Chapter 18

Listening-Based Sales Prediction

Social media listening provides us with signals that can be used to sense the near-term future. Our ability to improve foresight contributes to better and more desirable company futures (Johansen 2007). We are learning from diverse studies on product categories like entertainment and consumer electronics, or elections for public office, that social media listening data can accurately predict near-term results, such as changes in sales, or winning candidates. The research chiefs, advertisers, and vendors with whom we spoke welcome that power and see listening-based prediction as an emerging business tool. Networked Insights CEO Dan Neely captured that thought when he told us:

Listening will help create models that can predict changes in markets and consumer trends. Most market research has a predictive element: We study the present and the past to inform what we will do in the future. But, the future will see predictive analytics that can match the vast data set of online conversation against an economic marker or other data set outside of social media—and predict business outcomes. (Neely 2010)

For almost all companies, the ability to anticipate change—even just a few days or weeks out—offers valuable guidance. Predicting the short-term future provides companies with closer-to-real-time data that factors into making marketing, operational, resources, and financial decisions that keep or sharpen competitive edges, and keep the company-customer relationship in sync. This chapter concerns the ways listening data contributes to predicting short-term sales and outlines the emerging plays that are beginning to shape strategy and influence positive business outcomes.

Sales forecasters nowadays possess sophisticated tools for estimation and analysis. Some of these include interviews with key buyers, customer surveys, and market-mix models. Some, the innovators, have recently turned to social media conversations as a source of data to use in straightforward and very sophisticated models to predict future sales within a narrow window of time. Here is a rundown of the types of social media data used to forecast sales:

• Post volumes
• Tweets and status updates
• Search trends
• Online advocacy
• Product reviews

The following pages discuss a variety of sales forecasting approaches using these data sources. You'll notice that most of them were conducted by academics or in company R&D labs on topics such as book sales, movie box office, DVDs, and consumer electronics. The primary reason for choosing these product categories is availability of data. Sources like Amazon provide a wealth of current and historical information on thousands of products, sales rankings, and publicly available consumer-generated content. Though we haven't yet seen research from companies exploring sales prediction yet, we expect that some of them have active research programs but are keeping these close to the vest for competitive reasons. We look forward to the time when their results become available, as their accessibility will stimulate interest in—and knowledge about—cutting-edge uses for social media listening-based sales predictions and their use in strategy.

Post Volumes

The study credited with popularizing listening-based prediction explored the question: Can online buzz predict book sales? Researcher Dan Gruhl and his colleagues at IBM's Almaden Research Center and at Google collected and analyzed about 500,000 sales-rank values for more than 2,000 books on Amazon over a four-month period (Gruhl et al. 2005). They then correlated the changes in sales rank for each book with online postings from a variety of relevant sources, and reached the following conclusions:

• Carefully hand-crafted search queries produced volumes of postings that predicted sales ranks—queries that, in many cases, could be automatically generated.
• Online discussion can successfully predict spikes in sales rank for individual titles.

This type of work, correlating post volumes to sales changes, remains popular today. Researchers have added refinements over the years, such as weighting posts according to factors like audience size, source authority (such as through Technorati rank), influence of the post's author, or sentiment. In the competitive marketplace of contact lens solutions, for example, the leading brand was forecast to lose share of category sales because its influence-adjusted share of the posts was declining over time (Onalytica 2009).

Tweets and Status Updates

Twitter's popularity, openness, and roots in broad public communication make it fertile research ground. Now that status updates can be posted on Twitter from popular social networks like Facebook and LinkedIn, tweets are becoming more reflective of the social networking population—and more valuable as a research source. For that reason, more and more market researchers have turned to Twitter to glean insights about what's on people's minds. October 2010 saw nearly 2.5 billion tweets posted on Twitter; that's about 80 million per day, or about 1,000 per second. And people don't just post; they search for what people are saying, 1,200 times per second (BuzzingUp 2010). Most of those posts will not be relevant to researching a specific category or product. But even if only 1/1000th of 1 percent of October's tweets are relevant, that's 25,000 comments and opinions—a substantial number, and one far larger than most surveys or focus groups would ever generate in a single month.

