ReviewScope: AI-Powered Review Integrity Analysis

Inspiration

Online reviews shape billions of dollars in purchasing decisions, yet up to $42\%$ of reviews on some platforms are estimated to be fake or manipulated. We were inspired by the challenge of detecting coordinated reviewing behavior—not to label reviews as "fake," but to help consumers and platforms identify patterns that warrant further investigation.

The breakthrough insight was treating review analysis as a multi-stage reasoning problem. Rather than using simple sentiment analysis or keyword matching, we designed a pipeline where each stage informs the next, mimicking how a human expert would analyze suspicious review patterns.

What We Learned

Multi-language analysis is crucial: Fake review operations often translate the same message across languages; our analyzer detects these cross-language patterns.
Timing patterns reveal coordination: Bursts of similar reviews posted within hours of each other are a key signal.
Probabilistic language matters: We learned to present findings as "patterns that may indicate" rather than accusations—this is both more accurate and more ethical.
AI orchestration is powerful: Breaking complex analysis into specialized stages produces far better results than a single monolithic prompt.

How We Built It

Architecture

$$ \text{User Input} \xrightarrow[\text{Language Detect}]{\text{Preprocessing}} \text{Pattern Discovery} \xrightarrow[\text{Gemini Call 1}]{\text{Analysis}} \text{Coordination} \xrightarrow[\text{Gemini Call 2}]{\text{Synthesis}} \text{Integrity Assessment} \rightarrow \text{Dashboard} $$

Multi-Stage AI Pipeline

Stage 0 - Preprocessing: Normalize text, detect language ($\text{en} / \text{de} / \text{fr} / \text{es}$), extract metadata.
Stage 1 - Pattern Discovery: Gemini 3 analyzes linguistic similarity, emotional exaggeration, and template phrases.
Stage 2 - Coordination Analysis: Gemini 3 examines timing patterns, rating-text mismatches, and suspicious bursts.
Stage 3 - Integrity Assessment: Gemini 3 synthesizes findings into Low/Medium/High manipulation bands.

Key Features

Real-time progress visualization during analysis.
Interactive results dashboard with tabs for Overview, Patterns, Clusters, and Timeline.
Review cluster visualization showing potentially coordinated groups.
Timeline density chart revealing suspicious posting bursts.
IP-based rate limiting for abuse prevention.
Past analyses history with full result reconstruction.

Challenges We Faced

JSON parsing from AI responses: Gemini sometimes returns markdown-wrapped JSON; we built robust extraction logic.
Cross-language detection: Simple word frequency wasn't enough, so we expanded to comprehensive word lists.
Non-accusatory framing: Designing prompts that identify patterns without making accusations required careful iteration.
Visualization of complex results: Making AI findings accessible through intuitive charts and cards.