Keyword Search Is Dying. Is Your Catalog Ready for AI Agents? — PayPal

频道: AI Engineer
视频: https://www.youtube.com/watch?v=YXb2Wx1r_Pk
原文语言: en
统计: 共 10 轮 · Nixon 10

[0:01] Nixon

[music] Hi everyone. Um, welcome. I I'm Nixon Den. I'm the director of product for Aenticcommerce at PayPal. And what that really means is I lead a lot of our efforts on how AI agents are reshaping shopping and payments. Um, and you know, most of my career has been on commerce and payments. And right now, I'm really excited about helping merchants navigate this shift uh towards like agent-driven commerce. And specifically, what we're going to talk about today is um the catalog piece and it being agentic ready and sharing some of our learnings on this on this journey. So I think it's very important we just kind of level set a little bit when we look at the evolution of commerce. We've seen it transform from you install brick and mortar transactions to the advent of e-commerce where this started off quite linear. Um a user would simply search and buy from e-commerce sites and then through like internet 2.0 O we've seen that transform further into more complex uh channels that involve email, SMS and like now social commerce for us like we see AI as the next evolution in that journey and with any evolution we know two things to be true. The first like all change it can seem like a threat to a business. Um but on the flip side of that can also be seen as a big opportunity.

[音乐] 大家好，欢迎。我是 Nixon Dinh，PayPal 的 Agentic Commerce（代理式电商，即由 AI agent 参与甚至代办的购物）产品总监。说白了，就是我在牵头做很多这方面的事：AI agent 正在怎样重塑购物和支付。我的职业生涯大部分都在做电商和支付，现在最让我兴奋的，是帮商家走过这场转向 agent 驱动电商的变迁。今天具体要聊的是商品目录（catalog）这一块——它怎么做到「agentic ready」（为 agent 准备就绪），以及我们一路上学到的一些东西。

我觉得有必要先把背景对齐一下，看看商业是怎么演化的。它从线下实体店的交易，走到电商出现——一开始路径很线性，用户就是在电商网站上搜索、然后购买；再经过 Web 2.0，演变成更复杂的渠道，有了邮件、短信，以及现在的社交电商。对我们来说，AI 就是这段旅程的下一次演化。而任何一次演化，都有两件事是确定的。第一，和所有变化一样，它对一家企业来说可能看起来像威胁；但反过来看，它也可以是一个巨大的机会。

[1:37] Nixon

And the second thing that we know is that whenever we have seen consumer behavior change because of a new technology, commerce tends to evolve alongside with it. And so we are starting to see that now with Aentic where not only is the new this is a new channel that's enabling uh merchants to engage with customers in a very deeply personalized way, but also how customers are searching and finding products is changing. And uh now is really the time that merchants are starting to look to invest on how to to change with with their customers and and uh succeed in the agentic era. So earlier I touched on how the consumer behavior is starting to change. Um but I just want to dive deeper a little bit on this. So we, you know, when we first uh when we think about consumer shopping, that behavior up to today has been now what we're calling the search era. Um it goes through a phase where like it's keywords, you're navigating a catalog, there's SEO strategies, and this is what most businesses have optimized towards over the last decade. And while for many businesses this worked, this model put the burden on the shopper to really find what they needed. Um, now what we're seeing is that we're moving into this phase called the intent error. So instead of typing keywords like I want school supplies, um, I might say something like, you know, my daughter's starting school in September in fifth grade and I need uh to to shop back to school. And that's a very completely different interaction.

第二件确定的事是：每当消费者行为因为一项新技术而改变，商业往往会跟着一起演化。我们现在在 agentic 上已经开始看到这一点——它不只是一个新渠道，让商家能以高度个性化的方式触达顾客；连顾客搜索和发现商品的方式也在变。现在正是商家开始考虑投入的时候：怎么跟着顾客一起变，在 agentic 时代里做成。

前面我提到消费者行为开始变了，这里想再往深讲一点。说到消费者购物，到今天为止的那套行为，我们现在把它叫作「搜索时代」（search era）。那个阶段是关键词、在目录里一层层翻找、各种 SEO 策略，过去十年大多数企业都是朝这个方向优化的。对很多企业来说这套确实管用，但这个模式把负担压在了购物者身上——得由他们自己去找到需要的东西。

