A social layer within an online sales process adds not only the possibility to create reach, as discussed in Chapter 1, but also knowledge and trust. Consider the following case:

This is an example of social commerce. The term social commerce was first coined by Yahoo! in 2005. It describes ecommerce transactions that are triggered or guided by social interactions or supported by social network data. Social commerce usually consists of one or more of the following elements:

Because the social media world is full of different tools, platforms, and functions, the social commerce world also has many names and faces: “T-commerce” stands for commerce via Twitter and “F-commerce” for sales via Facebook. We are sure there are many more terms to come. Social commerce seems to have become the “next big thing,” and the numbers appear to justify this hype. There are over a billion users on Facebook, and more than 170 million active Twitter accounts. Let’s assume we could sell each of them something for only $2—that would result in a revenue of $2.3 billion.

Social commerce is sometimes presented in terms that practically guarantee raining money, if you only implement “F-”, “T-”, or whatever-commerce. The reality, of course, is often different from this, and many merchants have lost money from “being social.” Should we thus refrain from social commerce?

Let’s be guided by lessons from history. For example, in the mid-1990s, people promoted the practice of having a page on the World Wide Web as a sure-fire way to sell products and services. A few years later, the crash of many dot-com startups taught us that the medium itself wasn’t the ticket to sales success. Today, social commerce has had the same glow of youth and promise for boosting sales, ever since the term was first coined. In all likelihood, history will prove these tools to have a similar trajectory to the growth of the Web.

Of course, websites and ecommerce not only work as selling tools but have also created a billion-dollar industry. They are now ubiquitous for anyone in business today, much like having a telephone. However, their adoption took longer than anticipated, and online capabilities are only an additional tool for the sales process. In much the same way, social commerce will develop. It is here to stay and will become a significant part of the online commerce world, but the adoption will take longer than expected. Social commerce is not the one-and-only savior; it is one additional tool to facilitate the sales process.

Using social commerce, we will also find an answer to the often-discussed question of ROI from social media. Sales are the ultimate link to ROI for efforts in social media. As discussed in Chapter 10, often it is hard to link effects from social media, such as an improved audience reach to an ROI. The “R,” or the financial return, is often too undefined, or a causation is too hard to prove (see Chapter 9). This is different with social commerce. One can create a direct link between the investment into a social sales channel and the return in terms of additional margin dollars earned.^[47]

Related to social commerce is the usage of data within the sales approach. But while the data might be taken from social interactions, this data is by no means the only source available to support sales activities. The second part of this chapter will touch on a few standard topics about the way data can be used to improve the sales experience. The underlying aim here is the same. Like with social commerce, we are using data to create one of the two major drivers in a sales process: knowledge and trust. Let’s look at the following cases:

Those two cases are using the data available about you. In both cases, an algorithm combines data taken from your user behavior—meaning what you have bought and done before on the site—with information about your friends (in the Netflix example) or with information about what the general public (in the Amazon example) would do. The data is used to create knowledge about you. We commonly refer to this process as recommendation engine. The second part of this chapter will explain what recommendation engines are and the main challenges in setting them up.

Social confirmation involves people online saying that a given product is good. It creates the trust that was missing so far in online interactions beyond big brand names. Examples of this include two forms of displaying actual purchasing intent.

A step up from social confirmation involves someone from a customer’s own personal network confirming his purchasing intent for products and services. This is the ultimate social confirmation, because it is not just anyone recommending a product, but a friend whom the customer knows. Many sites facilitate this process by displaying pictures of people from an individual’s network who bought something—or, more recently, as in the case of Facebook’s Open Graph Search functionality, by creating the ability to mine data through one’s social connections (for example, a list of books that are liked by journalists).

This aspect of social media is especially helpful if a service or product depends on network externalities. Take the adoption of a messenger service like Whatsapp or Skype. Both services would benefit highly from a tool that tells the user which of her friends are on it. Another example is the decision to purchase a ticket for a concert. If a person already knows that some of his friends are going, he will make sure that he is there as well, so as not to miss out.

This process leverages the network of an individual, and thus people with a bigger network can be used for more social confirmation than others. Are those people influencers? As we discussed in Chapter 1, the concept of an “influencer” is often overrated. The person with the biggest network might be used the most to create a social confirmation. But the network size does not say anything of its ability to create an intention. Intention might be created at a local level by only a few very specific personal contacts. A well-designed social commerce setup will need to learn this kind of network information that goes way beyond any network size description.

