ML models

The SEO agent trains five models each week: content decay (seasonal trend decomposition), query clusters (embeddings and HDBSCAN), a search-intent classifier, learned recommendation priority, and related articles. No ML runs inside a web request; the site reads the stored results.

A new version becomes active only if it beats the current rule on held-out data (and, for some models, after your review). Its output counts as fresh for 14 days.

• Override auto (default): the model when an active version is fresh, otherwise the rule.
• Override rule: always the rule.
• Override model: the model whenever its output is fresh, even if it never beat the rule.

Every number on the site says where it came from: model vN, rule, rule (insufficient history) or rule (stale model output). Labels lets you review training labels; Run now trains outside the schedule.