These stats portray active, engaged, chatty social networkers, and beg the question: Can Twitter activity be analyzed in a way that allows sales to be forecasted? Two researchers from HP Labs sought to answer this question for movie box office receipts (Asur and Huberman 2010). Movies generate research interest for several reasons: People avidly discuss them, both online and off; they have known release dates, which, therefore, permit before-and-after comparisons; and trade sources publish their gross revenues. Importantly, several services exist that forecast movie receipts, enabling comparisons of social-media-derived forecasts to competitors and to actual figures.

Asur and Huberman studied 24 movies over a three-month period, capturing and analyzing nearly 3 million tweets from 1.3 million people. Their analysis looked at the “tweet rate”—the number of tweets per hour—concerning a particular movie both before and during the first two weeks after its release. They studied the ratio of positive to negative sentiment (polarity) to determine its impact on predicting second-week sales. Briefly, their findings were:

• Tweet rate is the best predictor of box office sales in weeks one and two.
• Prior to release, tweets contain a higher proportion of links referencing the movie's publicity (trailers, photos, news, or blog posts, for example), but retweets are a minority. However, these do not predict the relative performance of the movie's box office.
• Following release, tweets become more subjective, as people write about their opinions and experiences. The ratio of positive to negative tweets made some improvements in forecasting sales, but “they were not as important as the rate of tweets themselves.”
• Tweet rate predicted box office results better than the Hollywood Stock Exchange (HSX), a prediction market that's long been considered the gold standard. HSX is a virtual exchange where traders buy and sell shares in movies. The movie stock prices can predict box office results: lower prices, lower box office; higher prices, higher box office.
• Tweet rate not only predicted opening weekend numbers, but accurately forecasted the revenue for all movies in distribution, some of which were two months old.

Overall, the authors concluded that: “The results have shown that the buzz from social media can be accurate indicators of future outcomes. The power of social media is illustrated by the fact that a simple linear regression model considering only the rate of tweets on movies can perform better than artificial money markets.”

The tweet rate method is applicable beyond movies, to just about any product or service actively discussed in online conversations, especially for those, like entertainment, cars, consumer electronics, or sports, that have an official launch, release, or event date.

Search Trends

Many companies plan marketing and investment activities according to news about business conditions, such as monthly retail, auto, and home sales, reported by the government or other organizations. However, timeliness and accuracy are two inherent problems with the monthly releases, as the government issues reports about one to two weeks after the month ends. For example, June reports come out in mid-July. Even then, data may not be final, as it may be revised up to two times subsequently. Improving data by overcoming these issues can equip businesses and policymakers with more accurate, timelier information about the current business environment, and can, therefore, improve the quality of their decisions.

Google economists took up the challenge to “predict the present.” While we know that search leads sales, these researchers sought to uncover whether search volumes on a particular category or brand would be helpful in predicting that category's or brand's monthly sales. To test that, Google economists Hal Varian and Hyunyoung Choi built two forecasting models to predict sales one month ahead. One model included Google Trends data; the other did not. Details on the models are available in Varian and Choi (2009). The models themselves are very straightforward; and, for those so inclined, the authors provide the software code to create them. Google Trends, available through Google, provides daily and weekly reports on query volumes related to most industries and individual brands. It is described more fully in the Appendix.

Varian and Choi's research analyzed areas like automobile sales and those of individual brands, home sales, and travel. After running its models many times, Google concluded that those that include relevant Google Trends data “tend to outperform” those that do not. In other words, Google Trends data leads to more accurate predictions of month-end sales. Ideally, Google's researchers state, prediction could be improved further by developing models that identify “turning points” in categories and brand sales within the month. That would give managers the ability to respond to real-time shifts in their market (Varian and Choi 2009).

Online Advocacy

Advocates, those passionate people who champion companies or products, contribute substantially to sales and market share. In the HDTV market, Samsung's market share leadership over Sony was correlated with advocacy. Although more people are familiar with the Sony name, Samsung outscored it in terms of conversational volume and favorability during the key selling season. Why? Samsung had a greater number of advocates than Sony or other competitors (Nail 2009).

