一笔订单的完整旅程,与利润的真相 · AI 与电商产业 · 2026-07One order's complete journey, and the truth about profit · AI & e-commerce · Jul 2026

AI 没有替代电商,它做了三件更狠的事:重分利润、搬走入口、重排人力——美工客服投放这些表层劳动被整层吃掉,而履约、账期、退货这五块硬骨头,一块都没啃动 AI didn't replace e-commerce — it did three harsher things: re-split the profit, moved the gateway, re-stacked the labour. The surface work of design, service and ads was eaten whole, while the five hard bones of fulfilment, credit terms and returns have not lost a single bite

诚实层=平台税的数学。商家净利 = GMV × 毛利率 − 佣金 − 投流 − 履约 − 退货逆向。把数字填进去:单品综合抽成 35–45%、抖音自播间投流常态占 GMV 50%+、双十一综合退货率 61.5%——等式解出来,绝大多数卖家的身份是给算法与物流打工的「数字佃农」。「开店致富」不是被 AI 杀死的,是被这道算术杀死的。The honesty layer = the math of the platform tax. Net = GMV × margin − commission − ad spend − fulfilment − returns. Fill in the numbers: a per-item blended take of 35–45%, Douyin self-broadcast ad spend routinely 50%+ of GMV, a Singles' Day blended return rate of 61.5% — solve the equation, and most sellers turn out to be «digital sharecroppers» working for the algorithm and the logistics network. «Get rich by opening a store» wasn't killed by AI; it was killed by this arithmetic.

范式层=流量入口的第三次迁移。搜索货架(人找货)→ 内容直播(货找人)→ AI 导购 / Agent 代客决策。当买方 Agent 与卖方 Agent 在后台互相比价,信息差利润不可逆消亡,价格回归绝对供应链性价比;写给人性弱点看的营销话术,在一台不冲动的机器面前全部失效。但别急着宣布完成——OpenAI 2026-03 收缩原生结账:这场迁移是「进行时」。The paradigm layer = the third migration of the traffic gateway. Search shelf (people find goods) → content & live (goods find people) → AI shopping agents deciding on your behalf. Once buyer agents and seller agents haggle in the backend, information-asymmetry profit dies irreversibly and price returns to raw supply-chain value; copy written for human weakness fails entirely before a machine that never buys on impulse. But hold the announcement — OpenAI pulled back native checkout in March 2026: this migration is «in progress».

主脊是一笔订单的八节点旅程(选品→采购→内容上架→流量→转化客服→履约→退货逆向→复购),每节点标「已完成时」(AI 已吃掉)vs 「进行时」(尚未跑通)。另含平台税瀑布、五块硬骨头、三次迁移、龙头分栏、产品指南(2C 消费者与小卖家 / 2B 商家与平台侧)。这是一张批判性行业解剖,不是开店教程。姊妹图:B2B 销售→sell、达人变现→creator、虚拟模特→beauty、take rate 模型→money。 The spine: one order's eight-node journey (selection → sourcing → listing → traffic → conversion/service → fulfilment → returns → repurchase), each node tagged «done» (AI ate it) vs «in progress» (not yet working). Plus the platform-tax waterfall, five hard bones, the three migrations, leader columns, and a product guide (2C consumers & small sellers / 2B merchants & platforms). A critical dissection, not a how-to-open-a-store guide. Siblings: B2B sales → sell, creator monetisation → creator, virtual models → beauty, take-rate models → money.