Most of us marketeers wish that social commerce would enable us to click a few buttons and suddenly we’ve created purchasing intent with our potential customers. We get this asked all the time: “What do I need to post so that my products sell more?” or “Which tool should I use so that my sales double?” Reality does not work this way, of course. Whether or not social commerce works depends on the customers and the product. We see, however, that some social commerce tools work better then others in most cases.

It also seems that average shoppers prefer structured data, such as ratings and reviews, as social confirmation. There is a demonstrable impact: one study by eccomplished found that the number of people using online ratings and reviews is about eight times higher than the number of people using only social media recommendations (see Figure 2-1). For us, this shows that the ordered format of an aggregated sum of online reviews is more successful in persuading someone to buy something than the relatively unstructured format of different blogs, tweets, and the like.

Figure 2-1. Reviews are more important than social media alone. (Courtesy of eccomplished.)

Unfortunately up to 38% of those reviews are actually fake or paid.^[50] Those fraudulent reviews are an issue for all social networks. As soon as consumers realize they were tricked, they’ll lose trust and the power of social endorsement and social confirmation will vanish. That is the reason why more companies offer what was specifically recommended by your connections. Those kinds of recommendations are so far not that common since there are often not enough reviews within one person’s network to provide an effective customer experience. This will change as social networks become a bigger part of our daily shopping experience. Friends’ recommendations will become some of the most effective ways to create intention. They can be easily transported into the world of social commerce even without the knowledge of social media tools or social graphs. For example, Leskovec and his colleagues described the success of a sales attempt in which the first person to purchase an item recommended through an emailed referral link got a 10% discount, and the recommender received a 10% credit.^[51]

Data such as these indicate that social confirmation and peer pressure do, in fact, influence our purchasing intent. This view is further validated by the large sums of money many large firms spend on mining and delivering such social confirmation. Though we feel that we are independent, and indeed we do not like being told what to do, the opinions of others often influence our own decision-making process.

Reviews and recommendations can influence consumer decisions about what to buy and whether or not to buy something. But how can one manage these in order to have the most positive impact on purchasing intent? For instance:

Those questions can be answered using recommendation engines or systems, the point where big data analytics and social media intersect. Recommendation systems select the product, review, or even site layout that is best for the consumer’s given aim. Compared with pure review sites such as TripAdvisor.com, which does not rank comments from visitors, a recommendation system openly facilitates a purchasing decision by suggesting things a consumer is most likely, based on data, to want.

Even if, on the surface, there are no social features—and thus no comments, no ratings, and no Facebook friends—recommendation engines use a collaborative approach to recommend products that others “like you” have bought. Thus it is a social group feature, as the engine has put similar consumers into a group and then suggested something for them to buy. The work of those engines often occurs within milliseconds, in a way that is invisible to the normal user.

Recommendation systems are perhaps the most critical link between social data and purchasing intent. The next sections explore the function, technology, and development of these systems.

Most of the examples we discussed in this chapter are based on collaborative recommendation systems. Sometimes we refer to them as crowdsourced systems, as they use patterns from other users or user actions. As a very simple example: if Pete and Tom both purchase items 1 and 11, but Tom also purchases 2 and 4, then 2 and 4 are likely to be recommended (rightly or wrongly) to Pete. Tom is also sometimes called the “identical twin,” as his behavior is similar to Pete’s. In reality, those systems will look not only for Tom to recommend products to Pete but to many others as well.

The most famous example of a collaborative recommendation system is Amazon’s listing of what people who bought the same book also purchased, as described previously. Amazon can take advantage of a database of millions of users to drive these recommendations.

Note that this collaborative approach is different from the content-based approach, which we will discuss next.

Each of us has already touched a content-based recommendation engine. At a shopping site, for example, a system might pull up items that are similar to the ones you recently purchased. For example, if you selected a red shoe for checkout, suddenly there might be a lot of other red shoes on your screen. Those shoes might be placed via advertisements on banners or via your online store, directly in the “recommended” section. The algorithm selected the shoes because it realized that you are about to buy a red shoe. As we saw in Chapter 1’s discussion of behavioral targeting, the chance of creating purchasing intent is greater the closer a consumer is to the point of sale. So if you already voiced the purchasing intent to buy a red shoe, why not have the system try to suggest other red shoes? This approach has nothing to do with the tastes of your friends or similar people. It has only to do with the content of the item selected, hence the name content-based recommendation engine.