In order to be predictive, we need to model the relationship between advocacy and sales. One step in that direction comes from full-service listening vendor MotiveQuest, which created an online measure similar to the Net Promoter Score, called the Online Promoter Score. Net Promoter Score is calculated by using answers to a single question: How likely is it that you would recommend Company X to a friend or colleague? Researchers compute the score by subtracting the percentage of detractors from the percentage of promoters. The Net Promoter Score links to company revenue growth rates (Reicheld 2006). For its computation, the Online Promoter Score, developed in conjunction with leading academics, listens to and analyzes social media conversations and sentiment (MotiveQuest 2009).

MotiveQuest's work with the MINI automobile utilized the Online Promoter Score to gauge the effectiveness of the brand's communications programs in generating sales. Following a successful U.S. introduction, MINI did not have a new product for its second year—a serious hurdle in a marketplace where sales momentum and new-car sales are driven by introductions, relaunches, and updates. MINI's agency, BSSP, built and implemented a progressive three-pronged communications program based on listening insights from MotiveQuest research (see Chapter 9 for a full description of this case). As the campaign rolled out, MotiveQuest tracked the relationship between the Online Promoter Score and sales from January 2006 through April 2007. Its analysis revealed that changes in the Online Promoter Score predicted changes in sales about one month in advance. The direction and size of the sales changes were just as important: Sales increased or decreased by 53 percent of the change in the Online Promoter Score. If the score went up by, for instance, 10 percent, then sales would increase by 5.3 percent. While the percentage change in sales will likely differ depending on the industry and brand, one result that emerged from this study is clear: The value of advocacy on sales is measurable and can be used to develop and evaluate marketing and advertising.

Product Reviews

Product reviews feature prominently as a source of online conversation. Internet reviews influence purchase, and retailers such as Best Buy and sites like Bizrate.com encourage their visitors to post appraisals and evaluate products. Amazon is especially aggressive in this area, ceaselessly developing new ways for its community members to post, read, and engage with one another. The wealth of comments and ratings attracts researchers who want to learn if, and how, user-generated content influences sales.

The first wave of research focused on the ratings and volume of reviews, and it solidly supports the notion that reviews influence product sales, box office receipts, and TV show ratings (Chevalier and Mayzlin 2006, Hu et al. 2008). However, those early studies did not capture the dynamics of evaluation, reading, and acting upon reviews that take place within a virtual conversation; conversations in which people pay attention to the qualities of both the review and the person writing it. Readers are listening for clues, and this second wave of research is looking for those same clues and gauging their impact. Understanding which features of reviews influence customers gives companies additional insight for making decisions such as which reviews to promote, which product features to emphasize, or which keywords to buy for a search engine marketing program. The emerging picture is one where customers weigh good and bad news and act in accordance with the judgment; people are not responding in knee-jerk fashion to high ratings.

An analysis of the related literature uncovered the following factors as those that influence sales:

• “Helpful” reviews: Helpful ratings may be a stronger motivator to purchase than the overall “star” rating for a product. People judge reviews as helpful in two ways: One is that review readers actually mark them as helpful, which is feedback that the retail site then counts and displays. The second method is to use readers’ own evaluations of the review's opinion (subjective) and factual (objective) statements.

  Reviews containing primarily objective comments are generally considered to be more helpful, while highly opinionated evaluations are usually less so—and have a downward impact on sales. However, the type of product makes a difference as to which kind of statements readers find most helpful. Readers value objective comments for feature-based products like electronics or washers, whereas more subjective statements are appreciated for experience-based products like DVDs. That makes sense, as the former relates to performance and the latter to how people felt about a movie (Chen et al. 2008, Ghose and Ipeirotis 2010). As more research is done in this area, marketers and advertisers will look forward to learning about the helpfulness of the mix of statements for products or services that combine experience and features in widely discussed categories such as autos, travel, and consumer electronics.

