The same people. The same products. Reviews posted close together in time. Our research asks how these patterns can help uncover coordinated reviewing.
WORKING PAPER · 2026
Can public traces help detect coordinated fake review campaigns on digital platforms?
Creator of Public Traces Academic affiliation: SMU Cox School of Business
I study how reviewers and sellers interact, and what their behavior can tell us about trust on online platforms. Public Traces is my independent project, turning research with my coauthors into tools for following review patterns and documenting observations locally. Shared contributions and research datasets are planned for a later release.
From a small search to patterns across the archive.
The original study tested how far one investigation could reach. A separate analysis measured recurring review patterns using the Amazon Reviews 2023 archive collected by McAuley Lab. These findings come from the completed research by Gijs Overgoor and coauthors.
A search opening 500 reviewer histories reached 92.2% of the study’s comparison group: reviewers who gave five stars to at least two products during observed review-recruitment campaigns.
What this number means
The search started from one randomly chosen reviewer. The comparison group contained 3,916 of the original study’s 39,524 reviewers. Researchers observed the promoted products and recruitment dates, rather than the identities of people recruited. Matching those products and dates is a behavioral link; it does not establish that each reviewer was recruited or paid. This percentage measures how many members of that group the search reached, not overall detection accuracy.
Patterns across all five-star reviews17.9%
In the 2023 Amazon archive, 17.9% of five-star reviews fell in clusters by the reviewer group identified by the method, with their authors recurring across at least two products.
What this number means
This result covers January 1–September 14, 2023 and uses 16,035,782 five-star reviews as its denominator. The method opens up to 100 histories per session. Reviewers enter the group after appearing in at least 10 different opened neighborhoods within a session. A cluster requires at least five group reviews within 28 days before or after the focal review; the author must recur across at least two such products. The paper calls this its L2 footprint. It is not a percentage of confirmed fake reviews.
Patterns among verified-purchase reviews12.2%
The same pattern appeared in 12.2% of verified-purchase five-star reviews when the full procedure was repeated using only verified purchases.
What this number means
This result covers January 1–September 14, 2023 and uses 13,742,903 verified-purchase five-star reviews as its denominator. The entire procedure was repeated on that restricted data, using the same settings as the all-review analysis. This is the paper’s L2 footprint, not a percentage of confirmed fake reviews.
The percentages describe different study groups and questions. They do not tell us the chance that an individual review is fake.
How we checked the results
For the original study, researchers observed which products were promoted in review-recruitment groups and when. Reviewers whose five-star activity matched at least two of those products and dates formed a comparison group. The study checked both how many of these reviewers the search reached and how many appeared near the top of its ranking.
89 of the top 100 reviewers matched observed campaigns
In the completed analysis, 89 of the 100 highest-ranked reviewers belonged to this comparison group: they gave five stars to at least two products during observed review-recruitment campaigns.
A sudden wave of reviews can have an ordinary explanation. So can people buying the same products. The research looks for these patterns occurring together, repeatedly, and keeps the supporting observations visible.
The results help identify activity worth examining. Missing reviews, incomplete histories, and the search settings affect what can be found. Establishing whether someone was recruited or paid requires further evidence.
McAuley Lab at UC San Diego collected and published this archive. Our reported audit uses reviews dated January 1–September 14, 2023; the archive itself contains reviews from multiple years. The displayed findings are our analysis of those records.
Sherry He · Brett Hollenbeck · Gijs Overgoor · Davide Proserpio · Ali Tosyali
Coauthor Brett Hollenbeck hosts the public dataset in his repository. The release also includes labels and variables from subsequent studies, described in its documentation.
The demos use generated records. Shared community contributions are not open in this release. Explore the data sources →
Use earlier findings to inform a new investigation.
A separate historical reference records reviewers whose activity matched the research’s 2023 patterns. This public release explains that reference but does not offer account-level membership checks.