ML Intern Builds Content Decay Predictor Using 79M Rows of Search Data

A machine learning intern at FlyRank developed an offline analytical pipeline designed to predict content decay before organic traffic drops, using historical search telemetry. The system processes anonymized search performance data and outputs a prioritized action queue recommending whether content should be refreshed, monitored, or expanded. The final model, a tuned Random Forest classifier, achieved a ROC-AUC of 0.750 and a Precision@50 of 0.74, far surpassing a naive recency baseline that scored just 0.24. A key architectural decision was replacing standard k-fold cross-validation with a strict chronological holdout split to prevent future data from leaking into training, making evaluation metrics more production-realistic. The builder acknowledges the model identifies statistical signals of decay but cannot account for unpredictable external events such as search engine algorithm updates or sudden shifts in user intent.