• Reviewer characteristics: There are differences between the top, or power, reviewers on sites and those who author individual posts. Two studies suggest that reviews by the top 1,000 reviewers are not important to people's purchase decisions. While it may be ego-gratifying to be designated as a “top reviewer,” their impact on sales is not retailer-gratifying. Review readers value expertise, readability, and the right mix of subjective and objective content. The reviews that the community rates as helpful are the ones influencing sales, period. Amazon, for example, “spotlights” the two most helpful reviews for a product, and these “super-reviews” impact sales more strongly than the average of all the other helpful reviews. Those reviews serve as shortcuts, enabling people to reduce the time they spend searching and evaluating (Chen et al. 2008, Hu et al. 2008, Ghose and Ipeirotis 2010).
• Product coverage: Reviews of newly introduced products or those that did not garner many reviews have greater impact on sales than those for widely covered products. For example, the power of an individual review for a Harry Potter book is less than a review for a less popular title.
• Sentiment: The accepted wisdom is that unfavorable reviews depress sales and positive reviews increase sales, but the relationship is actually more nuanced than this. In fact, there are times when negative reviews spur sales, as well as instances in which both negative and positive sentiments have no impact.

  Two factors appear to positively influence sales even when the reviews are less than favorable: product awareness and perceptions of review quality. Jonah Berger and his colleagues investigated awareness and found that while positive reviews increased book sales for all authors they studied, negative reviews affected authors differently. Specifically, the more critical assessments hurt the sales of well-known authors, but increased sales for lesser-known authors by raising their profiles and interest in their work. Berger theorizes that negative word of mouth may work in a similar way for lesser-known products or services (Berger et al. 2009).

  Elements that indicate a quality review—one that is informative, fair, balanced, well-written, and spell-checked—are also associated with increased sales, even when the review itself is negative. Reviews that lay out the strengths and weaknesses provide information that readers use to evaluate the product from their perspective and interests. For example, a reviewer who claims that a particular product is difficult to hold in her small hands might not be a showstopper for someone with normal-sized hands. If criticisms or objections are not as important to the reader, increased sales can result (Ghose and Ipeirotis 2010).

  There have been several studies of the effects of “buzz” on box office sales. Some of these indicate that the volume of postings or tweets about a movie impacts sales more powerfully than sentiment does. In other words, the mere fact that people are discussing a movie is often more important than the opinions expressed during these discussions. The reason may be, as noted earlier, that people are reading or discussing them for their helpfulness, balancing the pros and cons in light of their own specific interests and preferences.

• Product features: There's a new method that uses text mining on product reviews to extract the features discussed. It then estimates their importance to consumers and the opinion around them, and models their contribution to sales. For example, by studying camcorders and digital camera sales on Amazon, Archok and colleagues (2010) determined that discussion over features had the greatest impact on sales—far greater than factors like product price, product age, trends, seasonal effects, and the volume and sentiment of reviews. What's especially interesting is the ability to include subjective features like “ease of use” and “design” into the models; these terms capture intangible qualities that influence decisions but have been hard to explain in a quantitative way. By including product features from the posts, researchers were able to build accurate predictive models of near-term future changes in sales.
• Passage of time: Reviews lose their influence over time; the early ones have the most substantial impact on sales. This is presumably because people have additional information available to them as time passes (Hu et al. 2008). Jonah Berger found that although the impact of negative reviews on lesser-known authors diminished over time, this was not the case for established authors. The reason: People may be more willing to buy afterward because the negative review raised awareness of the author—which correlates with purchase—and their memory of the review's disparaging tone had faded. In other words, when a famous author comes out with a clunker, people write it off and look forward to the next one; but when a book by an emerging or promising author comes out and gets a negative review, buyers think more about the author than they do the specifics of what reviewers did not like.

