The Coming Disruption to E-Commerce Search
When I was a kid I loved going to bookstores. Actually, no.
I loved books, but I hated bookstores. In the 1980s, buying a book in Shanghai, China was like buying a piece of jewelry. Don’t get me wrong, books were cheap. But you were not allowed to take books from bookshelves by yourself. Instead, you had to point at a book and ask the pokerfaced salesperson behind the counter to get it for you. You would have to decide at the counter if you would buy it after flipping a few pages. If you didn’t want it, you would have to give it back to the salesperson before you could ask for another book. If this went on for more than 2 or 3 books, my comrade salesperson would become impatient, and to this day I still remember his annoyed face behind his thick framed glasses.
Only after I came to the US in 1996 did I start enjoying my book shopping experience. I spent many weekends in the bookstore on 3rd Street Promenade in Santa Monica, which was just steps from the beach. I would sometimes get as many as 10 books, find a cozy place to sit down, order myself a cup of coffee, and spend as much time as I need to decide if I want to buy any of them. Life was beautiful.
But of course, Amazon killed it. All the bookstores I loved are gone now. In return, Amazon offered a new experience. It allows me to search for any book that ever existed. I can do it in my bed, on my phone, at work, or on the beach. It tells me what other readers around the world are saying about the book. Once I make the purchase, it uploads the book to my Kindle in less than a second. Many times when someone was still in the middle of recommending me a book, I already started reading it.
E-Commerce experience hasn’t changed for years
Shopping is not just about buying. It’s an experience.
Even for something as simple as books, I have just described the authoritarian-jewelry-buying experience, the bookstore-as-a-coffee-house experience, and the even-a-dog-can-online-order-a-book experience. A good experience is not just an effective one. It also needs to be enjoyable.
The current e-commerce experience can be summarized as search, click, and ship. This experience hasn’t changed for so long that it will put the e-tailers in danger — the same type of danger the antiquated taxi industry put itself into before Uber came along. For sure, the Uber of e-commerce is around the corner.
So what is wrong with the current search-click-ship experience?
First, the search-click-ship experience is sometimes broken. A customer communicates with the e-commerce search engine mainly through keyword queries. But too often the search engine fails to understand a customer’s intent.
Second, even if search-click-ship works as designed, this experience is insufficient to fulfill a broad range of customer needs that I will describe in more detail. For a majority of these needs, the current experience is neither effective nor enjoyable.
A broad range of customer needs
Let us consider the following 4 types of customers and their particular needs.
- Customers who have an exact product in mind, for example, a customer who wants to buy the newly released iPhone X 64gb gold.
- Customers who have a category of products in mind, for example, a customer who wants to buy a rice cooker or a refrigerator. However, the customer does not know which particular product has the right features he needs. Moreover, it is also possible that after some painstaking research, he may realize, instead of a rice cooker, what he really needs is a steamer.
- Customers who have a symptom or a situation but do not know what products can help, for example, a customer who is dealing with insomnia or heartburn, who is concerned with how to get rid of a raccoon in his garden, or who is panicking at the eleventh hour about a Valentine’s day gift.
- Customers visit malls even if they do not have anything in mind to buy. They believe in serendipity. They want to be informed, inspired, enlightened and entertained, but generally they just want to spend some time and have fun. How would e-commerce serve this need?
Current e-commerce experience
E-commerce platforms provide a search-centric experience: A customer describes his shopping intent by a keyword query, and the search engine interprets the query and returns relevant products to the customer.
For customers who know the exact product they want to buy, keyword search does a reasonable job. The problem is, customers do not always know the exact product they want to buy. More likely what they have in mind is a type or a category of products, for example, rice cooker or red wine.
In this case, a customer may choose to conduct some research outside the e-commerce platform to narrow down to an exact product. For example, for rice cooker, through Google, the customer may land on a web page that recommends (in a very authoritative tone) Panasonic SR-DF101 5-Cup Fuzzy Logic Rice Cooker. The customer then comes back to the e-tailer and places an order.
