Abstract:
Online reviews dominate purchase decisions, yet the share of AI-generated feedback on Google
Maps alone has jumped from about 5 % in 2019 to nearly 19 % in 2024, eroding consumer
trust. Google says it removed more than 240 million policy-violating reviews in 2024 41 %
more than the year before and recent surveys show trust in reviews has fallen from 79 % to 42
% in just five years. ShopFinder answers this credibility crisis with a lightweight Flask service
that ( uses Playwright to scrape exactly the number of fresh Google-Maps reviews a user
requests; (ii) filters out AI-written text via a fine-tuned DistilBERT detector that achieves 96.5
% accuracy / F1 on a balanced test set well above the 90–94 % usually reported for state-of-
the-art models; and recalculates a transparent shop score with a hybrid DistilBERT-embedding
+ XGBoost regressor that predicts human-only star ratings at 0.85 F1 while Optuna-guided
tuning keeps inference fast. Every decision is paired with SHAP and LIME visualisations,
exposing the tokens and metadata that most influenced the model’s call. Users can further filter
results to businesses that will be open tomorrow, combining authenticity with geospatial
relevance a gap regulators now demand platforms address. By uniting high-recall AI detection,
real-time scraping and word-level explanations in a single, non-containerised service,
ShopFinder restores confidence in review-based decision-making without heavy infrastructure.