传统节点Traditional

AI 已吃掉 · 表层劳动(已完成时)AI ate it · surface labor (done)

AI 导购 · Agent 第三次迁移(进行时)AI shopping · agent migration (in progress)

平台税 · 退货黑洞 · 危机Platform tax · returns · crisis

硬骨头 · 供应链上游 · 真正杠杆Hard bones · upstream · real leverage

61.5%

2025 双十一综合退货率;服装直播超 80%、女装大促极端 90%。一件 100 元女装退一单亏 50–80 元——前端卖 3 单,才抵 1 单退货2025 Singles' Day blended return rate; live-stream apparel tops 80%, womenswear promos hit 90%. One returned ¥100 dress loses ¥50–80 — it takes 3 sales to offset 1 return

35–45%

单品平台综合抽成(佣金+FBA+广告,$25 商品,Marketplace Pulse);亚马逊综合成本 2016 约 1/3 → 2022 超 50%。利润的大头不归卖货的人Per-item blended platform take (referral+FBA+ads, a $25 item, Marketplace Pulse); Amazon's blended cost went from ~1/3 in 2016 to 50%+ in 2022. The lion's share does not go to the seller

+693%

2025 假日季生成式 AI 导流零售网站同比涨幅(Adobe,超 1 万亿次访问),AI 来源访客转化率高 31%——第三次迁移的实测流速YoY surge in generative-AI referrals to retail sites in the 2025 holidays (Adobe, 1T+ visits), with AI-sourced visitors converting 31% higher — the third migration's measured flow rate

45.22%

2025 上半年商家投诉中「任意仅退款」占比第一(网经社);2025 年 3–4 月主流平台密集取消强制仅退款——平台规则的政治性,一夜能改写你的生意模型«Refund-only at will» topped merchant complaints in H1 2025 (100EC); in Mar–Apr 2025 the major platforms scrapped forced refund-only — the politics of platform rules can rewrite your business model overnight

口径警告:本页是批判性行业分析,不是开店教程 / 投资建议;厂商自述与独立数据并置。货币化率 / 平台抽成多口径冲突(拼多多 4.55% vs 7.6% 因分母 GMV 不同;亚马逊 22–29% 占平台收入 vs 单品 35–45% vs 占卖家收入 1/3→50%,统计对象不同,已逐条标注)。退货率 61.5%(双十一综合)/ 女装 80–90% / 美国线上 19–20%,类目地区口径不同。平台自报 AI 效果(CTR +25% / Listing +40% / 90% 几乎不改 / Rufus 120 亿增量)多为厂商口径、未经独立审计;市场规模(McKinsey 3–5 万亿)为 2030 预测。agentic commerce 归「进行时」(OpenAI 2026-03 收缩结账)。卡片右上角 A/B/C/D=证据强度。本图纪律:能核实的说规模,不能核实的不说效率。这一段里流传最广的几个数字——某先买后付公司称 AI 客服省 4000 万美元、替代 700 名客服、咨询机构给的收入提升与生产力百分比、某物流大脑的效率指标——目前均无公开的第三方量化验证,全部是厂商自报或咨询估算;可核实的一侧则是形态与规模事实:生成式购物助手 Rufus 2024 年从 Beta 全量上线至全美用户,新一代履约中心的机器人数量达标准中心 10 倍、新库存系统可容纳超 3000 万件商品。 Basis warning: a critical industry analysis, not a store guide / investment advice; vendor claims sit beside independent data. Take rate / platform cut carries conflicting bases (PDD 4.55% vs 7.6% differ by GMV denominator; Amazon 22–29% of platform revenue vs 35–45% per item vs 1/3→50% of seller revenue — different objects, flagged one by one). Return rates 61.5% (Singles' Day blended) / 80–90% womenswear / 19–20% US online differ by category and region. Platform-reported AI effects (CTR +25% / listing +40% / 90% barely edited / Rufus $12B incremental) are largely vendor figures, unaudited; market sizes (McKinsey $3–5T) are a 2030 forecast. Agentic commerce is filed as «in progress» (OpenAI pulled back checkout, Mar 2026). Card badges A/B/C/D = evidence strength. The discipline here: state scale where it can be checked, stay silent on efficiency where it cannot. The most-quoted numbers in this section — one buy-now-pay-later firm's claim of $40M saved and 700 support agents replaced, a consultancy's revenue-lift and productivity percentages, a logistics «super brain's» efficiency metrics — have no public third-party quantification; all are vendor-reported or consultancy estimates. The checkable side is form and scale: the generative shopping assistant Rufus went from beta to all US users in 2024, and the newest fulfilment centre runs 10× the robots of a standard site with an inventory system holding over 30 million items.

