Is Google’s 20-year search dominance about to end?

谷歌长达20年的搜索引擎霸主地位即将终结了吗？

The rise of ChatGPT-like AI applications has profound implications for internet use

类似ChatGPT的人工智能应用兴起对互联网的使用有着深远的影响

Near the bay in Mountain View, California, sits one of biggest profit pools in business history. The site is the home of Google, whose search engine has for the past two decades been humanity’s preferred front door Ito the internet—and advertisers’ preferred front door to humanity. Every second of every day, Google processes perhaps 100,000 web searches around the world—and, thanks to its “PageRank” algorithm, serves up uncannily relevant answers. That has conferred onto Google verb status. It also adds up to billions of daily opportunities to sell ads that the searchers see alongside the results of their queries. The results’ accuracy keeps users coming back, and rivals at bay: all other search engines combined account for barely a tenth of daily searchers (see chart 1).

在加州山景城的海湾附近，坐落着商业史上最大的利润池之一。这里是谷歌的所在地，谷歌搜索引擎在过去的二十年里一直是人类首选的互联网门户，也是广告商的首选。每一天每一秒，谷歌都会处理全球大约10万次网络搜索——并且，得益于其“PageRank”算法，谷歌能提供了不可思议的相关答案。这也让“谷歌一下”成为搜索的代名词。这也意味着谷歌每天有数十亿的机会出售广告，搜索者可以在搜索结果旁边看到这些广告。搜索结果的准确性让用户成为回头客，竞争对手只能望而却步：所有其他搜索引擎加起来只占每日搜索量的十分之一（见图表1）。

Advertisers pay handsomely for access to Google’s users, not least because they are typically only charged when someone actually visits their website. The revenue of Google’s parent company, now called Alphabet, has grown at an average annual rate of over 20% since 2011. In that period it has generated more than $300bn in cash after operating expenses (see chart 2), the bulk of it from search. Its market value has more than trebled, to $1.4trn, making it the world’s fourth-most-valuable public company. Unlike Apple and Microsoft, its bigger middle-aged tech rivals, it has not felt the need to reinvent itself.  Until now.

广告商为将广告投放给谷歌用户，往往支付丰厚的费用，尤其是因为他们通常只在有人真正访问他们的网站时才被收费。自2011年以来，谷歌母公司Alphabet的收入年均增长超过20%。在此期间，扣除了运营费用后，它产生超过3000亿美元的现金（见图表2），其中大部分来自搜索业务。公司市值增长了两倍多，达到1.4万亿美元，成为全球第四大市值上市公司。与苹果和微软这两个规模更大的保守派科技竞争对手不同，谷歌并不觉得有必要重塑自己。直到现在。

The reason for the soul-searching in Mountain View is Chatgpt, an artificially intelligent chatbot designed by a startup called Openai. Besides being able to have a human-like conversation, Chatgpt and others like it can draft equally human-seeming poems, history essays, computer code and just about anything else that people write down, paint, compose or sing. ubs, a bank, reckons that since its launch in November Chatgpt has gained around 100m monthly active users, a feat that had taken TikTok, the world’s fastest-growing social-media platform, nine months. Bill Gates, co-founder of Microsoft, has called this sort of “generative” ai “as important as the pc, as the internet”.

谷歌进行自我反省的原因是ChatGPT，这是一家名为OpenAI初创公司设计的人工智能聊天机器人。除了能够进行近似人类的对话，ChatGPT和其他类似的软件还可以起草同样类人的诗歌、历史论文、计算机代码以及人们书写、绘画、作曲或演唱的任何东西。瑞士银行估计，自去年11月推出以来，ChatGPT每月活跃用户增加了约1亿，这是TikTok(全球增长最快的社交媒体平台)9个月来的成就。微软联合创始人比尔·盖茨称这种“生成性”人工智能“与个人电脑和互联网一样重要”。

To Eric Schmidt, who used to run Google, Chatgpt is the “first broadly visible example” of what a human being’s ai friend might look like. To his former employer, it is the first broadly visible threat to Google’s search dominance. For Chatgpt can also answer the sorts of questions that people might have asked Google. And Chatgpt’s creator, Openai, has teamed up with Microsoft, which is covetously eyeing Google’s gleaming pool of profits.