而我们现在看到的，是正在进入一个叫「意图时代」（intent era）的阶段。我不再敲「我要学习用品」这样的关键词，而是可能会说：「我女儿九月开学，上五年级，我要给她采购开学用品。」这是一种完全不同的交互。

[3:22] Nixon

Um, so you're now expressing your needs in your own words and AI is now asked to meet you where you are. And so the brands are coming to you instead of you having to go find them. And that's a very very different interaction model. Um, and most businesses now are starting to find that they need to set themselves up for this interaction model. Um so what we believe will happen next is after we go through the intent error what will have to happen is customers need to start trusting the recommendations that they're getting from the from the agents and once that trust starts to build we do start to believe that they will start delegating more and more actions to their agents and that's also where as PayPal we're building towards is um continue to build on that trust and moving to a world where humans start to delegate more and more of their task in commerce and shopping to their agents. So although we talk about this delegation era and we think it sounds like it's so far into the future, you might be asking like why this matters now. And so we have couple pretty big data points that we want to share to really underscore why this matters. The first is uh through a study at Bay & Co. um they project that by 2030 15 to 25% of all commerce will be executed by an agent and furthermore within a decade agents are predicted to outnumber humans online and so these numbers don't just indicate that this is like not meaningful now it is a very marginal shift that businesses can't ignore and that this is a very structural transformation happening entire entire e-commerce ecosystem So to go further I think it's very good to understand why PayPal feels like we are bullish on this opportunity. Um as a industry we expect uh to be some very tangible opportunities coming forward. We expect 1.1 trillion dollars of uh commerce to happen uh in US retail to be agentic and then we expect that to be 3 to 5 trillion globally and what

你现在是用自己的话表达需求，而 AI 被要求到你所在的地方来迎合你。于是变成了品牌来找你，而不是你去找品牌。这是一种非常非常不同的交互模式，大多数企业现在也开始发现，自己得为这种交互模式做好准备。

那接下来会发生什么？我们的判断是：走过意图时代之后，必须发生的一件事是，顾客要开始信任自己从 agent 那里拿到的推荐。一旦这种信任开始建立起来，我们确实相信，他们会把越来越多的动作委托给自己的 agent。这也是 PayPal 正在努力的方向——持续把这份信任筑牢，走向一个人们把越来越多电商和购物任务委托给 agent 的世界。

我们讲这个「委托时代」（delegation era），听上去好像还很遥远，你可能会问：那为什么现在就重要？我们有几个挺有分量的数据点想分享，用来说明这一点。第一个来自 Bain & Company（贝恩公司）的一项研究：他们预测到 2030 年，全部商业交易的 15% 到 25% 将由 agent 来执行；另外，十年之内，网上 agent 的数量预计会超过人类。这些数字说明，这不是一个企业可以忽略的边角变化，而是整个电商生态正在发生的一场结构性转变。

再往下说，我觉得有必要讲清楚 PayPal 为什么看好这个机会。作为一个行业，我们预期会出现一些非常实在的机会：我们预计美国零售里会有 1.1 万亿美元的交易变成 agentic 的，全球则是 3 到 5 万亿美元。而

[5:46] Nixon

we've have seen is that there are over two billion users um projected to use like AI to start their shopping journey over the next few years. And so when we think through this, what does this actually mean for businesses and merchants? When we talk to our customers and what we talk to our partners, what we've seen is six they what they've reported is that over 693% uh referral traffic from AI engines. And when we did consumer studies, we found that 39% of US shoppers are already using AI for shopping today. Um, and I'm sure many of you have probably defaulted to that pattern today. And what we have found also is that when you do start and interact with an agent first, businesses have tend to see up to four times higher conversion on a completion of an order because now they can get more rich and in-depth data on a product. And so with that all happening and the traffic moving, we want to break this down further onto like what this really means ac deeper for merchants. In the search era, this was very like SEO based and you might optimize for like a sum dress. In the intent era, you have to think through how do I be discoverable for where an agent is representing my my business. And then in the delegation era, we need to think through uh how I can convert that demand safely and trusted in a trusted way.