The customer may also choose to search or explore rice cookers on the e-commerce platform directly. E-commerce search engines do not necessarily understand the intent of a query, but with the help of a search log, which connects head queries such as rice cooker or red wine to historical purchases, the customer will likely get a list of popular rice cookers or red wines. One problem is that sometimes the results look like a random set of products to the customer. Thus, the search engine may need to present their results in a meaningful and authoritative way, as customers want to understand why a particular product is recommended.
But anything beyond described above could be problematic. E-commerce search engines haven’t invested much in understanding the semantics of a query. For queries not in the search log, the algorithms simply fall back to keyword matching. In terms of effectiveness, it often looks like the search engine tries to guess a customer’s intent. Then, the customer, after being presented unsatisfactory results, tries to guess how to modify his keyword query so that the search results become more relevant. The e-commerce experience degenerates into a frustrating game of back and forth guessing.
For example, consider “red wine $30.” Now a tail query, most e-commerce search engines fall back on the simplest keyword search method: Any item whose description contains the 3 tokens is a match (e.g., a pair of jeans of color wine red and size 32x30 is a match). As you can see in the screenshots below, Amazon, eBay, and Walmart produce irrelevant results to this query.
The figure below shows the result of similar queries on Taobao (China’s biggest e-commerce website) and Google. Taobao and Google do a better job. In particular, Google combines product search and web search. The top web search result 50 Best Wines under $50 not only shows Google understands the query well, but also the value of information on the web to e-commerce platforms.
Google’s superiority in understanding user intent is no accident. In fact, Google poses an imminent threat to e-tailers such as Amazon. There is a trend for customers to shop on Google, and then follow Google’s leads to specific retailers. In a recent blog titled Help shoppers take action, wherever and however they choose to shop, Google announced an initiative to reinforce this trend.
What is the difference between Google and e-tailers when it comes to product search? Instead of treating “red wine $30” as three meaningless tokens, Google’s algorithm knows its intent is to find red wines whose price is around $30. But a more fundamental difference is their scope. Google operates at web scale and handles open domain queries. E-commerce platforms thought they can satisfy customer needs by conducting search inside the product catalog data. The knowledge Google has accumulated by operating in a much larger scope enables Google to better understand a wide variety of user queries and better handle a broad range of customer needs. Unless e-tailers improve their algorithm, expand their scope, and gain more knowledge about products, customers and the world, they will not be equipped to tackle the threat from Google.
Visual search, as an alternative to keyword search, is on the rise. Visual search is effective in two ways. First, a user provides a picture and gets products that look similar to the picture. Second, when presented with a list of pictures, a user flags those that are close to what he is looking for. Based on the visual feedback from the user, the search engine narrows down exactly what the user has in mind. Visual search is great for products (e.g., soft-lines, furnitures, etc) that are much harder to describe in words than in a picture (“I know it when I see it”). But in terms of applicability, visual search is probably less general than keyword search, as for most products customers care much more than how they look like.
Knowledge: Researching before Buying
Customers often need more information than currently available on e-commerce platforms before they make a purchase. For queries such as rice cooker, e-commerce search engines return relevant products, but web articles such as 2018’s Top 5 Rice Cookers and Rice Cooker Buying Guide are extremely useful to the customer. For queries such as insomnia, heartburn, or how to get rid of a raccoon, e-commerce search engines do not even provide relevant results, but again there is plenty of information on the web that may educate customers what products may help.
According to various studies, 85% of customers conduct research before making a purchase online, and the most frequently used channel of research is web search. E-tailers risk losing a large chunk of the 85% of customers as customers are not able to conduct research on their platform.
Information on the web is, of course, unstructured text. The burden of doing research is thus on the customers. For queries such as rice cooker and red wine, Google suggests products but does not explain their suggestion, and the customers need to dig deep into them. For queries such as insomnia, heartburn, or how to get rid of a raccoon, Google does not suggest products in any explicit or direct way, so the customers need to read documents and follow links in order to find solutions. Overall, researching on Google to find relevant products is a tedious and time-consuming process.
Is there a way for e-commerce platforms to automatically acquire, organize, and present knowledge related to a customer’s shopping needs to improve his shopping experience?
To fundamentally improve the shopping experience, we need extensive knowledge ranging from public knowledge about the world to personalized information about a user. Where do we start? Here is an example. I will describe one particular type of information that may help handle queries such as insomnia that current e-commerce platforms do not support well.