对于曾经运营谷歌的埃里克·施密特来说，ChatGPT是“第一个广泛可见的例子”，展示了人类的人工智能朋友可能是什么样子。对他的前雇主来说，这是对谷歌搜索主导地位的第一次显而易见的威胁。因为ChatGPT还可以回答人们可能会问谷歌的各种问题。ChatGPT的创建者OpenAI已经与微软合作，后者正贪婪地盯着谷歌闪闪发光的利润池。

On February 7th Microsoft, which recently announced an investment of $10bn in Openai, showed off how it plans to go after those profits. Results from Microsoft’s search engine, Bing, will now be accompanied by an ai-generated side box summarising pertinent information. Bing will also get its own chatbot based on Openai’s models. Microsoft executives showed off neat tricks, such as producing a shopping list based on a week of planned meals and creating a music-trivia quiz. “It is a new day in search,” declared Microsoft’s ceo, Satya Nadella. Baidu, which dominates search in China, where Google is banned, plans to launch its own ai-boosted search in March. In a pre-emptive response, and to steal some of Mr Nadella’s thunder, Alphabet has in recent days unveiled its own chatbot, Bard, and reportedly invested $400m in Anthropic, another generative-ai startup. On February 8th, while presenting some non-chatty ai search features, it confirmed that Bard will be integrated into search within weeks. Investors were unimpressed; Alphabet’s share price tumbled by 8% after the announcement.

2月7日，最近宣布向Openai投资100亿美元的微软展示了它计划如何追求这些利润。来自微软搜索引擎必应的搜索结果现在将伴随着一个人工智能生成的侧框，用于总结相关信息。必应也将基于OpenAI的模型获得自己的聊天机器人。微软的高管们展示了一些巧妙的技巧，比如根据一周的饮食计划制作一份购物清单，以及制作一个音乐问答游戏。“这是搜索引擎的革新，”微软首席执行官塞特亚·纳德拉宣称。在谷歌被禁的中国，主导搜索的百度计划在3月份推出自己的人工智能搜索。作为先发制人的回应，也为了抢纳德拉的风头，Alphabet最近几天推出了自己的聊天机器人Bard，据报道，还向另一家生成性人工智能初创公司Anthropic投资了4亿美元。2月8日，在展示一些非聊天型的人工智能搜索功能的同时，谷歌证实Bard将在几周内集成到搜索中。投资者不为所动；声明发布后，Alphabet的股价暴跌了8%。

The battle of the bots is, then, brewing. It will transform the way people find things on the internet. In doing so, it may upend the lucrative business of search.

互联网大厂的人工智能之战正在酝酿之中。它将改变人们在互联网上找东西的方式。这样做可能会颠覆利润丰厚的搜索业务。

The last time online search was disrupted was at the dawn of the consumer internet in the late 1990s. As the number of webpages grew into the millions, useful information became harder to find. A number of search engines, such as AltaVista and Yahoo!, made things a bit easier. But it was Google, founded in 1998, that revolutionised the industry. Its algorithm ranked webpages based on the number of other websites linking to them, which turned out to be a good proxy for relevance. Soon after, Google worked out that it could display ads related to users’ search keywords alongside the results.

