We use a robust combination of automated and manual processes to protect the integrity and trustworthiness of our online review community. Key to this is our software that detects fake reviews, backed up by a dedicated Compliance Team.
Trustpilot’s customized software
Staying true to our roots as an innovative tech start-up, when Trustpilot needs a new tool, we usually find a way to customize or create something using the skills of our technical specialists.
In order to detect and remove fake reviews, we have a team dedicated to building and refining unique fraud detection software. We’ve seen promising results so far, but we’re always working to improve it.
Much like Google doesn’t openly promote its algorithms, we can’t reveal exactly how our software works because that would defeat its purpose. Providing all the details would give the small number of people who want to cheat the exact tools they need to game the system.
But what we can do is tell you about what we’re doing and why we believe in it:
The strength of independent systems
As a starting point, all reviews are treated equally by our software. It doesn’t have access to, or factor in, information about whether a company subscribes to Trustpilot’s services or not. In line with our commitment to trustworthiness, we believe that’s how it should be.
The unique software that we’ve built runs independently and tirelessly 24/7 across our platform, examining a large range of different behavioral parameters. The system then uses an algorithm to calculate the likelihood that reviews are fake. This is represented by a ‘fake score’.
If the score for an individual review is very high, our software will automatically move the review offline and send an email message to notify the reviewer. At this stage, none of our employees are directly involved - everything is handled by the software alone.
Why start with behavioral factors to assess reviews?
Usually, untrained readers of reviews focus on the language, rhythm and tone of the writing - linguistic indicators - to determine trustworthiness. In contrast, our software looks at behavioral aspects based on raw data. It’s been suggested that linguistic indicators may be the less accurate of the two methods** because good writers can copy expressions used in genuine reviews in order to deceive readers.
In any event, it’s always necessary to look at a full range of indicators. We never remove a review based on one parameter alone - it’s always multiple factors that lead to taking something offline.
And while the process tends to be extremely accurate, we acknowledge the importance of giving reviewers the opportunity to contact us if they believe there has been a misunderstanding. This happens very rarely, but because we move reviews offline rather than delete any that are deemed fake by our system, we can always reinstate them at a later date.
Supporting manual processes
To really make sure our processes are robust, we also arm our people with technology. On top of Trustpilot’s fully automated software, our specially trained Compliance Agents and Investigators also have access to powerful tools that help them examine review patterns for anomalies. These systems highlight any unusual and persisting patterns using a range of data.
Our team can use these factors to investigate and decide whether exceptional patterns indicate problems or have a logical explanation. This takes the best of both worlds - technology flags patterns that the human eye may not see, and our Compliance Team uses these elements to investigate further.
We take seriously our responsibility to uphold the credibility of reviews and always investigate cases that are brought to our attention as suspicious.
**Mukherjee, A.; Venkataraman, V.; Liu, B.; Glance, N. What Yelp Fake Review Filter Might Be Doing? International AAAI Conference on Web and Social Media, North America, jun. 2013. Available at: http://www.aaai.org/ocs/index.php/ICWSM/ICWSM13/paper/view/6006. Accessed: 18 July 2016.