Fake Reviews Aren't a Content Problem. They're a Shape Problem.

Based on the research ofSherry He, Brett Hollenbeck, Gijs Overgoor, Davide Proserpio & Ali Tosyali, "Detecting Fake Review Buyers Using Network Structure: Direct Evidence from Amazon," arXiv, 2024

In 2024, Amazon proactively blocked more than 275 million suspected fake reviews before a single shopper saw them, and over the prior two years it pursued legal action against 115 review brokers. That scale did not come from a smarter reader. A 2024 study by Sherry He, Brett Hollenbeck, Gijs Overgoor, Davide Proserpio, and Ali Tosyali, built on Amazon's own transaction data, found that the strongest signal of a fake review has almost nothing to do with what it says. It is where the product sits in the network of who reviewed it, and who else those reviewers reviewed. Products that buy fake reviews cluster tightly together, a pattern the researchers detected from structure alone, without ever labeling a single review real or fake.

For most of the past two decades, platforms policed reviews the way a copy editor would: flag suspicious phrasing, watch for repeated wording, discount brand-new accounts. Fraud brokers adapted faster than the copy editors did. Generative writing tools now produce reviews that clear every content filter a platform can build, varied in tone and length, indistinguishable sentence by sentence from an honest customer's post. What does not change so easily is the trail a purchased review leaves in the transaction graph: the same handful of accounts, a burst of five-star ratings inside a 48-hour window, one seller buying from one broker again and again.

Call this graph governance: policing a marketplace by the shape the relationships among buyers, sellers, and listings form, rather than by judging any single transaction on its own content (see the exhibit). Three marketplaces, two decades apart, show how it works in practice.

eBay Proved the Model Before Anyone Trusted It

In 2007, Carnegie Mellon researchers Shashank Pandit, Duen Horng Chau, Samuel Wang, and Christos Faloutsos built NetProbe, a system that modeled eBay buyers and sellers as a network and used belief propagation, a method for updating each account's fraud likelihood from its neighbors' likelihoods, to flag probable fraud rings. On a real dataset of nearly 700,000 transactions among more than 66,000 eBay users, NetProbe surfaced hidden fraud networks in about six minutes, without reading a single auction listing or feedback comment. The idea worked. It sat mostly unused for another decade because most platforms still built trust systems around content filters and reputation scores, not relationship maps.

Fraud brokers can fake the words, but they cannot fake the company those words keep in the transaction graph.

Amazon Scaled It to a Platform With Billions of Listings

The 2024 study shows why that finally changed. He, Hollenbeck, Overgoor, Proserpio, and Tosyali built a bipartite graph connecting every product to every reviewer who rated it, then applied unsupervised clustering, meaning the model needed no pre-labeled examples of fraud to find it. Products that bought fake reviews showed up as dense, isolated clusters, tightly linked to each other and to a small set of repeat reviewers, unlike the sparse, non-overlapping pattern that organic reviews produce. That structural signature held even when individual review text looked entirely legitimate, one reason Amazon could block 275 million suspected fake reviews in 2024 alone while still reporting that more than 99% of its reviews came from genuine buyers. The same read now anchors enforcement in court: Amazon won a judgment against the operators of more than 75 fraudulent websites that sold fake reviews and seller accounts, cases built on coordinated patterns rather than on parsing any one post.

Alibaba Turned the Same Logic Into Standing Infrastructure

Alibaba runs graph governance as daily operations rather than a periodic audit. Its risk systems scan tens of millions of listings a day, and the company has used them to remove roughly 380 million fake or counterfeit listings and shut down about 180,000 seller accounts, work documented in Alibaba's own published fraud research. It got there not by hiring more content reviewers but by treating every transaction as one edge in a graph large enough to make coordinated fraud rings visible as outliers, the same principle NetProbe proved on a dataset one ten thousandth the size. The pattern is not Amazon's alone: Trustpilot removed 3.3 million fake reviews in 2023 out of 54 million submitted, and Tripadvisor blocked a record 2 million fake reviews the same year, roughly four in five of them before they ever appeared.

The machines now do most of this catching without a human in the loop (see the exhibit): Trustpilot's automated share of removals climbed from 82% in 2023 to 90% in 2025, and Tripadvisor now blocks the large majority of fakes before they post.

You might object that graph governance demands data and engineering talent most marketplaces, especially smaller ones, do not have. That is a fair worry, but the Amazon paper's central finding cuts against it: the clustering method needed no labeled fraud data, only the transaction graph every platform already logs for each order. A marketplace with a fraction of Amazon's scale can start with the same coarse signal: flag any burst of five-star reviews from accounts with no other purchase history on the platform. Yelp shows how far the coarse version travels: its automated recommendation software, which weighs reliability and reviewer-activity signals rather than prose, flagged 17% of all reviews as not recommended in 2025. Regulators are catching up too. The FTC's rule banning fake reviews, effective since October 21, 2024, allows penalties of up to $51,744 per violation for platforms and brokers that knowingly traffic in them. Structural evidence changes what counts as a violation: a broker ring that shows up as one dense cluster is not one case to build, it is grounds for hundreds.

The lesson for any platform governing trust at scale is not to read faster or filter harder. It is to look at the shape transactions make before reading a word of what anyone wrote. Graph governance will not replace content moderation, but it catches precisely what content moderation is built to miss: fraud that reads exactly like the truth.