Every page we publish here is built on landing pages we scored ourselves. roast.page runs each URL through three pipelines — screenshots at desktop and mobile viewports, an HTML extractor that captures every heading, button, and trust signal, and Google PageSpeed Insights for the Core Web Vitals. A Claude vision model then scores eight weighted dimensions on a 1–10 scale and writes the recommendations.
Studies sample from this scored dataset. Where we make claims like “outcome-driven headlines score 14 points higher,” we mean: we hand-classified the H1 of every page in the sample, ran a two-sample comparison, and report the mean difference. Where a finding is correlational, we say r. Where the sample is small or skewed, we say so. We don’t round inconvenient numbers up.
Citations are welcome. Please link back to the specific study, not the homepage, so readers can verify the methodology. We update quarterly as the dataset grows.