Data for content-based recommendations often takes one of two forms: either data based on consumers’ shopping, purchasing, or search history or a specific user profile that enables the consumer to define her preferences for easier shopping. For example, US shoe retailer Payless allows consumers to enter their gender and shopping preferences, and keeps track of their past ordering history. This data is used to customize in-store recommendations or mail advertising. It is purely based on the content consumers said they like and not based on the patterns of others.

Content-based systems have one disadvantage. To identify similar products, you need a specific description of any product. When looking at a product (clothes, ladies, evening wear), you can easily suggest an evening dress from the same category. However, these categories need to exist in the first place; if you have a product that falls into fewer categories (clothes, ladies), then the system might choose a pair of gym shorts to display next to the evening dress.

When customers do not have the ability to rate items, and companies do not have an infrastructure for evaluating the purchasing history of similar buyers, businesses are often left with content data as the only criteria for making purchasing recommendations. Such data is often less accurate than collaborative data and requires careful planning of data categories to avoid wildly inconsistent recommendations. At the same time, these content-based systems are often much easier to implement, particularly for applications with a smaller audience and/or limited access to social data.

Collaborative recommendation systems are based on “supervised learning.” In order to make recommendations based on what others have done or purchased, the program needs to analyze similar historical data. Therefore, at the beginning of a program, the computer doesn’t know anything, and the likelihood is high that its recommendations will be wrong. Being wrong is a bad thing, because your early customers could lose trust in your ability to make recommendations. This is why sites such as Amazon had to wait until they were running for a while and had more data to work with before they could start making recommendations.

For example, FOUNDD is a young Berlin-based startup that has built a recommendation engine for the movie industry. Despite its brilliant algorithm, the cold-start issue is a serious problem. Without having a long purchase history, such as Netflix would have, the algorithm will not be able to recommend anything useful. Fully aware of it, the founder created a “hot or flop” page in the beginning. Each customer has to rate 10 movies before the system aims to recommend anything. “It has to be fun,” said Lasse Clausen, one of the founders. “Otherwise the user won’t do it, and we will not get up and running.” In order to bypass the cold-start problem, Lasse and his team are are initially using a content-based approach by pre-clustering similar movies according to their genre or popularity.

Only after 5 to 10 movies are rated can the system start recommending. Are 10 movies enough? Most likely not for a superb recommendation, but the more the users rate the movies, the better the system will become at detecting each individual’s preference.

The combination of pre-clustering and initial user input is the best way to avoid the cold-start issue as demonstrated by FOUNDD. Alternatively, if possible, initial “learning” data can be drawn from external sources and then improved upon over time with a company’s own customer data.

A much bigger issue than the cold-start issue has yet to be faced by FOUNDD, the startup we just discussed. They have focused on a set of movies, but when you have thousands of items in your store, it is difficult to figure out which ones are best to recommend.

This issue is often called data sparsity. It is similar to the cold-start issue but broken down in terms of each product. If a retailer has millions of items, it is hard to find good ways to suggest items from the “long tail,” or the millions of items that are not best sellers but still get sold. To recommend products from this long tail, the algorithm needs to include a minimum number of purchases for each of these products.

According to one of the top five online retailers in Germany, the biggest sales come from a changing online assortment, thus the data for their long tail is quite small despite having the potential for many users and many purchases. As a result, the recommendation engine could only suggest its biggest non-changing sales item, which happens to be socks. Of course, recommending socks with each and every purchase did not drive increased incremental sales. The retailer eventually addressed this issue by having a student write manual recommendations for more appropriate goods instead of using a machine-learning approach.

The more complex the system is, and the more items there are to recommend, the more difficult it is to make a good recommendation due to the potential issue of data sparsity. For example, LinkedIn serves employment advertisers by recommending the right people for their posted job ads. For this functionality, LinkedIn needs a tremendous amount of data; such a service was not feasible even with 10 million users. But now that they have more than 100 million users with highly variable data points, it must be addressed as a big data problem. In the extreme, issues such as tastes in human style or matching specific body shapes to clothing choices can also require millions of matching data points in order for a machine to begin to make an accurate recommendation.