我们看到的是，未来几年预计会有超过 20 亿用户用 AI 来开启自己的购物旅程。

那想到这里，这对企业和商家到底意味着什么？我们和客户聊、和合作伙伴聊，他们反馈的情况是：来自 AI 引擎的引荐流量增长了超过 693%。我们做消费者调研时发现，39% 的美国购物者今天已经在用 AI 购物了——我相信在座很多人今天大概也已经默认这么做了。我们还发现，当购物确实是先从和 agent 互动开始的，商家在订单完成这一步看到的转化率往往能高到四倍，因为这时候能拿到关于一件商品更丰富、更深入的信息。

既然这一切正在发生、流量正在迁移，我们想把它对商家更深一层的含义拆开讲。在搜索时代，这件事非常以 SEO 为中心，你可能会针对「sundress（太阳裙）」这样的词去优化。到了意图时代，你得想清楚：在 agent 代表我的生意出场的那些地方，我怎样才能被发现？再到委托时代，我们要想的是：我怎样才能以安全、可信的方式把这份需求转化成订单。

[7:31] Nixon

And so the challenge becomes to think through is like as COVID moves to AI agents, how do you as a business be visible where the consumers interact? The issue with this is that cataloges weren't really built for agents. Um most merchants today maybe advertise on Google. Um they have a catalog they syndicate to there or maybe to Meta or Facebook or other advertising platforms. But the spec wasn't built for agents. And so agents now um you know if you it's not optimized for agents to discover products. Um it's not as readable for agents as well. And they're really just built for humans search. And then to compound that even more, the models that agents use are still for the most part a black box. Agents can use different varying models to make their decisions. Um, different models use different ranking and recommendations and and and uh algorithms to to determine their recommendation. Uh, vector database retrieval is also not always deterministic. And because of all of this, how do you decide as a business what model do you do to you optimize for? In the old SEO world, there was a clear strategy. In this new world, the strategies are still remain largely unclear. And so what do you do when you don't know? You run experiments. And so we we ran some experiments. We enriched uh some of our merchant product data. And we ran experiments to see how agents uh uh searched uh and and recommended products.

所以难题就变成了：当电商转移到 AI agent 上，你作为一家企业，怎样才能在消费者发生交互的地方被看见？

问题在于，商品目录当初压根不是为 agent 建的。今天大多数商家可能在 Google 上投广告，他们有一份目录分发到那里，或者分发到 Meta、Facebook 或别的广告平台。但那套规范（spec）不是为 agent 设计的。所以现在的情况是：它没有针对 agent 发现商品做过优化，对 agent 的可读性也没那么好，实际上只是为人的搜索而建的。

雪上加霜的是，agent 用的模型大体上仍然是个黑盒。agent 可以用各种不同的模型来做决定；不同的模型用不同的排序和推荐算法来决定推荐什么；向量数据库的检索也并不总是确定性的。因为这一切，你作为一家企业怎么决定该针对哪个模型去优化？在旧的 SEO 世界里，有一套清晰的打法；在这个新世界里，打法大体上还不清楚。

那不知道的时候怎么办？做实验。于是我们做了一些实验：我们对一部分商家的商品数据做了富化（enrich，即把商品信息补全、写厚），然后跑实验，看 agent 是怎么搜索和推荐商品的。

[9:18] Nixon

And before I get into the experiments that we ran, um I think very good to level set with everyone how like the differences between keyword and semantic search and it it trickles through to the themes I talked through earlier with keyword search being you know very matching the the search that the user is entering in. So, in a classic search, it might match literal words. Um, for example, I'm searching for blue running shoes, and you only get results that contain those words. The good part about this, it's very precise, it's fast, it's predictable. But the downside is it misses intent. And so, if you ever search footwear for jogging, it may not even return those shoes. Uh, and that's only simply because the words don't match. Whereas with semantic search, it matches based off meaning. Um, what it does is it converts the query into vectors that create uh meanings of intent. And so you might type something, oh, I'm searching for something comfortable for marathon training. And you might find your running shoes even though none of those words exist in the search query. Um so the good part about the that is it broadens and how it can search but the downside is if you inject too much meaning it can lose focus and hallucinate. So there is a balance of how much you want to enrich and how much you want to inject into your enrichment.