◆ 诚实层 · 平台税的数学The honesty layer · the math of the platform tax

100 元 GMV,流向了哪里?Where does ¥100 of GMV actually go?

看电商的任何「AI 革命」,先做一道减法题。下面是一件商品售价的示意拆解——货品成本、平台佣金、营销投流、履约、退货逆向层层剥完,商家净利常只剩一条细缝。把这条细缝放在眼前再看 AI:它并没有把蛋糕做大多少,它改变的是哪一项成本被压缩、哪一项被抬高、哪一项从显性变成隐性。谁的成本项被 AI 压掉,谁就短暂受益;谁的收入项恰好是别人被压掉的成本,谁就出局。Before believing any «AI revolution» in e-commerce, do one subtraction. Below is an illustrative breakdown of a single item's price — peel off goods cost, commission, ad spend, fulfilment and returns, and the seller's net is usually a sliver. Hold that sliver in view and AI looks different: it barely grows the pie; what it changes is which cost gets compressed, which gets raised, and which turns from visible to hidden. Whoever's cost line AI compresses briefly wins; whoever's revenue line was someone else's compressed cost is out.

▚ 一件商品售价的去向(示意 · 非精确,锚定行业区间)Where one item's price goes (illustrative, anchored to industry ranges)

综合抽成 35–45% + 退货损耗 + 履约 + 货本,挤出「数字佃农」的真实薄利A 35–45% blended take + returns + fulfilment + goods cost squeeze out the «digital sharecropper's» true thin margin

38%货品成本goods

16%平台佣金commission

20%营销投流ad spend

11%履约fulfillment

11%退货逆向returns

货品成本(供应链)Goods (supply chain)

平台佣金 / 抽成Platform commission

营销投流(抖音自播 50%+)Ad spend (Douyin 50%+)

履约仓配Fulfillment

退货逆向损耗Returns loss

商家净利(细缝)Seller net (the sliver)

货币化率的两个时代:起步期国内平台以低税率换生态(阿里 3.66% / 京东 5.5% / 拼多多 2.17%,对照亚马逊 16% / eBay 10.82%);如今低税率时代终结——拼多多货币化率约 7.6%(2024)几乎是淘天 3.9% 的两倍。更隐蔽的一层是巨头的防御性营销军备:阿里 2025 财年营销费用激增 77.67% 至 1916 亿元、京东增 75.07% 至 839.5 亿、拼多多 1253 亿——存量互搏烧掉的每一分钱,都通过更贵的竞价广告系统性转嫁为商家的获客成本。结果就是:年销破亿的女装商家,扣掉近千万退货费用,净利只剩两三百万。Two eras of take rate: China's platforms once traded low taxes for ecosystem (Alibaba 3.66% / JD 5.5% / PDD 2.17%, vs Amazon 16% / eBay 10.82%); that era is over — PDD's take is ~7.6% (2024), nearly double Taotian's 3.9%. The stealthier layer is the giants' defensive marketing arms race: Alibaba's FY2025 marketing spend jumped 77.67% to ¥191.6B, JD +75.07% to ¥83.95B, PDD ¥125.3B — every yuan burned fighting over existing users is systematically passed on as sellers' acquisition cost through pricier auction ads. Net result: a womenswear seller past ¥100M in sales, minus ~¥10M in return costs, keeps ¥2–3M.

