智能体电商爆发：AI 替你下单，跨境卖家如何被机器选中（2026）
The Agentic Commerce Boom: How Sellers Get Picked by AI Shopping Agents (2026)
> 📌 TL;DR
> 2025 到 2026 年，"让 AI 替你买东西"从演示变成了现实：ChatGPT 在 2025 年 9 月上线站内结账，Google 在 2026 年 5 月的 I/O 上发布"万能购物车"（Universal Cart），亚马逊、Perplexity 各有自己的购物 Agent。当下单的不再是人、而是 AI 智能体时，卖家面对的是一套全新的游戏规则——从"取悦人眼"变成"被机器读懂"。这篇帮你把这场变革讲清楚，并给出 6 件现在就该做的事。
先看一个正在发生的场景
2026 年的某个深夜，一位美国消费者对 ChatGPT 说："帮我找一双 40 码、白色、200 美元以内、评分 4.5 以上的跑鞋，今晚下单。"几秒钟后，AI 比对了十几家店铺的价格、库存、评价和退货政策，直接在对话框里完成了支付——这位消费者没有打开任何一个购物网站，没有看任何一张产品主图。
这不是科幻。这就是智能体电商（Agentic Commerce）：AI 智能体代替人，自主完成"搜索 → 比价 → 决策 → 下单 → 支付"的全过程。对卖家来说，这意味着一个残酷的事实——你的产品页可能再也不会被人眼看到，决定成交的，是一个读结构化数据的机器。
这两年到底发生了什么：一条时间线
过去 18 个月，几乎所有巨头都把"购物 Agent"推上了牌桌：
| 时间 | 事件 | 谁 |
|------|------|-----|
| 2024 年 11 月 | Perplexity 推出 "Buy with Pro"，PayPal 通行密钥一键下单 | Perplexity |
| 2025 年 1 月 | OpenAI 发布网页操作 Agent "Operator" | OpenAI |
| 2025 年 4 月 | Visa Intelligent Commerce、Mastercard Agent Pay 同月发布 | Visa / Mastercard |
| 2025 年 9 月 | ChatGPT "Instant Checkout" 上线，Etsy 首发（当日股价 +16%） | OpenAI + Stripe |
| 2025 年 9 月 | Google 公布 AP2 智能体支付协议，60+ 合作伙伴 | Google |
| 2026 年 5 月 | Google I/O 发布 "Universal Cart"，跨 Amazon/Shopify/Walmart 购物 | Google |
（来源：OpenAI、Google Cloud、Stripe、CNBC 官方公告，2024 年 11 月–2026 年 5 月）
值得单独说说亚马逊：购物助手 Rufus 在 2025 年服务了超过 3 亿用户，带来约 120 亿美元的增量年化销售额（亚马逊官方数据，2025 年）。但就在 2026 年 5 月 13 日，亚马逊宣布"砍掉 Rufus 聊天机器人"，把 Alexa for Shopping 扶上购物 AI 战略的 C 位（CNBC，2026-05-13）。可见连巨头的打法都还在剧烈变动。
数据说话：这不是小众实验
- 流量端：Adobe Analytics 报告，来自生成式 AI 的美国零售网站流量在 2025 年 3 月已同比暴涨 1,200%；2025 年假日季再 +693%（11 月 +769%，12 月 +673%）（Adobe，2025-03 / 2026-01）。
- 转化端："AI 来的流量不下单"这个偏见正在被打破——Adobe 数据显示，2025 年假日季 AI 引流的转化率比其他渠道高出 31%，几乎翻倍；进入 2026 年 Q1，AI 引流仍同比 +393%（TechCrunch 援引 Adobe，2026-04）。
- 大盘预测：麦肯锡估算，到 2030 年 AI 驱动的商业可在美国创造约 1 万亿美元、全球最高 5 万亿美元 的收入；Gartner 则预测到 2028 年 90% 的 B2B 采购 将由 AI 智能体居间完成（Gartner，2026-01；McKinsey，2026）。
三类购物 Agent，别用同一种打法
并不是所有购物 Agent 都一样。粗略可以分成三类，卖家的应对策略完全不同：
- "搜索优先"型（ChatGPT、Perplexity、Gemini）：用户带着模糊需求而来，Agent 跨品牌比价、综合评价。这里拼的是被发现——你的结构化数据和口碑，决定了你进不进 Agent 的候选名单。
- "购买优先"型（亚马逊 Rufus / Alexa for Shopping）：用户打开 App 时购买意图已经很强，Agent 帮你在站内快速成交。这里拼的是站内排名与转化，逻辑更像传统平台运营的延伸。
- "跨平台代理"型（Google Universal Cart）：一个界面跨 Amazon、Shopify、Walmart 下单。这里拼的是你在多渠道的数据一致性——同一个 SKU 在不同平台价格、库存、描述打架，Agent 只会挑那个最"干净"的。