Machines only learn based on popular choices. Thus, they are more likely to recommend the obvious popular choices that are safer to recommend.

Let’s say you have just finished the fourth book of the seven books in the Harry Potter series. What will be a likely recommendation from a machine-learning algorithm? The correct answer: books one through three. But that is not useful, since you most likely have already read them. Similarly, recommending book five would not be helpful, as you probably already know that it exists. A good recommendation would be a book in the same genre or style or maybe based on one of the other attributes of the book that you appreciated the most.

A good bookseller will understand your needs quickly. A recommendation engine will, at least so far, only recommend what others have read. This is the reason Amazon cannot replace the recommendation of a human bookseller—at least not yet.

If you were to scratch the surface of the major recommendation engines, you could envision a set of complex analyses guarded by a top-secret formula. At the top of the online retailing spectrum, there is in fact some truth to this stereotype. But a more important point is that, for most people, recommendation systems revolve around making use of clear and observable data.

Let’s look at the leaders in recommending emotional content, such as Amazon and Netflix. Both have great recommendation systems. Their sites suggest things to their users such as movies to watch or books to read, and the recommendations are very important to sales. At Netflix, 75% of all user choices are based on recommendations made for them by their own system. During the Strata 2012 conference, Xavier Amatriain (@xamat) from Netflix explained some of the inputs that are used by their recommendation system:

Those are many variables, and there are surely more. But more importantly, each of those six areas are influenced by several indicators taken out of millions of data points from 23 million subscribers in 47 countries with over 5 billion recommendations. Mind-blowing? Probably not! If you look further into your own company, you will find similar amounts of data lying around. Most of it is probably not used; or even worse, it might not be easily accessible as it is within Netflix. However, you have a complexity of your own, and the question you must ask yourself is how to use the data.

Let’s have a look at how Netflix has used the data. It did not start out with many different variables; it started out with just a simple method of recommendations. As discussed in Chapter 9, the strategy it used is highly suitable because it is:

The lesson learned from Netflix is to start simple. Only after the correlations were proven did Netflix start to expand the system step by step. It started by optimizing a few metrics where causation was highly predictable and then built from there. Even today, when according to Netflix, 75% of what people view is recommendation-based, it continues to refine features and algorithms using a simple baseline of viewer opinions. Looking at your own process of using recommendations, you could boil the process down into three key guidelines:^[56]

“Hi Lutz, I am Brad your community manager. I wanted to check in with you to see how you like our product.” Brad’s Facebook picture smiles at the top of his email. It is a family photo of him at the beach. It is difficult to know whether Brad is a real person, a robot, or perhaps a human-guided robot, but this company clearly understands that personal relationships are an important part of the sales process, even if—or maybe especially when—someone shops online.

Netflix, as we mentioned earlier, benefits from the ability to tell customers which of their friends watched a movie and uses it as a trustworthy recommendation. However, not every company has products that are so commonly bought by many friends. If you are a bike retailer, you would need to be lucky to find some friends of your customers who just happened to have bought a similar bike and thus could be used as an endorsement to instill purchasing intent.

Therefore, in the future, we will see more and more attempts to create personal online connections between salespeople and customers in order to facilitate sales. The need for personal connection is not only needed for the sales process, but also for retention purposes. We will see in Chapter 4 how customer-care efforts in social media help to create a more personal face for a company and thus satisfy this human need for connection.

Research shows that users are highly sensitive to the reasons something is recommended to them. Providing an explanation seems to trigger an automated response pattern similar to the one observed by Robert Cialdini (see Give Me a Reason). The best recommendations are, of course, often those that can be linked to personal data, but this is not necessarily a requirement. Take Amazon, which analyzes our purchasing habits without knowing that our friends are, say, mountain bikers. Amazon’s explanation of why these book recommendations are being made—because others liked them as well—is still much more powerful than just displaying the recommendations alone.

How can you use social media processes or social media data to support your sales processes? To get you thinking, look at the following workbook questions:

Recommendation engines are an effective way of getting to the fourth “V,” the value. Setting them up is not always easy, but it coud be the small steps that make the difference between success and failure:

Have you found a new way of engaging within your sales channel? Engage with us via @askmeasurelearn or write on our LinkedIn or Facebook page.