Dead Internet Theory: A 2GIS Review Fraud Investigation

The "Dead Internet Theory" suggests that much of what we see online — reviews, comments, ratings — is not written by real people at all, but generated by bots or paid for by those who profit from inflated reputations. I decided to check whether this is true for 2GIS, one of Russia's most popular mapping and review services, and what I found was unsettling.

How the Investigation Started

I built an open antifraud tool called af2gis.ru to analyze review networks and user behavior patterns on 2GIS. The tool processes user account age distributions, cross-company review correlations, and statistical anomalies in review timing and rating distributions. For the analysis I collected data on approximately 2 million users and 60,000 companies in Novosibirsk alone.

Pattern 1: Accounts Created Just to Leave One Review

The most common pattern involves accounts registered on the same day or within days of each other, each leaving a single glowing five-star review on the same business. A dental clinic in Yekaterinburg showed 97% of its reviews coming from newly-registered accounts with suspiciously uniform positive ratings. These accounts had no other activity on the platform whatsoever.

This is the digital equivalent of hiring a crowd of strangers to stand outside your restaurant applauding. The accounts exist solely to move a rating needle.

Pattern 2: Cross-Promotion Between Related Businesses

Some business owners apparently coordinate. A cluster of seemingly independent businesses — a barbershop, a nearby cafe, a car wash — share a pool of the same reviewer accounts. The same ten people leave positive reviews for all of them in rotation. This mutual back-scratching inflates every business in the ring simultaneously.

Pattern 3: Geographic Impossibilities

Some accounts reviewed venues in Moscow and Vladivostok on the same day, or left reviews for locations in cities thousands of kilometers apart within hours of each other. Either these are the most well-traveled people in Russia, or the accounts are not being operated by real users visiting real places.

The Tbank Verified Review Anomaly

2GIS allows reviews to be marked as "verified" via Tbank (formerly Tinkoff Bank) payment confirmation — a signal that the reviewer actually paid for something at the establishment. This sounds like a fraud-prevention mechanism. In practice, it became the most interesting data point in the analysis.

Across the dataset, Tbank-verified reviews comprised between 1% and 10% of total reviews for most businesses in any given city. Statistically, 96% of companies that received Tbank-verified reviews saw their average rating increase as a result. That alone is suspicious — genuine payment-verified reviews should reflect the honest experience of paying customers, who are at least as likely to be disappointed as pleased.

But the real red flag: certain specific outlets showed Tbank-verified reviews making up 80–97% of all their reviews, while every competitor in the same category showed near zero. A single hot dog stand received 93% of its reviews through Tbank verification. Its direct competitors on the same street: essentially none.

This concentration is not explainable by organic customer behavior. It points strongly to a scheme in which Tbank-verified reviews are being manufactured — possibly by running actual micro-transactions through affiliated accounts to obtain the "verified" badge.

Banking Institutions: Synchronized Nationwide

Some of the most striking patterns appeared in reviews for major banking institutions. Branches of the same bank across dozens of cities received synchronized waves of positive reviews — accounts posting on the same days, with similar phrasing, across hundreds of locations simultaneously. The probability of this occurring organically approaches zero.

The Tool: af2gis.ru

The antifraud analyzer I built is open source. It produces visual network graphs of reviewer-to-business relationships, highlights statistical outliers, and flags accounts whose behavior matches the patterns described above. The methodology deliberately avoids AI or machine learning — everything is based on statistical analysis of observable behavioral patterns, which makes the results easier to explain and audit.

Key signals the tool looks for:

• Account age at time of review (new accounts reviewing old businesses)
• Cross-business reviewer overlap graphs
• Review velocity clustering (many reviews on the same day)
• Geographic impossibilities in review history
• Anomalous concentration of verified-payment reviews
• Rating distribution outliers versus city-wide norms

What This Means

The Dead Internet Theory in its extreme form — that most online content is bots — is probably overstated. But in the specific domain of local business reviews on 2GIS, the data suggests that a meaningful fraction of what looks like authentic user feedback is manufactured. Businesses are paying for it, platforms are (at minimum) failing to stop it, and consumers are making decisions based on ratings that have been deliberately corrupted.

The Tbank verification mechanism, intended as a trust signal, appears to have been turned into just another tool for manipulation — arguably more dangerous than unverified fake reviews precisely because it carries an official badge of authenticity.

Whether 2GIS takes action on this data remains to be seen. The tool is available for anyone to use at af2gis.ru.