Decoding Social Media Data
- by Wendy Moe
Yes, social media has changed how we interact with our friends and with brands. We post pictures of our experiences to share. We rave or rant about a recent service experience. No one argues that social media has not changed the landscape of our social lives. But the question that companies are trying to answer is: How should social media change our business practices.
Some organizations treat social media strictly as a communications vehicle. They use social media as another outlet to push out messages or as a customer service hotline. But social media has the potential to do much more. With all of the data that social media generates, other companies are turning to social media as a source of customer intelligence, treating posts as valuable insights into the voice of the customer.
This sounds great, right? Social media will be the solution to all of our business challenges, helping us measure customer satisfaction, respond to complaints and design better products based on customer opinions. So what’s the problem?
First, there’s more data than many of us know what to do with. In my opinion, even the leading social media monitoring platforms are struggling with this. Second, often what social media metrics tell us don’t align with our offline intelligence. In a recent study published in the Journal of Marketing Research, a colleague and I showed that popular metrics of social media sentiment were not at all correlated with traditional offline brand tracking surveys. For a brand manager who has used tracking surveys for years, this lack of correlation casts serious doubt on the validity of social media as a source of intelligence.
Watch Prof. Wendy Moe describe a new social media measure of brand health that strongly correlates with offline brand tracking surveys, during a presentation at the Marketing Science Institute.
So what’s causing the discrepancies between offline surveys and online social media?
• Opinions expressed don’t necessarily reflect majority opinion. There are significant biases in the opinions that people choose to share on social media. People with extreme opinions are more vocal than people with more moderate opinions. People with negative opinions also try to dominate the conversation. And what do the majority in the middle do, those individuals who make up the bulk of our customer base? They tend to stay silent. This selection effect is one that survey design experts go to great pains to control, but we do little to control for these biases in our social media metrics.
• Opinions systematically differ across venues. Opinions on microblogs like Twitter tend to be more positive. Opinions posted to discussion forums tend to be more negative and exhibit a negative trend. Blogs provide more moderate or mixed opinions. However, few metrics explicitly acknowledge these differences. Instead, we average across venues, or we choose to monitor or give more weight to opinions expressed on one venue over another. Either way, we end up with metrics that don’t necessarily reflect the majority opinion of our customers.
What’s the solution? There are three:
1. Minimize the bias. When we encourage a larger variety of opinions, even if some of those opinions are negative, the dynamic that tends toward bias on online opinions is minimized. Plus, the impact of any one negative opinion is reduced as other voices balance it out. Rather than inviting only your best customers to post an opinion online, encourage everyone to post. It cultivates a more vibrant conversation around your brand.
2. Control for the bias. We will never be able to eliminate some bias, but we can account for it in our metrics. We can do this by examining how expressed sentiment varies depending on the product attribute being discussed or on the venue to which the opinion is posted. Understanding the sources of variance helps separate bias from true underlying opinion. Models that can explicitly measure the effects of these biases on expressed opinion will then be able to provide an unbiased measure of underlying customer opinion.
3. Cast a wide net. We can account for venue bias by monitoring a wider variety of venues and explicitly acknowledging the differences across venues. When we monitor only one venue, our metrics will be biased depending on what types of opinions and dynamics that venue attracts. But if we monitor a wider variety of venues, we can see the variation in opinion across venues and try and isolate average opinion, independent of the venue.
In the recent study published in the Journal of Marketing Research, my co-author and I did exactly that. We measured the effects of each of the biasing factors (including venue effects) on expressed opinion to separate bias from underlying opinion. The resulting measure of underlying opinion was highly correlated with offline brand tracking surveys. In other words, tremendous potential exists in social media data to serve as a source of intelligence, but only if we analyze the data carefully.