在讲我们跑的实验之前，我觉得有必要先和大家对齐一下关键词搜索和语义搜索的区别，它也呼应我前面讲的那几个主题。

关键词搜索，就是去匹配用户输入的搜索词。在经典搜索里，它匹配的是字面上的词。比如我搜「blue running shoes（蓝色跑鞋）」，你只会拿到包含这几个词的结果。好处是非常精确、快、可预期；坏处是它抓不住意图。所以如果你搜的是「footwear for jogging（慢跑穿的鞋）」，它可能连那双鞋都不会返回——仅仅因为词对不上。

而语义搜索是按意思来匹配的。它做的是把查询转换成向量，用向量来表达意图的含义。所以你可能输入「我在找适合马拉松训练、穿着舒服的东西」，就能找到你的跑鞋，哪怕这些词一个都没出现在搜索词里。所以它的好处是把搜索的范围拓宽了；坏处是，如果你注入了太多含义，它会失焦、会产生幻觉（hallucinate）。所以你想富化多少、想往富化里注入多少东西，这里有一个平衡。

[10:46] Nixon

So to sum it up keyword search is quite precise. It's it's very literal. Semantic search is just it's smart but it's very fuzzy. And realistically, the real answer isn't an either or. It's it's combining both, which is exactly what our hybrid approach does. Um, and that's kind of where we're headed next. So, now that we understand what semantic search is, it's important to highlight that there is no singular way to really enrich product data for semantic search. I think it's really important for you as a business to understand the kind kind of catalog data you already have. um do you have a thin catalog? Do you have a catalog of products that are more spec heavy? Depending on where your business is today, you should think through what you might want to enrich. And so for the experiment though, we thought through different patterns of enrichment. Um and I won't go through all of them in detail, but things around like filling attributes, making sure there's description depth, um make sure there's buyer context and trust signals, and even some product identity. I think the main take away from this slide is that enrichment isn't one sizefits-all. The right approach really depends on the catalog you're starting with and that's the framing we'll build on next.

总结一下：关键词搜索相当精确，非常字面；语义搜索聪明，但很模糊。而现实里真正的答案不是二选一，而是把两者结合起来——这正是我们的混合方案（hybrid approach）在做的事，也是我们接下来要讲的方向。

既然理解了语义搜索是什么，有一点必须强调：为语义搜索富化商品数据，并没有唯一的做法。我认为对一家企业来说，真正重要的是先搞清楚你手上已经有的是哪种目录数据。你的目录是不是很单薄（thin catalog）？还是你的商品偏重规格参数（spec heavy）？取决于你的生意今天处在什么位置，再去想你可能要富化什么。

不过就这次实验来说，我们设想了几种不同的富化模式。我不逐个细讲，大致是这些：补全属性（filling attributes）、保证描述有深度（description depth）、加入买家情境（buyer context）和信任信号（trust signals），甚至还有一些商品身份标识（product identity）。这页幻灯片最主要的结论是：富化不是一刀切的。合适的做法真的取决于你起步时手里的目录是什么样，这也是我们接下来要沿用的框架。

[12:03] Nixon

And so here's just an example of enrichment in action for us. Um example on we're starting with we have men's uh men's shoe model 8543 that's blue and then the description will be I have blue running shoes. um available multiple sizes. there's an attribute that's blue keyword matches maybe blue running or shoe and then um the semant for semantic there's you know very little context for that when you enrich there now we might have the title as a men's lightweight running shoe um the description is now more enriched like it's cushion it's breathable uh there are further attributes we've pumped into that as well so there's like true to size it's it's the the material is mesh um we also put in some reviews here as well so enrich that with some review data and now what the search sees is um it matches running, it matches training, it matches meshes. There's a lot more that it matches. So with semantic there's a lot of rich meaning that's put there uh to kind of enhance the search. And so here the the some of the examples we use was attribute filling or bio context or even uh the trust signals around the reviews as well. And so ultimately at the end of the experiment um you know our thesis was generally speaking does enrichment help boost uh recommendations and at its core yes it does. Um what we saw was an unenriched baseline like when it went headto-head um it just completely underperformed. We found that merchants with the thinnest cataloges tend to gain the most. Uh merchants with weak product data also had the most room to improve. Um but I think what we needed to also understand is that it's all about content quality that drives the results and it's not the schema that actually and actually piling on you unstructured data can actually dilute the signal as well. And so what do we mean by like unstructured data? like we just don't want to stuff the the context of anything any words we want. We don't