On a highly abstract level, the knowledge base we need to construct for this particular type of queries consists of triples in the form of
(key phrase, relationship, {objects})
where key phrase is an ngram, objects are either products, product types, or other key phrases. Essentially, the knowledge base is a graph of key phrases and their related products.
For example, the knowledge base may contain
(heartburn, medicine-for, {antacids, h2 receptor blockers, proton pump inhibitors})
which lists three medicines for heartburn, and
(2017 sci-fi movies, top-10-of, {okja, blade runner 2049, Thor: Ragnarok, Marjorie Prime, …})
which list top 10 sci-fi movies in 2017, and
(depression, treatment-of, {stay connected, exercise, eat healthy diet, get sunlight})
which lists possible treatment for depression. Note that here the list does not contain products directly, but other key phrases in the knowledge base.
With a knowledge base like this, we may suggest a list of products for a query, and also provide explanation for our suggestion. Of course, we still need to work on query understanding. For example, assume raccoons control is a key phrase in the knowledge base. We need to map queries such as how to get rid of raccoons and raccoons removal from backyard to the key phrase raccoons control so that we can take full advantage of the content in the knowledge base.
Finally, how do we automatically construct such a knowledge base? It takes three steps: 1) Identifying candidate key phrases; 2) Identifying candidate products or product types; and 3) Mining text corpora for relationships between key phrases and products as well as relationships between two key phrases. There are lots of details, and in particular, Step #1 is worth some extra attention. Since the total number of phrases is infinite, how do we identify candidate key phrases? We may consider two approaches: i) Start with some highly valuable domains, for example, the medical domain, where a key phrase is a symptom or a condition. Quite a few websites provide comprehensive lists of symptoms and conditions (e.g., http://www.webmd.com or http://www.draxe.com), and some of them, such as dandruff and food allergy are very common queries on a product search engine. ii) Identify open-domain key phrases. First, we consider popular verb phrases, such as lose weight, fall asleep, quit smoking, as key phrases. Second, We consider noun phrases that appear in patterns such as how to {deal with, prevent, treat, get rid of, stop} [noun phrase]. More generally, we can first come up with a set of verbs of interest such as {learn, work, invest, improve, …} and then use how to [verb] [noun phrase] to find key phrases. In order to find key phrases using the above mentioned patterns, we may need the search log of a general purpose search engine, and if that is not available, we may rely on a web corpus.
Communication: Conversational AI
A great online shopping experience depends on effective two-way communication between the customer and the e-commerce system. The goal is for the customer to easily express their intent, and for the e-commerce system to effectively seek clarification, offer comparison, make recommendations, etc.
As of now, the communication mechanisms for e-commerce are indirect, inconsistent, and sometimes confusing. As I mentioned above, very often the search engine has to guess customers’ intent. The customers, after seeing results that are not satisfactory, have to guess how to revise their queries so that the search engine may come back with better results.
Much effort has been made to improve the situation. Assume you search for red wines on Amazon. Amazon understands the query. Besides showing you a few bottles of wines, it also provides, on the left panel, a long list of options to help refine your query.
The list of options is a way of communication. It helps but it’s not perfect. First, queries such as red wine $30 get options unrelated to wines. This can be fixed when query understanding gets better. Second, working through a long list of options is not necessarily a good experience. In particular, different products have different options, which feels like a different experience every time. Third, on mobile devices, it might not be an option to communicate via a long list of options.
We need something more drastic to fundamentally improve communication: After two decades, it is time to consider moving away from the keyword search experience for e-commerce.
The trend is already underway. On the web search front, Google is offering better support for queries in the form of natural language questions. In 2017, people asked ‘how’ questions on Google more than ever before. It is predicted that 50 percent of search will be voice search by 2020. With the rise of Alexa, Siri, and Google assistant, it is obvious that online shopping will no longer be dominated by keyword search in the future.
An intelligent shopping assistant that communicates with customers in natural language (voice or text) will be the future of e-commerce. It will not only eliminate all the guesswork that plagues e-commerce search now, but also improve customer experience to a new level. On the other hand, it is not trivial to build a general purpose intelligent shopping assistant that can handle a broad range of products. In the following, I consider a two-step approach in this direction.
