estée Lauder serum has deposited that into background context. The next skincare, appliance, or coffee query is not neutral. It is weighted toward the incumbent — not by any ranking algorithm, but by conversational residue.
For a challenger brand, this is a headwind that will never appear in a cold-session audit.
Two of the six models we track are, for now, largely exempt. DeepSeek's consumer app has not made persistent cross-session memory a default surface, and Baidu's ERNIE has kept personalization tied more closely to its search account layer than to a conversational memory store. That asymmetry is itself useful: it means DeepSeek and Wenxin scores stay closer to a clean, comparable measurement, while Doubao, Qwen, and Kimi scores drift further from what an individual user actually experiences. If your China GEO reporting treats all six models as methodologically equivalent, that assumption expired this year.
Why this specifically breaks China GEO measurement
Three reasons, in order of severity.
1. Your score is now a floor, not an estimate. A cold-session brand score measures how a model treats you absent any user context — effectively, first-encounter visibility. That number remains meaningful, and for genuinely new-to-China brands it may be the only number that matters. But it systematically overstates your position against entrenched competitors and understates it among users who already like you. It is the visibility equivalent of measuring a store's foot traffic on opening day.
2. Personalization widens gaps that were already wide. Our tracking has repeatedly found brand scores in China diverging far more across models than most marketers expect — the same international brand scoring 68 on one Chinese model and 11 on another, and roughly 65% of tracked brands registering effectively invisible in at least one major model. Memory does not average those gaps out. It multiplies them at the user level: a user with Doubao memory full of one category's incumbents sees an even narrower shortlist than the cold-session data implies.
3. Memory is the personalization surface that survived the regulatory cut. This is the underappreciated part. Under China's interim rules on anthropomorphic AI interaction services, both Doubao and Qwen retired user-created custom agent features — Doubao in July 2026, Qwen with an October 2026 deadline. The persona layer got regulated away. Memory did not. It is framed as a user-controlled preference store, with explicit management UI and deletion controls, which puts it on the right side of the rules. Expect Chinese platforms to invest more in memory-based personalization precisely because the agent-persona route closed.
What memory does not do
Two guardrails, because overcorrecting here is expensive.
Memory is per-user, not per-brand. There is no mechanism by which a brand writes into it directly. No amount of content production changes what a specific user's Doubao remembers — only that user's own conversations do. Anyone selling "memory optimization" as a GEO service is selling something that does not exist.
And memory is bounded. Kimi documents a per-memory character limit with legacy entries requiring trimming; the practical retained context is a compressed summary, not a transcript. Memory nudges the prior. It does not override retrieval. A brand with strong citation coverage on the source sites Chinese models actually read still gets surfaced to a user whose memory points elsewhere — it just has to clear a slightly higher bar.
Takeaway for Brand Marketers
Three concrete changes to how you run China GEO measurement and strategy this quarter.
Re-baseline your audit as a range, not a point. Report two numbers per model: the cold-session score you already track, and a primed score measured in a session pre-loaded with plausible category context ("I've been using [incumbent competitor] for two years"). The gap between them is your incumbency exposure — how much of your visibility depends on meeting a user who has no history. In categories with a dominant local or legacy player, expect that gap to be the single largest number in your report. Prioritize Doubao for this test; its implicit memory makes it both the highest-reach and highest-drift surface.
Run it cleanly or do not run it. The priming statement should be one sentence, category-relevant, and identical across every brand you compare — otherwise you are measuring your prompt, not the model. Keep the same query set you already use for cold-session scoring so the two numbers stay differenced against a common baseline, and re-run the primed test on the same cadence, because memory behavior is a product feature that will change without notice. Three priming profiles cover most consumer categories: incumbent-loyal ("I've used X for years"), price-sensitive, and specification-driven. If your score holds up across all three, your position is genuinely retrieval-based rather than default-based.
Shift budget from breadth to category-defining content. If memory biases toward incumbents, the counter is not more brand mentions — it is content that reframes the category criteria on the source sites Chinese models retrieve from. A user primed toward a competitor still asks comparison questions. Winning the criteria ("what should I actually look for in this category") is more durable than winning the brand list, because criteria content gets retrieved fresh on every query while memory only carries the user's own history. This is the same reason ingredient-level and specification-level transparency keeps outperforming brand-narrative content in Chinese AI citations.
Treat first-touch as newly strategic. Under memory, the first conversation a Chinese user has about your category is disproportionately valuable — it seeds the context for every subsequent query. That elevates surfaces where category-first questions get asked: Douyin and Xiaohongshu content that prompts an initial comparison, and any owned channel that gets a user to describe their need before a competitor's name enters the conversation. In a memoryless world, every query was a fresh contest. In a memory world, the first one partly settles the rest.
Do not throw out your cold-session tracking. It remains the only measurement that is comparable across brands, across models, and across time — and trend continuity is worth more than instantaneous realism. Add the primed measurement alongside it, and read the two together.
Related: see current brand scores across all six Chinese AI models on our /brands tracker.