Lucas Vandenberg, chief executive of Los Angeles-based social media agency Fifty & Five, and his staff sift through thousands of tweets each day, eavesdropping on any mention of their clients' brands or products.
The agency operates Twitter accounts for many of those brands. Some of them may ask Fifty & Five to respond to complaints it comes upon, or to locate negative tweets about competitors and then present the client's brand as a better alternative.
But wading through all that online conversation can be challenging for a small firm like Fifty & Five. So it is turning to an emerging industry of companies that help businesses such as Vandenberg's churn through thousands of Facebook and Twitter posts to better understand what customers really think. Some rely on linguistic algorithms, while others try to blend in a bit of psychology and more.
Fifty & Five, for instance, uses natural-language processing software to sort tweets. The presence of words such as "good" and "love," along with a brand name, might throw a mention into the positive bucket, and "hate" or "bad" into the negative. Vandenberg favors a service from San Francisco-based public relations software firm Meltwater, although other tech firms, including SAP and Salesforce, offer similar technology.
But purely linguistic algorithms can't always understand context, Vandenberg said. The system he uses is trained to look for the word "virus" because one of his clients sells anti-virus software, but the results can also include tweets in which people mention infections that have nothing to do with the computer sort.
Natural-language processors can also be confused by slang or sarcasm -- if a user calls a product "sick" or "bad," they could actually mean "good," he said, explaining that he and his staff can manually tweak Meltwater's algorithm to recognize terms such as "sick" as positives.
"As powerful as the software is, you still have to have that human touch," he said.
To compensate for stumbling blocks in language processing, a newer psycho-linguistic approach is emerging -- pairing basic linguistic algorithms with psychological principles.
David Sackin, analytics director at ad agency BBDO, uses software called Decooda, which categorizes tweets as positive, negative, or neutral but backs the classification with its own psychological research -- for instance, what kinds of tweets, Facebook posts, or blogs do people create when they are frustrated or disappointed?
For example, when reflecting on a life-changing product they like, users commonly use the word "hate" to express their frustrations with rival products.
"In a post, what you'll find is it's disproportionately negative -- the most common word is hate -- but it's a positive post" for the alternative product, Decooda chief executive David Johnson said.
So Atlanta-based Decooda's algorithm gives added weight to both the frequency of certain words as well as their likelihood, according to psychological research, to indicate positive or negative brand sentiment.
Sackin said BBDO considers such sentiment as one element of a larger picture of consumer behavior. The firm uses it in conjunction with many other software platforms, such as measuring Facebook shares, ad views, or clicks.
"We find [brands] crave the psychological framing" when deciding which campaigns were most effective, Sackin said.
IBM, meanwhile, has been developing a psycho-linguistic product analyzing individual Twitter users' personalities. Using "the big five personality traits" -- a common psychological paradigm scoring a subject's openness, conscientiousness, extroversion, agreeableness and neuroticism -- the software creates a portrait based on a user's tweet history, noting characteristics such as positive or negative outlooks, self-consciousness, or if they're "open to experience," among other characteristics.
IBM is betting that brands will use this information to tailor their customer service to the personality of individual customers.
For instance, if a generally positive Twitter user expresses extreme anger toward a brand, that complaint could be more important to the brand than one from someone who complains about everything, said IBM research scientist Michelle Zhou.
IBM is basing its model largely on text submitted to its psychology team by hundreds of thousands of volunteer subjects.
"Knowing a person's emotional stability or [tendency] toward negative emotion, the business entity has a better handle on how to interact with this person," Zhou said. "If this person has a high level of self-consciousness or a high level of vulnerability, you may want to start [a response] with 'don't worry about it,' to give a sense of security to this person."