Cross-border e-commerce is how most foreign brands now reach Chinese buyers. No local entity, no Chinese business licence, goods shipped from a bonded warehouse or straight from abroad. It is the cheapest door into the market. It is also the door where AI has stopped being a slide in a deck and started changing unit economics.
This article is not a general tour of artificial intelligence. It looks at the five places where AI actually moves numbers in a cross-border operation: product listings, demand forecasting, customer service, returns and fraud, and visibility inside Chinese AI assistants. Some of it works. Some of it quietly destroys conversion. I will be specific about which is which.
Written by Olivier Verot, founder of GMA. I have been based in Shanghai since 2012 and I run cross-border stores on Tmall Global, JD Worldwide and Douyin for European and American brands. Most of what follows comes from listings my teams have fixed after an AI tool broke them.
The market AI is being pointed at
The numbers justify the attention. According to China’s General Administration of Customs, 140 million people in China bought through cross-border e-commerce platforms in the first half of 2026. Total cross-border e-commerce trade reached 2.84 trillion yuan in 2025, up 4.8% year on year. That is a buyer base larger than the population of Japan and Germany combined, ordering foreign goods on their phone.
Three things about those buyers matter for anyone deploying AI. They shop almost entirely on mobile. They read reviews before they read your product page. And they treat foreign origin as a quality signal, which means a listing that sounds foreign in the wrong way, translated badly, kills the exact advantage you paid to have.
Chinese customs policy keeps making this easier. Duty-free allowances have been raised, bonded zones have expanded, and clearance times have dropped. The rules did change in 2026 though, and the annual per-person spending threshold is not what most brands assume. We covered that in detail in our guide to the 2026 cross-border e-commerce law for foreign brands.
Automatic listing translation: fast, cheap, and the fastest way to lose a sale
This is where every brand starts, and where most of the damage happens. You have 300 SKUs in English. A translation API turns them into Chinese in an afternoon for the price of a lunch. The store goes live. Traffic arrives. Nobody buys.
The reason is that a Chinese product page is not a translated Western product page. It is a different document. Western listings sell benefits in prose. Chinese listings on Tmall and JD sell through a long scrolling image stack with parameters, certifications, ingredient lists, usage scenarios and comparison charts baked into the graphics. Machine translation produces correct sentences that sit in the wrong container.
Then there are the specific failures we see every month:
- Category terms that are technically right and commercially dead. The literal translation of your category is often not the term buyers search. Get this wrong and you are invisible in on-platform search, which is where most Tmall Global purchase journeys start.
- Regulated claims translated verbatim. Words like “anti-ageing”, “medical”, “cures” or “whitening” carry legal exposure in China that they do not carry in Europe. An AI translator has no idea it just wrote you a fine.
- Sizing and units left alone. European sizing on a Chinese apparel listing without a body-measurement conversion table is a guaranteed return.
- Tone. Chinese e-commerce copy is warmer and more direct than the polished neutral register AI defaults to. It reads like a manual, and the buyer scrolls past.
The workable method is to use AI as a first draft and a Chinese copywriter as the editor, not the reverse. Let the model handle the parameter tables, ingredient lists and specification blocks where accuracy matters and style does not. Have a human write the first screen, the headline, the claim structure and the review-response templates. On the accounts we manage, that split cuts listing production cost roughly in half without touching conversion. Pure machine output does not.
Demand forecasting and bonded warehouse stock
Bonded warehouse stock is the part of cross-border that hurts when you get it wrong. Goods sitting in a Ningbo or Zhengzhou bonded zone are paid for, shipped, and not yet sold. Too little and you go dark during 618 or Double 11. Too much and you are holding cash in a warehouse on the other side of the world.
Forecasting models built on your Western sales history will fail here, because Chinese demand is festival-shaped rather than seasonal. Volume concentrates violently around 618, Double 11, Chinese New Year and 520. A model that has never seen those spikes will smooth them away.
What works is feeding the model platform-side signals instead of your own history: category search volume on the platform, competitor pricing moves, livestream slots already booked by your KOLs, and add-to-cart data in the pre-sale window. Platform tools now do a lot of this natively. On the sourcing side, Alibaba’s 1688 launched a cross-border AI agent called Aoxia in 2026 that pulls demand data from over 20,000 product categories across 23 countries; 36Kr reported it compresses a product-selection analysis that took an experienced buyer at least 100 minutes down to one or two minutes.
Treat that speed honestly. It does not make the analysis smarter. It makes it cheap enough to run weekly instead of quarterly, and weekly is what a bonded warehouse needs.