上一次在线搜索收到冲击是在20世纪90年代末消费者互联网出现的时候。随着网页的数量增长到数百万，有用的信息变得更难找到。一些搜索引擎，如AltaVista和雅虎，让事情变得简单了一些。但是，是成立于1998年的谷歌彻底改变了这个行业。谷歌运用的算法根据链接到网页的其他网站数量对网页进行排名，这被证明是相关性的优秀运用案例。不久之后，谷歌发现它可以在搜索结果旁边显示与用户搜索关键词相关的广告。

In the past few years challengers have tried to take on Google. Some are startups offering ad-free subscription search for consumers, such as Neeva and You.com. And Alphabet’s big-tech rivals are getting in on the action. Amazon, which has become the place where many shoppers start looking for products, has seen its share of the American search market jump from 3% in 2016 to 23% today. Apple’s ads business, consisting of searches for apps on iPhones, has gone from nothing a few years ago to 7%. Google’s own research shows that two-fifths of 18-to-24-year olds favour Instagram, Meta’s photo-sharing app, or TikTok over Google Maps when searching for a nearby restaurant.

在过去的几年里，挑战者们试图挑战谷歌。一些初创公司为消费者提供无广告的订阅搜索，如Neeva和You.com。Alphabet的大型科技竞争对手也加入了进来。亚马逊已经成为许多购物者开始寻找产品的地方，其在美国搜索市场的份额从2016年的3%跃升至今天的23%。苹果的广告业务，包括在iPhone上搜索应用程序，已经从几年前的零增长到7%。谷歌自己的研究显示，在搜索附近的餐馆时，五分之二的18至24岁年轻人更喜欢Instagram、Meta的照片分享应用程序或抖音，而不是谷歌地图。

As a result of this ferment, Google’s share of revenue from search advertising in America will fall to 54% this year, down from 67% in 2016, according to eMarketer, a research firm. But these challengers never presented an existential threat to Google. The same cannot be said of chatbot-assisted conversational search. After Chatgpt’s launch Mr Pichai reportedly declared a “code red” at his company.

根据研究公司eMarketer的数据，由于这种动荡，谷歌在美国搜索广告收入中的份额将从2016年的67%下降到今年的54%。但这些挑战者从未对谷歌构成生存威胁。聊天机器人辅助的对话搜索就不一样了。据报道，在ChatGPT推出后，皮查伊先生宣布他的公司进入“红色警报”。

To understand why Google fears the chatbot and its copycats, first consider the technology behind them. Chatgpt works by predicting the next word in a sentence that is a response to some query. These predictions are based on a “large language model”, itself the result of prior analysis of millions of texts, harvested from the internet. Once trained on this giant corpus of natural language, the chatbot can, when prompted by users’ instructions, produce a fluently written answer rather than merely serving up a list of links.

要理解谷歌为什么害怕聊天机器人及其模仿者，首先要考虑它们背后的技术。ChatGPT的工作原理是预测句子中的下一个单词，该单词是对某个查询的响应。这些预测是基于一个“大型语言模型”，这个模型本身是对从互联网上收集的数百万文本进行事先分析的结果。一旦在这个庞大的自然语言语料库上接受训练，聊天机器人就可以在用户指令的提示下，产生流利的书面答案，而不仅仅是提供一系列链接。

Applied to search queries, this means that responses could in principle contain many more variables. Want to go on a day trip somewhere off the beaten track that is cheap, child-friendly and educational? Unless you chance upon a travel blogpost, today finding a precise answer on Google (or Bing or Baidu) requires comparing dozens of websites and skim-reading tonnes of text. By contrast, Chatgpt spits out a list of decent options in moments. Users can then add extra considerations or ask for more information with follow-up queries.

应用于搜索查询，这意味着响应原则上可以包含更多的变量。想去一个便宜、适合儿童且有教育意义的人迹罕至的地方进行一日游吗？除非你偶然发现一篇旅游博客，否则今天在谷歌（或必应或百度）上找到一个准确的答案需要比较几十个网站和浏览大量文本。相比之下，ChatGPT很快就会列出一系列不错的选择。然后，用户可以添加额外的考虑因素，或者通过后续查询询问更多信息。

Changing how people search will, in turn, change what they search for. As well as seeking existing information, users can use conversational search to generate completely original content. Chatgpt writes poetry and essays—in the style of your favourite author, if you wish. On January 26th Google published a paper describing Musiclm, a new large language model that can make music from text. GitHub, a Microsoft-owned platform which hosts open-source programs, has a chatbot called Copilot that can churn out lines of code. That opens up all-new search-adjacent markets, argues Mark Shmulik of Bernstein, a broker, most immediately in productivity tools for businesses (like helping desk jockeys write presentations).

