A Lie Detector Test for Online Reviewers

Researchers develop new ways to detect fake reviews

In February, the owner of a home improvement company in Texas posted an advertisement online for “a writer who can write and post 25 positive reviews” on eight popular websites, including Yelp, Google Places, and Citysearch. A man in Chittagong, Bangladesh, won the gig and agreed to do 200 fake write-ups for $100. Within days, rave reviews for the company started popping up online. The business owner, who spoke on the condition of anonymity, says he doesn’t feel bad about the deception. He says clients call all the time extolling his service, but they don’t post reviews.

The proliferation of fake reviews is a huge problem for e-commerce and recommendation sites that depend on user ratings. “At the end of the day, if consumers don’t trust the content, then there is no value for anyone,” says Vince Sollitto, a spokesman for the local review site Yelp. It’s not just consumers who become suspicious. British regulators are investigating the complaint by a company representing hoteliers that there are millions of phony reviews on the travel site TripAdvisor, which says it’s cooperating. For some products, up to 30 percent of reviews can be fake, says Bing Liu, a computer science professor at the University of Illinois at Chicago.

Several big tech companies are sponsoring academic research on fake reviews. Liu worked with two researchers at Microsoft, which put up $138,000 to study nearly 6 million Amazon.com reviews, among other things. They found many duplicate or near-duplicate reviews—a clear sign of fraud. Within those fake reviews, the team found telltale patterns. Spammers are likelier to post multiple reviews for one product or various products made by a single brand. They write very soon after products launch, and their views often differ sharply from most other users. The fake reviewers aren’t usually longtime members of a site. To expand the research, Liu is working with two engineers from Google, which put up $50,000.

At Cornell University, researchers focused on finding semantic tics unique to fake reviews. Like the Texan, they went online and hired people to write 400 fake reviews of hotels. Then they used a computer model to compare those reviews with 400 real ones. The truthful reviews tended to talk about the actual physical space, using specific nouns like “floor” and adjectives like “small.” Since spammers weren’t familiar with the look of the hotel, they spent more time talking about themselves, the reasons they went on a trip, and their traveling companions. “The word ‘husband’ is very indicative of deception,” says Myle Ott, a PhD candidate and co-author of the study. When the researchers trained a computer to look for the linguistic signs, it detected 90 percent of the fake reviews. Ott says several websites, including TripAdvisor, have inquired about the team’s findings.

For the most part, tech companies are loath to talk about how they detect fraudsters, for fear of inadvertently spreading hints for gaming the system. TripAdvisor says only that it considers 25 factors in determining whether a review is fake. Yelp runs its reviews through an anti-fraud filter, with impressive results; every fake review the Texan bought was flagged by Yelp’s algorithms, though his fraudulent reviews remain up on the seven other sites.

In an effort to fool the filters, some job ads for fake reviewers insist that writers use unique user names and e-mail addresses for each entry so they don’t look like they all come from the same person. Others demand the writers use computer servers based in the U.S., spread out reviews over the course of weeks, and post at different times of day. Some offer up to $80 for reviews from Yelp users that have achieved Elite status, which makes them trusted members on the site. Adam Medros, TripAdvisor’s vice-president for global product, says his company trolls the sites where businesses advertise for fake reviewers. “I can’t think of a recent time where somebody has surprised us with instructions on how to evade our systems,” he says. “We have pretty much seen it all.”

The cat-and-mouse game is likely to continue, since there’s real money involved. The Texan says that while the ratings are nice, what really matters is that the reviews contain keywords that boost his Google ranking. He’s now one of the first sites listed for the search terms relevant to his business. “We’ve only been online for a couple of years,” he says, “yet we have basically conquered all of our competition.”


    The bottom line: As much as 30 percent of online product reviews can be fake, prompting companies to sponsor research to detect and combat fraudsters.

    Before it's here, it's on the Bloomberg Terminal.