Multilingual customer service that runs while you sleep
Chinese buyers expect an answer in minutes, not hours, and the pre-sale chat is part of the purchase itself. On Tmall a customer will ask three questions before adding to cart. Time zones make this brutal for a European brand. Customer service typically eats 15% to 25% of operating cost for a cross-border seller, and it scales linearly with orders unless you automate.
AI handles the repetitive 70% well: shipping timelines, customs and duty questions, ingredient and material checks, size guidance, return procedure. It handles complaints badly. Route anything with a negative sentiment score to a human immediately, because a bot arguing with an unhappy Chinese customer produces a public review you will pay for repeatedly.
Voice is the newer part. Livestream replays, short video and voice-assistant answers all need natural Mandarin audio, and Chinese text to speech tools now produce output that passes for human on a phone speaker. That matters more than it sounds: buyers scrolling Douyin at 11pm listen before they look.
The other shift is that the assistants themselves now live inside the apps buyers already use. WeChat’s AI agent reaches over a billion users without anyone downloading anything, which changes what a brand service account is expected to do. We wrote about that in what WeChat’s DeepSeek integration means for brand marketing.
Returns and fraud: the number nobody puts in the business plan
Return rates in Chinese e-commerce are far above what Western operators budget for. Free-return policies are near universal, the platforms enforce them aggressively, and a share of buyers order three sizes with the intention of keeping one. On cross-border goods a return also means the item leaves the bonded chain, which usually means you cannot resell it.
Here is where the rates actually sit in 2026, based on platform benchmark data compiled by Chinese e-commerce analytics providers:
| Segment | Typical 2026 return rate | What drives it |
|---|---|---|
| Taobao / Tmall | 10% to 25% | Shelf-style browsing, considered purchase |
| JD.com | 8% to 20% | Higher trust, more 3C and appliances |
| Pinduoduo | 10% to 30% | Price-led, low-consideration orders |
| Douyin | 15% to 40% | Impulse buying in livestream |
| Kuaishou | 12% to 35% | Impulse buying, lower-tier cities |
| Apparel and footwear | around 30% | Sizing, fit, colour on screen |
| Beauty and skincare | around 12% | Shade mismatch, ingredient concerns |
| 3C and electronics | around 8% | Spec mismatch, rare |
| Food and snacks | around 3% | Low value, consumed on arrival |
Read the table before you pick a channel. A brand selling apparel through livestream is planning for a third of its gross orders to come back. That is a business model decision, not an operations detail.
AI helps in two ways here. First, transaction-level fraud scoring on address patterns, payment behaviour and device fingerprints, which the platforms largely run for you. Second, and more useful, return-reason clustering. Push six months of Chinese-language return comments through a model, cluster them, and you usually find that 60% to 70% of returns trace back to three or four fixable listing problems. Sizing chart missing. Colour photographed under the wrong light. A material described in a way that set the wrong expectation. Fixing the listing beats chasing the refund.
AI virtual hosts in livestream
Digital human hosts have moved from novelty to standard equipment. A whitepaper published by iResearch with JD.com in April 2026 put China’s AI digital human market growth at 85.3% in 2024, with more than 1.14 million enterprises deploying the technology at scale. For a cross-border brand the appeal is obvious: an avatar covers the 2am to 8am slots that no human host will take, at a fraction of the cost, in fluent Mandarin.
The limits are equally clear. Avatars convert well on repeat purchases and commodity SKUs, poorly on discovery and premium positioning, where buyers want a face they trust. Use them to hold the slot, not to launch the brand. We go into the tooling, the costs and the platform rules in a dedicated piece on AI-powered virtual hosts in e-commerce live-streaming.
GEO: being the answer when a Chinese buyer asks an AI
This is the change most foreign brands have not priced in yet. A growing share of product research in China no longer starts on a search engine or a marketplace. It starts in a chat window.
The scale is already serious. QuestMobile reported that in Q1 2026 China’s AI-native apps reached 446 million monthly active users, with Doubao at 345 million, Qwen at 166 million and DeepSeek at 127 million. Doubao users opened the app 54.8 times per month on average, DeepSeek users 41.7 times. That is habitual daily use, not experimentation.
When a Chinese buyer types “which foreign fish oil brands are worth buying” into Doubao, the model answers with three or four brand names. If you are not one of them, you are not in the consideration set, and no amount of Tmall ad spend fixes that. Generative engine optimisation, GEO, is the work of becoming one of those names.