改变人们的搜索方式反过来会改变他们搜索的内容。除了寻找现有的信息，用户还可以使用对话式搜索来生成完全原创的内容。ChatGPT写诗歌和散文——如果你愿意，可以用你最喜欢的作家的风格。1月26日，谷歌发表了一篇描述Musiclm的论文，这是一种新的大型语言模型，可以从文本中制作音乐。微软旗下托管开源程序的平台GitHub有一个名为Copilot的聊天机器人，可以大量生成代码行。伯恩斯坦研究公司的经纪人马克·什穆里克认为，这打开了全新的搜索市场，最直接的观点是企业生产力工具（比如帮助办公桌工作人员写演示文稿）。

A little more conversation…

再聊一会儿……

As a new area, conversational search is attracting hopeful newcomers, buoyed by the prospect of an expanding market for search and generative content. “When I started two years ago, people said I was crazy. Now the sentiment has massively shifted,” says Richard Socher, founder of You.com, which also offers an embedded ai-powered chatbot. Neeva has also added a chatbot to its subscription search. Sridhar Ramaswamy, its co-founder, hopes this will help it reach 5m-10m subscribers, up from almost 2m today (not all of them paying subscribers), and become financially self-sustaining. c3.ai, a business-software firm, has begun offering a chatbot to help companies search their in-house data. Travel companies, including Booking.com, are toying with chatbots, too.

作为一个新领域，受到搜索和生成内容市场不断扩大的前景鼓舞，对话式搜索正在吸引有希望的新来者。“当我两年前开始的时候，人们说我疯了。现在这种情绪已经发生了巨大的变化，”You.com的创始人理查德·索彻说，该公司也提供嵌入式人工智能聊天机器人。Neeva还在其订阅搜索中添加了一个聊天机器人。其联合创始人斯里达尔·拉马斯瓦米希望这将帮助其达到500万至1000万活跃用户，而目前只有近200万用户（并非所有用户都是付费用户），并实现财务自给自足。商业软件公司c3.ai已经开始提供聊天机器人来帮助公司搜索他们的内部数据。包括Booking.com在内的旅游公司也在调制聊天机器人。

The most serious threat comes from Microsoft. Its big advantage is that it already has the search infrastructure, including oodles of computing power, storage systems and armies of web-crawling programs that constantly scrape information from the internet. Setting all this up from scratch to compete with Google would, according to one estimate by the Competition Markets Authority, a British trustbusting agency, cost between $10bn and $30bn.

最严重的威胁来自微软。最大优势是它已经拥有了搜索基础设施，包括大量的计算能力、存储系统和大量不断从互联网上抓取信息的网络抓取程序。根据英国反信任机构竞争市场管理局的一项估计，从零开始建立这一切以与谷歌竞争，将花费100亿至300亿美元。

At present Bing’s share of America’s search-advertising market is a measly 5% (see chart 3). Microsoft hopes that Bing’s new bag of tricks will change that. Importantly, the company seems to have solved some of Chatgpt’s shortcomings. One was keeping the bot up to date. Chatgpt’s underlying ai, called gpt-3.5, has been trained on data from 2021 and thus has no knowledge of anything on the internet after that point. Ask Chatgpt about recent news or today’s weather forecast and you get an apology. Bing’s ai, by contrast, decides how to gather the most relevant information and then uses search tools to find it. The data are then fed back into the model, which uses them to compose a fluent answer. Other firms, including Neeva, are using this method, too.