Emerging Plays to Predict Changes in Sales

We identified four emerging plays that aid in predicting changes in sales:

• Match the sales prediction approach to the type of product or service marketed, as well your organization's capabilities. Sales prediction methods will continue to be tested and improved, and new ones will appear as the discipline evolves. We have already learned that, one, all social media sources are not alike; two, types of products or services, such as functional or hedonic products, generate a variety of discussions and are evaluated differently; and, three, that accurate predictions can use simple trends, like the tweet rate, or combine very sophisticated text mining and econometric modeling. Increase your brand's confidence in listening-based sales predictions by triangulating; utilize and compare several listening-based approaches and/or combine them with traditional sales predictions.

  Companies of all sizes can conduct sales forecasting; several of the methods described in this chapter can be adapted by even the smallest business. Predictive approaches grounded in analyzing volumes are applicable and, importantly, relatively easy to do well. For example, every company should take advantage of programs like Google Trends or Google Insights for Search, free tools that provide reports and downloadable data on current and historical trends in search terms.

  By importing the search data into a spreadsheet and adding a column for sales, users can discern lead-and-lag relationships among search volumes and sales—no special expertise or additional investment required. To add a little more insight, integrate search volumes for competitors, to learn how competitive searching affects your search volumes and sales. You can also perform similar analyses when people tweet, blog, or otherwise discuss your category, brand, product, or service in public places.

• Leverage sales predictions and develop strategy linked to sales drivers. Sales prediction models reflect people's behavior, the way they interpret discussions, and how they act. Use those listening-based insights to give clear, specific directions for tailoring or adjusting marketing and advertising strategy to grow sales. Consider some of the strategic and tactical guidance available from our review:
  • Detect and respond to changes in demand in near-real time.
  • Identify and promote the most helpful reviews.
  • Emphasize the features people value most highly.
  • Match emotional and rational messaging to reflect the product or service type.
  • Select “hot button” terms for search engine marketing keyword purchases.
  • Reach out to the right people for direct engagement.

  Designing strategy and implementing tactics that are grounded in factors that motivate sales promises to bring about greater coherence in marketing and advertising. You'll want to predict and adapt continuously as goals, programs, peoples, and resources become more closely aligned.

• Challenge the conventional wisdom around ratings, influence, and sentiment. People do not blindly respond to high ratings, influential reviewers, or positive buzz. Research shows that they instead take a more balanced approach to conversation content; they weigh the pros and cons, and take the writer or speaker into account. Even negative reviews do not always lead to negative outcomes, and can actually stimulate sales under certain circumstances. For these reasons, question the received wisdom about influencers, ratings, and sentiment, and determine whether, and how, it applies to marketing and advertising your products and services.

  For example, Amazon's elite top 1,000 reviewers have shown to be unimportant to book sales, whereas a reviewer that the community had deemed “helpful” can inspire sales. Companies that are tempted to base an influencer social media strategy on such lists may find their programs underperforming. The key seems to be locating those individuals the community regards as influential, not choosing reviewers based on their volume of activity.

• Share the predictions with colleagues and management. Predicting sales changes from listening data transforms conversation into valuable management metrics and points the way to the near-term future that colleagues and senior management readily grasp. The sales prediction is a common currency. As valuable as consumer insights are for the marketing and advertising functions, their power is often unrecognized or underappreciated by the C suite. However, executives and higher-ups will see the connection of listening-based sales predictions to the bottom line right away. Communicate predictions throughout the company, to open possibilities for more shared understanding of customers, greater collaboration, better strategy and tactical programs, and more concerted action for building business.

Summary

Social media listening data is being used to accurately predict sales in the short term, providing businesses with signals about their current environment. These forecasts allow companies to take action by responding to expected changes in demand. Initially based on buzz volume, prediction models have matured to capture the content and characteristics of online discussion that drive sales. Marketers and advertisers can leverage those qualities to create and implement strategies and tactics that resonate with customers and link directly to revenue growth.

Because customers and prospects are not automatons, you must carefully consider the pros and cons in discussions, consider the writer's or speaker's perspective, and act according to their judgment. Conventional notions of the impact of influence, sentiment, and product ratings do not apply across the board, and companies need to reflect upon the applicability of these ideas to their specific situations. Last, listening-based sales predictions provide measures that nearly every function appreciates and understands. Sharing them across your organization can bring the value of listening to the entire enterprise.