一句话：搜索优先看口碑，购买优先看排名，跨平台看一致性。
看不见的管道：一场"协议战争"
AI 要替你付款，背后需要一套"我确实得到了用户授权"的信任机制。现在赛道上至少有三套互相竞争的标准：
- ACP（Agentic Commerce Protocol）：OpenAI 与 Stripe 联合开发，开源，驱动 ChatGPT 的站内结账（2025 年 9 月）。
- AP2（Agent Payments Protocol）：Google 主导，2025 年 9 月发布，用三张签名"授权书"（意图 / 购物车 / 支付 Mandate，基于 W3C 可验证凭证）锁定"所见即所付"；现已捐给 FIDO 联盟，2026 年 4 月的 v0.2.0 还加入了"用户不在场"自动下单。
- Visa / Mastercard 的代币化方案：把信用卡凭证安全地交给 Agent 使用。
> ⚠️ 卖家要注意
> 协议没统一，意味着短期内你可能要同时面对多套接入标准。但有个捷径：如果你在 Shopify、Etsy 这类已经接入主流协议的平台上开店，很多"被 Agent 买到"的能力是平台层面自动获得的——这正在悄悄成为独立站（DTC）的生死分水岭。
为什么这是卖家的"地震"
传统电商，你优化的是人：好看的主图、有说服力的文案、情绪化的促销。智能体电商，你优化的是机器：
| 维度 | 给人看的电商 | 给 AI 看的电商 |
|------|------------|--------------|
| 入口 | 搜索框 + 广告位 | AI 的推荐与比价 |
| 决策依据 | 主图、品牌、情绪 | 结构化数据、评分、退货政策 |
| 比价 | 人懒得比 3 家以上 | 机器毫秒级比 50 家 |
| 致命伤 | 页面加载慢、图丑 | 数据脏、价格/库存不同步 |
一个典型的新冲突：2025 年 11 月 4 日，亚马逊起诉 Perplexity，指其 Comet 浏览器 Agent 未经授权访问平台，法院随后发出初步禁令（多家媒体，2025-11）。平台愿不愿意让 Agent 进来、以什么条件进来，本身就是新战场。
对中国跨境卖家意味着什么
对中国卖家，这既是机会也是风险：
- 机会：AI 不看你的广告预算，只看数据和真实表现。一个把产品信息、评价、退货政策做扎实的中小品牌，第一次有机会和大牌"同台被 AI 比较"，靠产品力而非烧钱被选中。
- 风险：Agent 毫秒级比价，会把"同质化白牌"逼进纯价格战——如果你唯一的卖点是便宜，AI 会替消费者找到更便宜的那一家。再叠加 2025 年美国取消 de minimis 免税、欧盟把 150 欧门槛从 2028 提前到 2026 年取消所带来的成本上升，"只靠低价"的路会越走越窄。
应对的核心，是从"流量思维"转向"被信任思维"：让机器更容易读懂你，也更有理由相信你。
现在就该做的 6 件事
1. 把结构化数据当一等公民：补全 Schema.org 产品标记（价格、库存、评分、GTIN、退货政策）。AI 读的是这些字段，不是你的主图。
2. 喂干净的产品 Feed：标题、属性、规格要机器可解析、无歧义。一个错的尺码或材质，会让 Agent 直接把你过滤掉。
3. 价格与库存实时同步：Agent 毫秒级比价，价格/库存不准 = 当场出局，甚至触发错单纠纷。
4. 把评价做成机器可读的信任信号：真实评分、问答、退货率，是 AI 判断"该不该推荐你"的核心依据。
5. 接入 agentic checkout 能力：评估你的平台是否已支持 ACP/AP2；独立站要主动规划接入路线。
6. 开始监控"AI 引流"：在 GA4 等分析工具里单独追踪来自 ChatGPT、Perplexity、Gemini 的流量与转化——你无法优化一个你看不见的渠道。
> ✨ 一句话记住
> 过去十年，电商的胜负手是"如何排在人的搜索结果第一位"；未来十年，是"如何被 AI 选进那个唯一的购物车"。看不见的客户，正在替看得见的客户做决定。
本文数据截至 2026 年 5 月，关键来源已在文中标注发布时间。
> 📌 TL;DR
> Between 2025 and 2026, "let AI buy it for you" stopped being a demo and became real: ChatGPT launched in-chat checkout in September 2025, Google unveiled the "Universal Cart" at I/O in May 2026, and Amazon and Perplexity each ship their own shopping agents. When the buyer is no longer a human but an AI agent, sellers face a brand-new rulebook — the game shifts from pleasing human eyes to being readable by machines. Here's what's happening, and 6 things to do about it right now.
