We got burned by a fake review scam in 2023. Bought a “highly rated” kitchen appliance that broke after two weeks. The 4.8-star rating was built on fake reviews. That pissed us off enough to build something about it.<\/p>\n\n
Fakespot and ReviewMeta exist. They’re decent. But they have problems: usage limits (can’t analyze more than 10 products per day on free tiers), no transparency (black box algorithms), and they miss sophisticated fakes.<\/p>\n \n
We wanted something unlimited, transparent, and more aggressive at catching manipulation. If that meant more false positives, fine. Better to warn about a good product than miss a scam.<\/p>\n \n
Also, we’re developers. Building stuff is what we do. This seemed like a good problem to solve.<\/p>\n\n
We chose Laravel (PHP framework) because we know it well. Fast development, good ecosystem, easy deployment.<\/p>\n \n
PostgreSQL for data storage. We need to store review data, analysis results, and cache responses. Postgres handles JSON well, which is useful for storing review metadata.<\/p>\n \n
Python with scikit-learn for machine learning. We run linguistic analysis on review text. Python’s NLP libraries are better than PHP’s.<\/p>\n \n
OpenAI API for advanced language analysis. Their models are good at detecting AI-generated text and analyzing sentiment. Costs money, but worth it for accuracy.<\/p>\n \n
The entire stack runs on a single VPS (for now). If we need to scale, we’ll add queue workers and load balancing. But 500 analyses per day fits on one server easily.<\/p>\n\n
Amazon doesn’t have an official API for reviews. We have to scrape their website. That’s technically allowed (public data) but practically difficult (they have bot detection).<\/p>\n \n
We use rotating proxies and randomized request timing. We also cache aggressively. Once we’ve analyzed a product, we cache the result for 24 hours. No need to scrape again if someone checks the same product twice.<\/p>\n \n
Scraping breaks occasionally when Amazon changes their HTML structure. We’ve rebuilt the scraper 3 times in 6 months. It’s maintenance overhead, but manageable.<\/p>\n\n
When you paste an Amazon URL, here’s what happens:<\/p>\n \n
Step 1: We extract the ASIN (Amazon product ID) from the URL. Step 2: We scrape all reviews (up to 5,000, which covers 99% of products). Step 3: We run timing analysis (check for spikes and patterns). Step 4: We run language analysis (detect AI text and generic praise). Step 5: We sample reviewer profiles (check account age and history). Step 6: We calculate statistical anomalies (compare to our database of 40,000+ products). Step 7: We combine all signals into a weighted score. Step 8: We convert the score to a letter grade (A through F).<\/p>\n \n
The entire process takes 5-15 seconds depending on how many reviews the product has.<\/p>\n\n
Transparency matters. If we’re telling you a product has fake reviews, you should be able to see how we reached that conclusion.<\/p>\n \n
That is why our result pages and methodology articles explain the signals behind each grade, including timing patterns, language markers, and reviewer behavior.<\/p>\n \n
User feedback also matters. Reports about questionable grades help us refine the analysis and investigate edge cases.<\/p>\n \n
Plus, we believe tools like this should be public goods. Everyone deserves access to review analysis, not just people who can afford subscriptions.<\/p>\n\n
We’ve manually verified 1,000 products to test our accuracy. Results: 87% accuracy on obvious fakes, 72% on subtle manipulation, 5% false positive rate.<\/p>\n \n
That false positive rate bothers us. 1 in 20 legitimate products gets flagged as suspicious. We could reduce it by being less aggressive, but we’d miss more fakes.<\/p>\n \n
We chose consumer protection over precision. If you skip a good product because we flagged it, you lose a purchase. If you buy a scam because we missed it, you lose money and trust.<\/p>\n \n
The trade-off favors caution.<\/p>\n\n
Fake review operations are more sophisticated than we expected. They use aged accounts, varied language, and timing strategies that bypass simple detection.<\/p>\n \n
AI-generated reviews are the new frontier. ChatGPT makes it trivial to generate thousands of unique, plausible reviews. Detection is an arms race.<\/p>\n \n
Users want simple answers. We tried showing detailed breakdowns (timing score: 65, language score: 78, etc.). People ignored it. They just want the letter grade.<\/p>\n \n
Caching is essential. Without it, our server would die from scraping load. With it, we handle hundreds of analyses per day on basic hardware.<\/p>\n\n
Running this isn’t free. Server costs $40\/month. OpenAI API costs $50-100\/month depending on usage. Domain and SSL are another $20\/year.<\/p>\n \n
We’re not making money from this (it’s free, no ads, no subscriptions). We cover costs out of pocket because we think it’s worth doing.<\/p>\n \n
If usage grows significantly, we’ll need to figure out sustainability. Maybe donations, maybe optional premium features. For now, we’re keeping it completely free.<\/p>\n\n
We want to add support for other platforms (Walmart, eBay, Etsy). The algorithms are platform-agnostic, we just need to build scrapers.<\/p>\n \n
We’re working on browser extensions. One-click analysis without leaving Amazon. Chrome and Firefox first, Safari if there’s demand.<\/p>\n \n
We’re also building a public API. Other developers can integrate our analysis into their tools. Free for non-commercial use, paid tiers for businesses.<\/p>\n \n
Long-term, we want to build a database of known fake review operations. Track sellers across products, identify patterns, share data with platforms and regulators.<\/p>\n\n
We can’t catch everything. Sophisticated operations that spread reviews over months, use real accounts, and write genuine-sounding text will slip through.<\/p>\n \n
We can’t verify product quality. A product can have authentic reviews and still be terrible. We only assess review authenticity.<\/p>\n \n
We can’t stop fake reviews from being posted. We can only help you identify them after the fact.<\/p>\n \n
These limitations are inherent to the problem. No tool will ever be perfect. We’re just trying to be good enough to be useful.<\/p>\n\n
Use the tool. The more products we analyze, the better our algorithms get. Each analysis adds to our database and improves detection.<\/p>\n \n
Report errors. If we grade a product wrong, tell us. We investigate every report and adjust our models.<\/p>\n \n
Send feedback. If you can improve the methodology, spot a recurring edge case, or identify a product we graded incorrectly, we want to know.<\/p>\n \n
Spread the word. The more people use review analysis tools, the less valuable fake reviews become. If buyers can easily detect fakes, sellers stop buying them.<\/p>\n\n
Fake reviews cost consumers billions annually. They prop up bad products, hurt legitimate sellers, and erode trust in online shopping.<\/p>\n \n
Platforms won’t fix this alone. Their incentives are misaligned. They need reviews to drive sales, even if some are fake.<\/p>\n \n
Consumers need tools to protect themselves. That’s what we’re building. Free, transparent, and aggressive about catching manipulation.<\/p>\n \n