这里是我们实际做富化的一个例子。起点是这样的：一双男鞋，型号 8543，蓝色；描述是「蓝色跑鞋，多种尺码可选」；有一个属性是「蓝色」。关键词能匹配上的大概就是 blue、running 或 shoe；而对语义搜索来说，这里几乎没什么上下文。

做了富化之后，标题可能变成「男士轻量跑鞋」；描述更丰富了，比如有缓震、透气；我们还往里补了更多属性，比如尺码标准（true to size）、材质是网面（mesh）；我们还放进了一些评价，也就是用评价数据做了富化。于是现在搜索看到的是：它能匹配 running，能匹配 training，能匹配 mesh，能匹配上的东西多了很多。对语义搜索来说，这里放进了大量丰富的含义来增强搜索。这个例子里我们用到的手法就是属性补全、买家情境，以及围绕评价的信任信号。

那么到实验最后——我们的命题大致是：富化到底能不能提升推荐？从根本上说，能。我们看到，没做富化的基线在正面对比（head-to-head）时完全落在下风。我们发现目录最单薄的商家往往收益最大；商品数据薄弱的商家，改进空间也最大。

但我认为还必须理解的一点是：驱动结果的全在内容质量，而不是 schema（数据结构）本身；实际上，堆砌非结构化数据反而会稀释信号。那我们说的「非结构化数据」是什么意思？就是我们不想随便什么词都往上下文里塞。我们不

[14:14] Nixon

want to put any boilerplate messaging around um potential like you know welcome to our store. We're familyowned business. Those things are irrelevant to an agent at times. And so how you enrich what you structure into the data really really matters as well. And so just generally speaking a sum up of our learnings was enrichment helps keyword search. Semantic retrieval requires a very different set of data structures. But there is a balance to how much you want to enrich uh to improve semantic retrieval. And while this was only one set of experiments, we did find that enriching just yeah head-to-head always improved recommendations. Um what we had found in some cases when we overenriched the agent tended to hallucinate more and it underperformed. Um and we don't have a concrete recommendation for every business but a good place to start is just assess your catalog and then figure out the right strategy for that. Um and so just to wrap up you know enrichment works assess your catalog structure and more text isn't better. um structured quality content, structured quality content is the one thing that wins out and AI will reward content quality and not brands. Um but it does optimize for other trust signals as well.

想放那些套话式的文案，比如「欢迎光临本店」「我们是家族经营的企业」之类。这些东西对 agent 来说很多时候是无关的。所以你怎么富化、往数据里结构化地放进什么，同样非常非常重要。

总的来说，我们学到的可以归纳为：富化对关键词搜索有帮助；语义检索需要的是一套很不一样的数据结构；但为了改善语义检索，你要富化到什么程度，是有一个平衡点的。虽然这只是一组实验，但我们确实发现，只要做了富化，在正面对比中推荐效果总是更好的。而在某些情况下，当我们过度富化时，agent 往往幻觉得更多，表现反而变差。我们没法给每一家企业一个具体的建议，但一个不错的起点是：先评估你的目录，再为它找出合适的策略。

最后收个尾：富化是有效的；评估你的目录结构；文字更多并不等于更好。结构化的高质量内容——结构化的高质量内容，才是最后胜出的那一样东西。AI 奖励的是内容质量，而不是品牌。不过它确实也会针对其他信任信号做优化。

[15:41] Nixon

So thank you everyone. Uh if you want to learn more about PayPal's Agent Commerce Services, please visit our booth at P11.

谢谢大家。如果想进一步了解 PayPal 的 Agentic Commerce Services，欢迎到 P11 展位来找我们。