Slot Filling
For each product type such as red wine and rice cooker, we may automatically define a semantic frame. A semantic frame contains a number of slots. For example, the red wine semantic frame contains slots such as wine vintage, grape variety, etc., which are nothing other than those options Amazon provides for red wine (this is also why we can automatically define semantic frames for a large variety of products).
Instead of showing a long list of options and asking customers to make choices, a shopping assistant or a chatbot will have a conversation with the customer in order to fill slots in the semantic frame. The chatbot may lead the conversation and start with the most important and relevant slots (the importance and relevance can be mined from search log data). As more and more slots are being filled, customers’ intent becomes clearer, and search results become more relevant.
A lot of work has been done in this area, particularly in the field of SDS (speech dialog systems). With the recent advances in deep learning, researchers and industrial practitioners are hoping for breakthroughs in conversational AI. In particular, end-to-end, goal-oriented dialog systems have become a hot research topic, and many of these systems are designed with e-commerce as a major application (see a mini survey on conversational AI — still a draft).
Intelligent Shopping Assistant
In the slot-filling approach, for each type of products, we pre-define a set of slots (e.g., vintage, grape variety, price, etc., for wines). These slots are modeled after the features of the product, which come from the product catalog. But a customer may ask questions beyond features. For example, “how does rice cooker A compare with rice cooker B?”, “how many people does a 3-cup rice cooker serve?”, “Do I need a rice cooker or a multi-purpose cooker?”, etc. These questions are not answerable by information in the product catalog.
Moreover, slot filling is for each product type. An intelligent shopping assistant is for a particular customer need, which is likely to involve multiple product types. For example, a customer who is thinking of buying a rice cooker may end up buying a steamer and a customer who initially has a scooter in mind may end up buying a bike. Thus, in order for the intelligent assistant to help customers decide what suits their needs the best, it must have knowledge across at least a related set of products.
Clearly, in order to build an intelligent shopping assistant, we need knowledge. A lot of such knowledge is on the web, but in unstructured form. For example, we need to know what are the alternative products for a given product; We need to know the top products in each product category (we cannot afford to understand every product); We need to understand what are customers’ preferences for each type of products, and this knowledge may be acquired by automatically processing various buying guides on the web; We may want to understand what a particular feature entails, for example, how many people does a 3-cup rice cooker serve?
Before we start building a knowledge base, we need to decide how such knowledge will be represented, and how the representation can generalize across different categories of products. This enables us to leverage transfer learning to scale up the development of intelligent shopping assistants for a variety of customer needs.
If we come this far, we may realize we are not merely building an intelligent assistant for shopping. Shopping is a particular need of a human being. A properly designed shopping assistant, empowered by knowledge, may already have the mechanism and capability to take care of other needs of a human being. With a different set of knowledge represented in similar ways, we may create an intelligent assistant to enhance the overall well-being of a person.
Serendipity
All the bookstores that I loved are gone. But whenever I pass by a new one, I will step in and take a look. Whenever I fly, I spend time in airport bookstores before boarding. I believe in surprises and serendipity.
As mentioned before, customers visit malls even if they do not have anything in mind to buy. There is no reason why they cannot do the same online. E-tailers should endeavor to inform, inspire, enlighten and entertain its customers.
E-commerce giants such as Amazon and eBay sell hundreds of millions of products, but in reality, only a tiny percentage of these products ever come out through search to the customer, and the huge inventory remains mostly a burden of the search infra. Meanwhile, an average user spends very little time on Amazon, Walmart, or eBay.
Consider Wish, an e-commerce platform that allows small businesses and manufacturers to sell goods directly to consumers. A user interacts with Wish mainly through browsing instead of searching. He may scroll endlessly on his screen to see as many novel and outlandish items as he likes, and not surprisingly, many are addicted and browsing Wish becomes their daily routine. Another interesting detail: Items on Wish are presented with pictures only (no captions), which fuels users’ curiosity to click through in order to find out what they are for.