The mechanism is not mysterious. These models are trained and grounded on Chinese-language sources: Baidu Baike entries, Zhihu answers, Xiaohongshu notes, industry media articles, and structured product data. To be cited you need consistent Chinese-language coverage across those sources, using the exact brand name and category terms the model will match on. One Chinese name for your brand, used everywhere. Comparison content that names your category. Third-party mentions with real detail, not press releases. It is a slow build, closer to reputation work than to advertising.
Search inside social platforms works on the same logic and deserves the same attention. See our analysis of Xiaohongshu as a search engine for brands in 2026, and our guide to SEO in China for the content and backlink side.
One case: a Finnish brand that fixed the wrong thing first
Teemu runs a Finnish outdoor apparel company, around 12 million euros in revenue, strong in the Nordics. He opened a Tmall Global flagship with 180 SKUs, all listings machine-translated from English, and bought traffic from day one.
Six months in the numbers were bad. Conversion sat at 0.6% against a category benchmark near 1.5%. The return rate hit 41%, well above the apparel average, and because the goods were bonded, most returns could not be resold. He was paying for traffic that converted badly and for stock that came back dead.
His first fix was to spend more. More ad budget, then a cheaper translation vendor to cut listing cost. Conversion did not move. The problem was never traffic volume.
What we changed: we rewrote the top 40 SKUs by hand, keeping AI only for the specification tables. We added a body-measurement conversion chart built on Chinese sizing data instead of a European size grid. We reshot the colour swatches, because three jackets were photographed under Nordic daylight and arrived looking like a different product. Then we clustered eight months of Chinese return comments and found that 64% cited fit or colour, which confirmed the listings were the fault, not the garments.
Over the following five months conversion went from 0.6% to 1.7% and returns fell from 41% to 24%. Nothing about the product changed. The listings stopped lying by accident.
What AI still does not solve
Some of the hard parts of this market are not information problems, so no model will fix them.
- Regulation. Product registration, ingredient restrictions, labelling and claim rules are legal work. AI can draft, it cannot take liability.
- Competition. Domestic brands respond to a price move in hours. Your approval chain takes a week. That gap is organisational.
- Trust. Chinese buyers verify through people: KOC reviews, Xiaohongshu notes, friends in a WeChat group. That has to be earned, not generated.
- Platform choice. Tmall Global, JD Worldwide, Douyin and Xiaohongshu behave differently and reward different products. Our guide to Chinese e-commerce platforms lays out which fits what.
FAQ
Can I run a Chinese cross-border store on AI translation alone?
No, and the failure is invisible until you look at conversion. Machine translation produces grammatically correct listings that use the wrong category terms, keep European sizing, and sometimes carry claims that are regulated in China. The workable split is AI for specification tables, ingredient lists and parameter blocks, human Chinese copywriters for the first screen, headlines, claims and review responses. That keeps roughly half the cost saving without paying for it in lost sales.
What return rate should I budget for in China?
It depends on channel and category, not on your home-market experience. Shelf platforms like Tmall and JD sit between 8% and 25%. Livestream channels like Douyin and Kuaishou run 15% to 40% because buying is impulsive. Apparel averages around 30%, beauty around 12%, food around 3%. On bonded cross-border stock, most returns cannot be resold, so model returns as lost margin rather than a logistics line.
What is GEO and does it matter for a cross-border brand?
GEO is generative engine optimisation: making sure Chinese AI assistants name your brand when a buyer asks for a recommendation. With Doubao at 345 million monthly users and DeepSeek at 127 million in early 2026, a meaningful share of product research now happens in a chat window. Models cite Chinese-language sources, so the work is consistent coverage on Baidu Baike, Zhihu, Xiaohongshu and industry media, under one fixed Chinese brand name.
Are AI virtual livestream hosts worth using for a foreign brand?
For overnight slots and repeat-purchase SKUs, yes. They cost a fraction of a human host, run 24 hours, and speak fluent Mandarin. For launching a brand or selling premium positioning, no. Chinese buyers discovering an unfamiliar foreign brand want a human face they can judge. Use avatars to keep the room open and humans for the sessions that actually build the brand.
Working with GMA on cross-border
GMA has been running cross-border e-commerce for foreign brands from Shanghai since 2012, on Tmall Global, JD Worldwide, Douyin and Xiaohongshu. We write Chinese listings that convert, manage bonded stock forecasting, and build the Chinese-language footprint that gets brands cited by Baidu and by AI assistants like DeepSeek and Doubao. If your store has traffic but not sales, that is usually a listing and visibility problem, and it is fixable.
Contact our team for a review of your cross-border setup.
