当你的AI助手,开始向你卖东西|经济学人

Why your Al assistant is sud- denly selling to you

Sponsored chats are changing the way digi- tal advertising works 

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

As chatbots become adbots,the future of two industries is at stake. For the arti- ficial-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 exis- tential question of how advertising will work if users move away from conven- tional search engines. Although it is early days,the outlines of a new kind of marketing are emerging.

Many of the Al 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 “Al Mode” since last year, and Microsoft has woven them into its Copilot. Amazon’s shop- ping 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-tar- geted plugs.

For OpenAI, which reportedly expects to burn $25bn of cash this year and twice as much in 2027,the need for more rev- enue is acute. Although it sells subscrip- tions, most of ChatGpT’s more than goom 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 immedi- ate cashflow worries,has reason to bring ads to Al 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.

So far the experiments are modest. Simi- larweb, a data provider, estimates that ChatGPT is showing ads in only about1% of conversations, and Google in an even smaller share of Al Mode queries. But they are due to rise fast. OpenAl is said to have told investors that it expects ChatcPT to generate $2.5bn in ad revenue this year and $ubn in 2027, with a goal of hitting $1oobn by 2030. The early evidence suggests that OpenAl 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 con- ventional 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 cus- tomer’s intent to become clear before they make their pitch.”This is not search. This is a new category of adver- tising,” argues Harel Amir of Similarweb. 

So far the results are mixed. ChatGPT’s targeting sometimes seems sophisticat- ed: for instance, advertising job-inter- view 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.OpenAl’s fledgling ad service so far offers advertisers little in the way of performance data or the abil- ity to target specific demographic groups, says Caelean Barnes of Gauge, which helps brands track how chatbots talk about them.

OpenAI and Google also seem to be going after different kinds of advertis- ers. Google’s Direct Offers allow companies to promote a discount code for a product the user is researching in Al Mode.OpenAl, 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 catego- ry,by Similarweb’s analysis. OpenAl may reckon that users deep in a coding or editing session are unlikely to break off to pursue an offer.

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 pitch- es. 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. Conversa- tions incorporating ads last an average of about 20 turns, the same as those without.

For advertisers, Al ads present two chal- lenges. 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 else- wherewill 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 Al 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.