目前，必应在美国搜索广告市场的份额只有区区5%（见图表3）。微软希望必应的新锦囊妙计能改变这种状况。重要的是，该公司似乎已经解决了ChatGPT的一些缺点。一个是让机器人保持最新。ChatGPT的底层人工智能名为gpt-3.5，已经根据2021年的数据进行了训练，它在那之后对互联网上的任何事情都一无所知。向ChatGPT询问最近的新闻或今天的天气预报，他会给你道歉。相比之下，必应的人工智能将决定如何收集最相关的信息，然后使用搜索工具找到它。然后，这些数据被反馈到模型中，模型使用这些数据来组成一个流畅的答案。包括Neeva在内的其他公司也在使用这种方法。

This in turn has helped Microsoft tackle a bigger technological problem: large language models’ tendency to make stuff up. Chatbots have no sense of what is true or false; they reflect what is on the internet, warts and all. These authorititive-sounding “hallucinations”, in computer-science speak, are innocuous when the chatbot is used for fun and games. When it is meant to give real answers to serious questions, they are a fatal flaw. Last year Meta had to take down its science chatbot, Galactica, after it was found to be spouting scientific nonsense.

这反过来帮助微软解决了一个更大的技术问题：大型语言模型倾向于编造东西。聊天机器人对于真假没有感觉；它们反映了互联网上的内容，包括其中所有的缺点。用计算机科学的话说，当聊天机器人被用于娱乐和游戏时，这些听起来很权威的“幻觉”是无害的。当打算对严肃的问题给出真正的答案时，它们会带来致命的缺陷。去年，Meta不得不关闭它的科学聊天机器人Galactica，因为它屡屡在胡说八道。

Giving the model access to up-to-date data has reduced, though not eliminated, the hallucinations rate for Bing’s chat feature. “A lot of the hallucinations were [the model] trying to fill in the blanks on things that had happened since the end of its training data,” explains Kevin Scott, Microsoft chief technology officer. He adds that his firm is using other techniques to lower the rate further. These include having humans tell the models which answers are better and what information is reliable; adding memory to the systems, so that the algorithms could learn from the conversations, which they currently do not; and posting links to sources in the ai-generated responses. Mr Schmidt similarly expects chatbots to be completely sober within a year or two.

让模型访问最新数据降低了必应聊天功能的幻觉率，尽管没有完全消除。微软首席技术官凯文·斯科特解释说：“很多幻觉都是模型试图填补自训练数据结束以来发生的事情的空白。”他补充说，他的公司正在使用其他技术来进一步降低幻觉率。这些包括让人类告诉模型哪些答案更好，哪些信息是可靠的；给系统增加内存，这样算法就可以从对话中学习，而目前还没有；并在人工智能生成的响应中发布链接。施密特同样预计聊天机器人将在一两年内完全清醒。

Solving the technological problems is only the first step to dislodging Google from the search pedestal. No less knotty is working out how conversational search will make money. Start with the costs, which for chatbots remain high compared with conventional search. Brian Nowak of Morgan Stanley, a bank, estimates that serving up an answer to a Chatgpt query costs roughly two cents, about seven times more than a Google search, because of the extra computing power required. He reckons that every 10% of Google searches that shift to an ai-powered application by 2025 will, depending on the number of words in an average response, add between $700m and $11.6bn, to Google’s operating costs, equivalent to between 1% and 14% of such spending in 2022.

解决技术问题只是将谷歌从搜索引擎王座上赶下来的第一步。同样棘手的是研究对话式搜索如何赚钱。首先是成本，与传统搜索相比，聊天机器人的成本仍然很高。摩根士丹利银行的布莱恩·诺瓦克估计，由于需要额外的算力，提供一个ChatGPT查询的答案大约需要2美分，大约是谷歌搜索的7倍。他估计，到2025年，每10%的谷歌搜索转向人工智能应用，根据平均回复中的字数，将为谷歌增加7亿至116亿美元的运营成本，相当于2022年支出的1%至14%。

To complicate matters further, many costly conversational-search queries will generate little or no ad revenue. Google has said that 80% of its search results do not contain lucrative ads at the top of the search results. Many of these ad-light searches are almost certainly “informational” (“what is the capital of Spain?”), precisely the sort of query where chatbots are most useful—and precisely the sort that advertisers are least interested in (it is hard to know what ad to place next to the word “Madrid”). For generative ai to make real money, it will need to find uses in “navigational” searches (looking for a site’s internet address by its name) and especially “commercial” ones (“Best new ski boots this season”).

