The epidemic of fake reviews and the ingredients needed to stop it

Among the many conveniences online shopping offers consumers today, such as shopping directly from home on any device, or scheduling repeated delivery, the availability of crowdsourced feedback from other customers might be one of the biggest reasons people make purchases on a particular ecommerce store aside from products themselves. 

Veteran ecommerce merchants are well aware that ratings and review averages are the most used proxies of quality for consumers. In today’s economy where any customer has multiple mediums to express their opinions on purchases, what customers are saying about a product can make or break a business.

Though because reviews hold so much sway over potential customers and their decisions to purchase, review listings are fertile grounds for abuse. 

In this post, we’ll cover some common ways review abuse is conducted, how some companies have approached stopping review abuse, and propose ways behavioral data acquisition can be leveraged as an effective means to ensure only legitimate feedback makes its way into ecommerce stores.

Product Reviews: Critical for Sales

Despite the many conveniences of shopping online, there are still risks a customer faces when making a purchase through an ecommerce store. They might buy a faulty product. They might buy a product that arrives not as advertised or promised. 

There’s no way of knowing how the product performs or looks in person when looking at its listing, and they might not even receive the product after making a full payment, then being forced to go through a frustrating refund process. 

Online reviews and ratings give potential customers detailed information on various aspects of the products they are evaluating, such as quality, usefulness, whether the physical product matches its online descriptions, etc., and thus serve as a key way consumers de-risk their own buying processes.

In many cases, and especially for new products or new businesses, reviews can make or break not only the velocity at which sales occur, but also whether the business makes any sales at all. 

In an economy where there are countless options for any product choice, and virtually every product can be sized up with a one to five start rating, initial reviews if negative can mean the death spell for a product’s future success.

Review Abuse: A Dirty Shortcut to Customer “Trust”

Most merchants are well aware of how reviews can influence buying behavior, and unfortunately many merchants succumb to engaging in review abuse, where fake reviews are used to boost a product’s reputation or disparage a competitor’s product artificially.

There are a few ways this happens; fake reviews can be authored by individuals or services that offer reviews for a fee, and can be written manually or posted using some degree of automation or bot scripting.

With large scale review abuse, first, hundreds of accounts may be created on platforms for the sole purpose of posting bogus reviews, and many times this is combined with legitimate inactive account stealing for the same purpose. 

Then, scripts can be used to generate various positive or negative reviews mentioning the product in bulk, which are then scheduled to be posted at specific times using the accounts under control.

On a smaller scale, individuals who engage in review abuse might have multiple accounts at their disposal, but author their reviews manually, varying points about the product they emphasize across reviews they post, without every actually buying or using the product.

Impacts and Countermeasures

The frequency of review abuse would surprise most consumers.

A study by Marketing Land found 61% of electronics reviews on Amazon were deemed fake. The percentages of fake reviews were similarly majorities in beauty products (63% fake), sneakers (59% fake), and supplements (64% fake). 

And in 2019, 99.6% of over two million unverified reviews on Amazon were reviews with five stars. Asked to confirm, Amazon had said that over half of all reviews on the site are fake.

In an investigation conducted by CBC in 2021, a single fake review network across North America was found to be comprised of 208 fake accounts which posted 3,574 fake reviews on 1,279 business sites. Many accounts were found to have full names and photos, but on closer inspection used identities of deceased historical figures.

According to a 2021 post, Google themselves stated they blocked or removed over 56M policy violating reviews and nearly three million fake business profiles in 2020 alone. 

Both Google and Amazon have made considerable efforts to develop systems which curb the amount of review abuse on their platforms. Google uses a combination of algorithms looking at automation and policy violation across accounts, traffic from known click farms, and various data types on user profiles to weed out fake reviews (ie. through tracking user locations or purchase history, Google can weed out reviews on places or products the user has never been to or never purchased). 

There is also a human layer in Google’s review system, where employees comb through certain reviews manually and make judgment calls on any reviews Google’s systems mark as indeterminate. 

Amazon in 2019 published a research paper on one of their methods to detect review abuse by framing abuse as a tensor decomposition problem, and in the paper outline their models for detection. For ecommerce companies that do not have the wealth of internet behavior data on users Google has, methods like these used by Amazon, when combined with behavioral data acquisition, are practical approaches companies can take to develop models for detecting review abuse in all its forms, automated or manual.

Filtering Out Abuse with Behavioral Data

Even without knowing a specific customer’s entire internet history and behavior, there is still a wealth of fine-grained, behavioral data that can be collected on users and devices in something as simple as a text input field, i.e a “leave a review” box.

If ecommerce platforms were to collect the right kinds of behavioral data like text input speeds, mouse and touchpoint interactions, data familiarity, and application fluency, they could use those data types to build user & device fingerprints to curb both automated and manual review abuse. 

We’ve covered how similar account takeover and creation scams are used in automated ticket scalping, and many of the same practices in gathering behavioral data on users to curb those cases can be applied in this case, where large-scale review services build their initial list of accounts before authoring fake reviews.

If a bot were somehow able to bypass detection in the account creation or takeover phase, behavioral data gathered during the posting of the review could still flag the abuse by measuring use of keyboard shortcuts (i.e cut and paste), or time average measurement between keystrokes (vs. legitimate users taking the time to write out their thoughts, often at erratic keypress speeds), as one example. 

Similar methods could be applied to detect individuals authoring fake reviews, looking for inputs from memory (each legitimate review should be assumed to be “new” to the user) or usage of keyboard shortcuts (abusers might have generic templates they cut, paste and alter slightly across products). 

Once an individual review is classified as fake or abusive, models can be developed further to assign reputational scores to specific devices or accounts connected across devices, effectively shutting down potential abuse by blacklisting the author behind the device completely, regardless of their device or account used.

These are but a few examples where using fine-grained, context-rich behavioral data can help e-commerce companies re-establish some modicum of trust from would-be customers, in an online economy where perceived reputation means everything to a business looking to succeed.