经济学人丨为什么你的AI助手突然开始“卖货”?

本篇刊次：2026年4月刊（4/4）

英文标题：Why your AI assistant is suddenly selling to you

文章摘要：为解决巨额的算力成本和变现难题，OpenAI与谷歌等科技巨头正将聊天机器人转变为推销员，在对话中植入广告；这种新型数字广告模式虽未引起用户反感，但在精准定位、效果测量和品牌安全上仍面临挑战。

阅读时长：6分钟

CHATBOTS ARE employed every day as teachers, counsellors, coders and escorts. Now they are taking on another role: salesmen. Advertisements are popping up ever more frequently in users’ conversations with large language models, punctuating chats with promotions. Consumers’ search queries, editing sessions and even intimate moments are increasingly at risk of interruption by sponsored messages.

聊天机器人每天都被用作教师、顾问、程序员和伴侣。现在他们正在承担另一个角色：推销员。广告在用户与大型语言模型的对话中越来越频繁地弹出，用促销信息打断聊天。消费者的搜索查询、编辑会话甚至私密时刻，都越来越面临被赞助信息打断的风险。

As chatbots become adbots, the future of two industries is at stake. For the artificial-intelligence business, ads represent a way to monetise a wildly expensive invention that most people currently use free of charge. For the ad industry, adbots are a possible answer to the existential question of how advertising will work if users move away from conventional search engines. Although it is early days, the outlines of a new kind of marketing are emerging.

随着聊天机器人变成广告机器人，两个行业的未来面临考验。对于人工智能企业来说，广告代表了一种将极其昂贵、且目前大多数人免费使用的发明变现的方法。对于广告业来说，如果用户不再使用传统的搜索引擎，广告将如何运作？广告机器人可能是这个关乎生存问题的一个答案。尽管还处于早期阶段，但一种新型营销模式的轮廓正在显现。

Many of the AI giants are piling in. In February OpenAI said it would begin testing ads in ChatGPT, the most widely used chatbot. Google has been trialling them in its search engine’s “AI Mode” since last year, and Microsoft has woven them into its Copilot. Amazon’s shopping assistant, Rufus, allows brands to sponsor its replies. Meta’s chatbot does not yet show ads, but since December it has been passing on insights from chats to its sister companies, Facebook and Instagram, to help them serve better-targeted plugs.

许多AI巨头正在涌入。今年2月，OpenAI表示将开始在最广泛使用的聊天机器人ChatGPT中测试广告。谷歌从去年开始就在其搜索引擎的“AI模式”中试用广告，微软也已将它们编织进Copilot中。亚马逊的购物助手Rufus允许品牌赞助其回复。Meta的聊天机器人目前还不显示广告，但自去年12月以来，它一直在将从聊天中获得的洞察信息传递给其姊妹公司Facebook和Instagram，以帮助它们提供针对性更强的推广。

For OpenAI, which reportedly expects to burn $25bn of cash this year and twice as much in 2027, the need for more revenue is acute. Although it sells subscriptions, most of ChatGPT’s more than 900m weekly users are on its free tier. Ads are a way to make money from them. And they may let the company make its best models available to users who do not pay, deepening their engagement, says Benedict Evans, a tech analyst. Even Google, with no immediate cashflow worries, has reason to bring ads to AI chats if these interactions are to replace search for some users. Already it is set to lose its crown as the biggest seller of digital ads: this year Meta will overtake it with ad revenue of $243bn, according to eMarketer, a research firm.

对于据报道预计今年将烧掉250亿美元现金、到2027年将烧掉两倍现金的OpenAI来说，对更多收入的需求非常迫切。尽管它出售订阅服务，但ChatGPT每周超过9亿的用户中，大多数使用的是免费服务。广告是从他们身上赚钱的一种方式。科技分析师本尼迪克特·埃文斯（Benedict Evans）表示，广告还可以让公司将最好的模型提供给不付费的用户，从而加深他们的参与度。即使是目前没有现金流担忧的谷歌，如果这些交互将取代部分用户的搜索，它也有理由将广告引入AI聊天中。它已经注定要失去最大的数字广告销售商的桂冠：根据研究公司 eMarketer 的数据，今年Meta将以2430亿美元的广告收入超过它。

So far the experiments are modest. Similarweb, a data provider, estimates that ChatGPT is showing ads in only about 1% of conversations, and Google in an even smaller share of AI Mode queries. But they are due to rise fast. OpenAI is said to have told investors that it expects ChatGPT to generate $2.5bn in ad revenue this year and $11bn in 2027, with a goal of hitting $100bn by 2030.

到目前为止，这些实验还很有限。数据提供商 Similarweb 估计，ChatGPT仅在约1%的对话中显示广告，而谷歌在“AI模式”查询中显示广告的比例甚至更小。但它们预计将快速上升。据说OpenAI已告知投资者，预计ChatGPT今年将产生25亿美元的广告收入，2027年将达到110亿美元，目标是到2030年达到1000亿美元。

The early evidence suggests that OpenAI and Google have different approaches to how a chatbot should sell. According to Similarweb, 98.5% of the ads in Google’s AI Mode appear in response to the user’s first query, rather like a conventional search engine. ChatGPT, by contrast, bides its time. Less than half its ads come in the chatbot’s first response, and nearly a third come after the tenth turn in the conversation—like a shop assistant waiting for the customer’s intent to become clear before they make their pitch. “This is not search. This is a new category of advertising,” argues Harel Amir of Similarweb.