更复杂的是，许多昂贵的对话式搜索查询将带来很少或压根没有广告收入。谷歌表示，其80%的搜索结果中，顶部并不包含利润丰厚的广告。这些轻广告搜索中有许多几乎可以肯定是“信息性的”（“西班牙的首都是哪里？”），这正是聊天机器人最有用的查询，也正是广告商最不感兴趣的（很难想出要在“马德里”这个词旁边放什么广告）。为了让生成性人工智能真正赚钱，它需要在“导航”搜索（通过名称寻找网站的互联网地址）中找到用途，尤其是“商业”搜索（“本季最佳新滑雪靴”）。

Some companies, such as Neeva, earn revenue from subscriptions. On February 1st Openai began signing up subscribers to the current version of Chatgpt. For $20 a month users get faster responses and access during peak times. Openai also plans to license the technology directly to other firms. But the big bucks are likely to lie in advertising.

一些公司，如Neeva，从订阅中获得收入。2月1日，Openai开始注册当前版本的ChatGPT用户。每月20美元，用户可以在高峰时段获得更快的响应和访问。Openai还计划将这项技术直接授权给其他公司。但是赚大钱的是广告。

Embedding ads in what is meant to feel to a user like a normal conversation will require deftness. One possibility is to display fewer adverts but charge advertisers more for each, says Mr Nowak. A chatbot is likely to offer up only a few suggestions in response to a search query about, say, blissful Hawaiian hotels. Hoteliers may be happy to pay more to ensure that theirs are among those suggestions, or displayed next to them. Microsoft says that it plans to test such a model in the new-look Bing.

将广告嵌入到让用户感觉像正常对话的内容中需要技巧。诺瓦克先生说，一种可能性是展示更少的广告，但向广告商收取更多的费用。一个聊天机器人很可能只提供一些建议来回应一个搜索查询，比如说，夏威夷的幸福酒店。酒店经营者可能很乐意支付更多的钱，以确保他们能出现在这些建议中，或者显示在问答框旁边。微软计划在新外观的必应中测试这种模式。

Microsoft may be betting that its chatbot-assisted informational search will lure more users, who will then use Bing for the more lucrative queries as well. This would almost certainly mean sacrificing margins, at least until costs can be brought down. It would only be worth it if they can wrest sizeable market share from Google. Microsoft expects that for every percentage point of market share they gain in search, its annual advertising revenue will grow by $2bn.

微软可能押注于其聊天机器人辅助的信息搜索将吸引更多用户，这些用户也将使用必应进行更有利可图的查询。这几乎肯定意味着牺牲利润，至少在成本降低之前是如此。只有当他们能从谷歌手中夺取相当大的市场份额时，这才是值得的。微软预计，他们在搜索领域每获得一个百分点的市场份额，其年度广告收入将增长20亿美元。

That is possible—but not guaranteed. Alphabet retains some formidable strengths. One is technology. Although Google has yet to integrate generative ai into its search engine, it has been deploying other ais in its search business for years. When it highlights a web-page snippet at the top of the search results, that is courtesy of models such as Bert and Mum. Its new generative ai, Bard, is likely to be no less impressive than Chatgpt. The reason the company has given for its foot-dragging is fear about unleashing a chatbot that spewed controversial content.

这是可能的，但不能保证。Alphabet保留了一些强大的优势。一个是技术。尽管谷歌尚未将生成式人工智能集成到其搜索引擎中，但多年来，它一直在搜索业务中部署其他人工智能。当它在搜索结果顶部突出显示一个网页片段时，这得益于Bert和Mum等模型。谷歌的新生成人工智能Bard模型可能不会比ChatGPT更令人印象深刻。该公司之所以发布的如此扭扭捏捏，是因为担心该聊天机器人会有争议内容。