▸ 退货黑洞 · 诚实指标The returns black hole · the honest metric

卖得越多,亏得越快The more you sell, the faster you bleed

内容电商的加速器装反了方向:直播+推荐流把「看到→下单」压缩到秒级,把「后悔→退货」也压缩到秒级。女装日常退货 50–60%、大促叠加主播激情与夸张滤镜飙到 80–90%;运费险从可选增值沦为事实标配,试错成本被抹平之后,试错就成了消费方式本身。每一单退货背后是双向运费+运费险+分拣清洗+残次折旧的隐性账——财报上 GMV 光鲜,现金流在仓库里枯竭。这不是中国特色:美国线上服饰退货率也达 22–26%(线下仅 6.2%),是渠道属性,不是国民性。Content commerce wired its accelerator backwards: live + feeds compress «see → buy» to seconds, and compress «regret → return» to seconds too. Womenswear returns run 50–60% daily and hit 80–90% in promos under host hype and heavy filters; return-shipping insurance slid from optional add-on to de facto standard — once the cost of trying is erased, trying becomes the consumption style itself. Behind every return sits a hidden bill of two-way shipping + insurance + sorting + cleaning + write-downs — GMV shines in the filings while cash dies in the warehouse. Not a China quirk: US online apparel returns hit 22–26% (vs 6.2% in-store) — it's the channel, not the culture.

▸ 仅退款 · 平台规则的政治性Refund-only · the politics of platform rules

从「合法劫掠」到集体纠偏From «legal looting» to a collective reversal

一条平台规则的生命史:拼多多 2021 年在生鲜品类首创「仅退款」,本意惩罚劣质商家;随后向全品类蔓延、被淘宝京东抖音跟进,异化为羊毛党与黑产「合法劫掠」商家的工具。2025 上半年商家投诉里「任意仅退款」占 45.22%(第一)、「任意罚款」11.94%、「扣押保证金」8.64%。直到 2025 年 3–4 月,监管与商家承压双重作用下,主流平台才密集取消强制仅退款。这段历史的教训比规则本身重要:在平台上做生意,你的商业模型随时可能被一次规则调整改写——这就是本页把「平台规则的政治性」列为硬骨头的原因。The life cycle of one platform rule: PDD invented «refund-only» for fresh food in 2021 to punish bad sellers; it spread to every category, was copied by Taobao/JD/Douyin, and mutated into a tool for scalper rings to «legally loot» merchants. In H1 2025, «refund-only at will» topped merchant complaints at 45.22%, «arbitrary fines» 11.94%, «seized deposits» 8.64%. Only in Mar–Apr 2025, under regulatory and merchant pressure, did the major platforms scrap forced refund-only. The lesson outweighs the rule: on a platform, your business model can be rewritten by one policy tweak — which is exactly why this page files «the politics of platform rules» among the hard bones.

◆ 判断层杠杆排序 · 结论卡The leverage stack · the conclusion

选品与供应链(上游)Selection & supply chain ＞运营技巧operations ＞视觉优化(已被 AI 商品化)visuals (AI-commoditized)

退货率与复购(诚实指标)Returns & repurchase ＞ GMV(可投流买出的虚荣指标)GMV (a vanity metric you can buy)

核心洞察:AI 吃掉的环节,恰恰是杠杆最低的环节。美工、话术、投放被 AI 商品化后,所有人都有,等于所有人都没有;真正的护城河退回到 AI 尚难渗透的选品与供应链上游。再逼真的 AI 模特、数字人、GEO 也救不了质量低劣、成本倒挂的商品——淘宝家居 AI 图能把点击率 +25%,但线上服饰 24.4% 的退货率足以吞掉最漂亮详情页的全部利润。判断一个电商 AI 工具的价值,先问它作用在哪一层杠杆上。Core insight: what AI ate is precisely the lowest-leverage layer. Once design, copy and ad ops are AI-commoditized, everyone has them — so no one does; the moat retreats upstream to selection and supply chain, where AI barely reaches. No photoreal AI model, digital human or GEO trick saves a shoddy, cost-inverted product — Taobao's AI home images lift CTR +25%, yet a 24.4% online-apparel return rate swallows the profit of the prettiest page. To price any e-commerce AI tool, first ask which leverage layer it touches.

从这张图看到的五条规律Five patterns this map makes visible