早期证据表明，OpenAI和谷歌对聊天机器人应该如何推销有不同的方法。根据 Similarweb 的数据，谷歌“AI模式”中98.5%的广告出现在对用户首次查询的回复中，这很像传统的搜索引擎。相比之下，ChatGPT则在等待时机。它不到一半的广告出现在聊天机器人的首次回复中，而近三分之一的广告出现在对话的第十个回合之后——就像店员在推销之前等待顾客的意图变得明确一样。“这不是搜索。这是一个新的广告类别，”Similarweb 的哈雷尔·阿米尔（Harel Amir）认为。

So far the results are mixed. ChatGPT’s targeting sometimes seems sophisticated: for instance, advertising job-interview coaching to a user drafting a professional email. At other times it is comically literal. One user asking about cryptographic “private keys” was served an ad for metal safety-boxes from Bed Bath & Beyond. OpenAI’s fledgling ad service so far offers advertisers little in the way of performance data or the ability to target specific demographic groups, says Caelean Barnes of Gauge, which helps brands track how chatbots talk about them.

到目前为止，结果好坏参半。ChatGPT的定位有时似乎很复杂精细：例如，向起草专业电子邮件的用户宣传求职面试辅导。但在其他时候，它又滑稽地死板（字面理解）。一位询问密码学“私钥（private keys）”的用户，收到了 Bed Bath & Beyond 提供的金属保险箱（需用钥匙）的广告。Gauge 公司帮助品牌追踪聊天机器人如何谈论它们，该公司的凯利安·巴恩斯（Caelean Barnes）表示，OpenAI刚刚起步的广告服务迄今在性能数据或定位特定人群群体方面，能为广告商提供的帮助很少。

OpenAI and Google also seem to be going after different kinds of advertisers. Google’s Direct Offers allow companies to promote a discount code for a product the user is researching in AI Mode. OpenAI, meanwhile, is focused on brand-building ads designed to burnish a firm’s image rather than trigger a click. So far 81% of its ads fall into that category, by Similarweb’s analysis. OpenAI may reckon that users deep in a coding or editing session are unlikely to break off to pursue an offer.

OpenAI和谷歌似乎也在追求不同类型的广告商。谷歌的 Direct Offers 允许公司为用户在“AI模式”中研究的产品提供折扣代码进行促销。与此同时，OpenAI则专注于品牌建设广告，旨在提升公司形象，而不是触发点击。根据 Similarweb 的分析，到目前为止，其81%的广告属于该类别。OpenAI可能认为，正沉浸在编码或编辑会话中的用户不太可能中断手头的工作去追逐某个优惠。

The good news for chatbot-makers is that users do not seem to be repelled by ads. Some competitors have steered clear of sponsored replies, fearing these will hurt trust: in February Perplexity stopped showing ads and Anthropic skewered its ad-toting rivals with a series of skits showing therapists and professors suddenly launching into sales pitches. Yet Similarweb finds that 72% of ads shown on ChatGPT are not on the final turn of the conversation—in other words, users keep chatting. Conversations incorporating ads last an average of about 20 turns, the same as those without.

对聊天机器人制造商来说，好消息是用户似乎并不排斥广告。一些竞争对手避开了赞助回复，担心这会损害信任：今年2月，Perplexity 停止显示广告，Anthropic 则用一系列短剧（展示治疗师和教授突然开始推销产品）讽刺了那些携带广告的竞争对手。然而 Similarweb 发现，在ChatGPT上显示的广告中，有72%并不是在对话的最后一回合——换句话说，用户仍在继续聊天。包含广告的对话平均持续约20个回合，与没有广告的对话相同。

For advertisers, AI ads present two challenges. The first is measurement. If users keep chatting rather than clicking through, “post-view attribution”—the murky science of determining whether a user eventually buys the product elsewhere—will become the main way to assess a campaign’s effectiveness. The second challenge is brand safety. Mark Zuckerberg has painted a picture of a future where advertisers simply tell Meta their objective and let the AI handle the rest. But letting a chatbot write its own creative copy is a gamble. Personalised ads have proved to be more effective than generic ones. Companies may nevertheless be wary of a silicon salesman who could hallucinate a deal too good to be true. ■

对于广告商来说，AI广告提出了两个挑战。首先是测量。如果用户继续聊天而不是点击进入（链接），“浏览后归因”——即确定用户最终是否在其他地方购买了产品的这门并不明朗的科学——将成为评估营销活动效果的主要方法。第二个挑战是品牌安全。马克·扎克伯格描绘了一幅未来的图景：广告商只需告诉Meta他们的目标，然后让AI处理剩下的事情。但让聊天机器人自己撰写创意文案是一场赌博。个性化广告已被证明比通用广告更有效。尽管如此，公司可能还是会对一个可能会“产生幻觉”、编造出好得令人难以置信的交易的“硅推销员”保持警惕。 ■

<翻译仅供参考>

1. Plugs: 推广，插播广告，非正式用语，指在节目、文章或对话中顺带提及或宣传某产品、服务或公司。

2. Fledgling: 刚起步的，羽翼未丰的，形容某个事物、组织或系统刚刚建立，缺乏经验。

3. Burnish: 擦亮，提升（形象），指改善某人或某公司的声誉，使其看起来更完美或更具吸引力。

4. Skewered: 严厉批评，讽刺，原意为用烤肉签串起来，比喻用尖锐的言辞无情地嘲笑或抨击。

5. Hallucinate: 产生幻觉，胡说八道，在AI领域特指大语言模型生成看似合理但事实上完全错误或捏造的信息。

「熊见」小调研

邀请一起建设更好的「熊见」

关注我们，每日更新