A scene that's already happening
Late one night in 2026, a US shopper tells ChatGPT: "Find me white running shoes, size 9, under $200, rated above 4.5 — and order them tonight." Seconds later the AI has compared prices, stock, reviews, and return policies across a dozen stores, and completed the payment right inside the chat. The shopper never opened a single store, never looked at a single product hero image.
This isn't sci-fi. It's agentic commerce: AI agents autonomously running the full "search → compare → decide → order → pay" loop on a person's behalf. For sellers, it implies something brutal: your product page may never be seen by human eyes again. The thing that closes the sale is a machine reading structured data.
What actually happened in 18 months: a timeline
Over the past year and a half, nearly every giant put a "shopping agent" on the table:
| Date | Event | Who |
|------|-------|-----|
| Nov 2024 | Perplexity launches "Buy with Pro," one-click via PayPal passkey | Perplexity |
| Jan 2025 | OpenAI ships web-operating agent "Operator" | OpenAI |
| Apr 2025 | Visa Intelligent Commerce and Mastercard Agent Pay launch the same month | Visa / Mastercard |
| Sep 2025 | ChatGPT "Instant Checkout" goes live; Etsy first (stock +16% that day) | OpenAI + Stripe |
| Sep 2025 | Google announces the AP2 agent-payments protocol, 60+ partners | Google |
| May 2026 | Google I/O unveils "Universal Cart," shopping across Amazon/Shopify/Walmart | Google |
(Sources: OpenAI, Google Cloud, Stripe, CNBC official announcements, Nov 2024–May 2026)
Amazon deserves its own note: its shopping assistant Rufus served 300M+ users in 2025 and drove roughly $12B in incremental annualized sales (Amazon, 2025). Yet on May 13, 2026, Amazon announced it was retiring the Rufus chatbot and making Alexa for Shopping the centerpiece of its AI shopping strategy (CNBC, 2026-05-13). Even the giants are still rewriting their own playbooks.
The data: this is not a niche experiment
- Traffic: Adobe Analytics reported that US retail traffic from generative-AI sources was already up 1,200% year-over-year by March 2025, and rose another 693% over the 2025 holiday season (Nov +769%, Dec +673%) (Adobe, 2025-03 / 2026-01).
- Conversion: The old "AI traffic doesn't convert" myth is breaking — Adobe found AI-referred traffic converted 31% better than other channels during the 2025 holidays, nearly doubling year-over-year. In Q1 2026, AI referral traffic was still up 393% YoY (TechCrunch citing Adobe, 2026-04).
- Forecasts: McKinsey estimates AI-driven commerce could generate ~$1 trillion in the US and up to $5 trillion globally by 2030. Gartner predicts that by 2028, 90% of B2B buying will be intermediated by AI agents (Gartner, 2026-01; McKinsey, 2026).
Three kinds of shopping agents — don't treat them alike
Not all shopping agents are the same. Roughly three types exist, and each demands a different playbook:
- Search-first (ChatGPT, Perplexity, Gemini): users arrive with fuzzy intent; the agent compares across brands and synthesizes reviews. The game here is discovery — your structured data and reputation decide whether you even make the agent's shortlist.
- Purchase-first (Amazon Rufus / Alexa for Shopping): intent is already strong when the user opens the app; the agent closes the sale on-platform. The game is in-platform ranking and conversion — closer to an extension of traditional marketplace ops.