A user-behavior-based recommendation algorithm is responsible for selecting and ranking millions of products for a customer. This is nothing new, as personalization and recommendation techniques have been part of search all along. But on traditional e-commerce platforms, personalization and recommendation are supplementary: They suggest a small number of items based on past purchases or searches. My family of 3 make over 200 purchases a year on Amazon, but Amazon’s recommendations are rarely useful. Basically, past purchases of a customer are not enough for Amazon to understand the customer well enough to anticipate his future needs.
Wish, on the other hand, made e-commerce browse-centric instead of search-centric. With this simple change, it turns an inventory of hundreds of millions of products into a source of surprises and serendipity, which is pleasant and addictive for the customer and very lucrative for the e-tailer.
Content
While I enjoy pleasant surprises and moments of serendipity in brick-and-mortar bookstores browsing rows and rows of books on display, most of my book purchases are triggered by other day-to-day events. For example, a fireplace chat with a friend working on his 5th startup led to the purchase of The Hard Thing About Hard Things, a post on Facebook about a New York Times book review led to When Breath Becomes Air, and a TED talk on introverts by Susan Cain convinced me to order Quiet.
Is it possible that these real life events can also happen online with an e-tailer?
The above screenshot is from Toutiao, a news and information content platform. By analyzing users, content, and users’ interaction with content, Toutiao generates a tailored content feed for each user. The company is China’s largest mobile platform of content creation, aggregation and distribution. As of June 2017, it has 120 million daily active users, and an average user spends 76 minutes on the site every day, resulting in 1.3 billion articles read.
E-commerce sites have as many users, but an average user spends less than 10 minutes when making a purchase. Note that the comparison is 10 minutes occasionally vs. 76 minutes daily.
We know well-organized content lead to purchases. For example, approximately 55% of Pinterest users in the United States use the platform for finding products. What if e-commerce sites also provide content — content that consists of general purpose articles and videos, not just content about specific products? In other words, what if we take the user base of a giant e-commerce website to bootstrap a content platform, and eventually create a mapping from content to products? We need to develop algorithms to analyze three types of entities {users, content, products} and their interplay, for example, how users interact with content, how content associates with products, how products satisfy user needs, etc.
In doing so, e-commerce should not think of its users as customers who already have intent to make purchases, but as people who are wandering in a mall, waiting to be informed, inspired, enlightened, and entertained by certain content. The goal is to increase the time spent of an average user on e-commerce sites. The conversion from time-spent to purchases will happen naturally, and algorithms that intelligently map content to products (e.g., through entity linking or other more advanced methods) will certainly facilitate and precipitate the conversion.
A more interesting and challenging question remains: Where does the content come from? Toutiao and other content platforms had the same challenge. There are a few ways to create content. First, users may contribute content of their own or share links to external content. There are many valuable customer reviews on e-commerce platforms. Second, to encourage content creation, we may introduce mechanisms so that readers can reward content creators directly (e.g., through micropayment). Another common practice of many content platforms is to hire professional writers to provide content on important topics. Third, new content can be created automatically by aggregating, summarizing, synthesizing existing content. For example, algorithms can summarize multiple articles on the same topic (say, top rice cookers) into one article.
Content creation could be sensitive and even controversial, especially on an e-commerce platform. It is vital for the platform to be a trusted and authoritative resource.
Conclusion
Jeff Bezos built Amazon around things he knew would be stable over time. Here is what Bezos said more than 20 years ago.
I very frequently get the question: “What’s going to change in the next 10 years?” And that is a very interesting question; it’s a very common one. I almost never get the question: “What’s not going to change in the next 10 years?”
And I submit to you that that second question is actually the more important of the two — because you can build a business strategy around the things that are stable in time. In our retail business, we know that customers want low prices, and I know that’s going to be true 10 years from now. They want fast delivery; they want vast selection.
It’s impossible to imagine a future 10 years from now where a customer comes up and says, “Jeff, I love Amazon; I just wish the prices were a little higher.” “I love Amazon; I just wish you’d deliver a little more slowly.” Impossible.
Indeed, the desire of the customer for low prices, fast delivery, and vast selection hasn’t changed.
But the customer also has another desire that probably won’t ever change either: an effective and enjoyable shopping experience. The customer wants to be informed, inspired, enlightened and entertained. It’s time e-tailers build a business strategy around that.