- Cross-platform proxies (Google Universal Cart): one interface ordering across Amazon, Shopify, and Walmart. The game is data consistency across channels — if the same SKU has clashing price, stock, or descriptions across platforms, the agent simply picks the "cleanest" one.
In short: search-first rewards reputation, purchase-first rewards ranking, cross-platform rewards consistency.
The invisible plumbing: a "protocol war"
For an AI to pay on your behalf, there has to be a trust layer that proves "the user really authorized this." At least three competing standards are now on the field:
- ACP (Agentic Commerce Protocol): co-developed by OpenAI and Stripe, open-source, powers ChatGPT's in-chat checkout (Sep 2025).
- AP2 (Agent Payments Protocol): led by Google, launched Sep 2025, using three signed "Mandates" (Intent / Cart / Payment, as W3C Verifiable Credentials) to lock in "what you see is what you pay." It's been donated to the FIDO Alliance, and v0.2.0 (Apr 2026) added "Human Not Present" autonomous purchases.
- Visa / Mastercard tokenization: securely hands card credentials to agents.
> ⚠️ Sellers, take note
> No unified standard means you may have to support several integration paths in the near term. But there's a shortcut: if you sell on platforms like Shopify or Etsy that already adopted the major protocols, much of your "buyable-by-agents" capability comes automatically at the platform level — which is quietly becoming the make-or-break line for standalone DTC stores.
Why this is an earthquake for sellers
In traditional ecommerce, you optimize for humans: beautiful hero shots, persuasive copy, emotional promos. In agentic commerce, you optimize for machines:
| Dimension | Commerce for humans | Commerce for AI |
|-----------|--------------------|-----------------|
| Entry point | Search box + ad slots | The AI's recommendation & price comparison |
| Decision basis | Imagery, brand, emotion | Structured data, ratings, return policy |
| Comparison | Humans rarely compare >3 stores | Machines compare 50 in milliseconds |
| Fatal flaw | Slow page, ugly photos | Dirty data, price/stock out of sync |
A telling new conflict: on Nov 4, 2025, Amazon sued Perplexity, alleging its Comet browser agent accessed the platform without authorization; a court later issued a preliminary injunction (multiple outlets, 2025-11). Whether — and on what terms — a platform lets agents in is itself a new battleground.
What this means for cross-border sellers
For cross-border sellers, this is both an opportunity and a risk:
- The opportunity: AI doesn't look at your ad budget — only your data and real performance. A small brand that nails product info, reviews, and return policy gets, for the first time, a shot at being compared head-to-head with the big names by an AI — and picked on merit, not ad spend.
- The risk: millisecond price comparison turns commodity "white-label" products into a pure price war — if your only edge is being cheap, the AI will find the shopper someone cheaper. Layer on rising costs from the US ending de minimis (2025) and the EU pulling its €150 threshold forward from 2028 to 2026, and "low price only" becomes an ever-narrower road.
The core response is a shift from a "traffic mindset" to a "trust mindset": make it easier for the machine to understand you, and give it more reason to believe you.
6 things to do right now
1. Treat structured data as a first-class citizen. Complete your Schema.org product markup (price, stock, rating, GTIN, return policy). The AI reads these fields, not your hero image.
2. Feed clean product data. Titles, attributes, and specs must be machine-parseable and unambiguous. One wrong size or material and the agent filters you out.
3. Sync price and stock in real time. Agents compare in milliseconds; stale price/stock = instant elimination, or worse, mis-order disputes.
4. Turn reviews into machine-readable trust signals. Genuine ratings, Q&A, and return rates are how an AI decides whether to recommend you at all.
5. Wire up agentic checkout. Check whether your platform already supports ACP/AP2; if you run a standalone store, plan your integration path proactively.
6. Start tracking "AI referral" traffic. Separately monitor traffic and conversion from ChatGPT, Perplexity, and Gemini in GA4 and friends — you can't optimize a channel you can't see.
> ✨ One line to remember
> For the last decade, winning ecommerce meant ranking first in a human's search results. For the next decade, it means being the one product an AI drops into the only cart that matters. The invisible customer is now deciding for the visible one.
Data current as of May 2026; key sources are dated inline.