How the ML is built, in simple words

Every Sunday the seo-agent runs the ML job. For each model it does the same four things: learn from your past data, test itself on data it has not seen, compare its score with the simple rule the site used before, and save the result. The website only reads the saved results; no ML runs while someone loads a page.

• Falling pages (content decay): looks at each page's daily clicks, removes the normal weekly ups and downs (and yearly seasons when there is enough history), and flags a page only when its trend has fallen steadily for three weeks and clearly more than usual. In tests it raised far fewer false alarms than the old 28-day rule.
• Topic groups (query clusters): turns each search query into numbers that capture its meaning (embeddings) and groups queries that mean the same thing. You check a sample of groupings on the Labelling page; the model is used only when it matches your answers better than simple word matching.
• Search intent: learns from intent labels Claude gave earlier (and your corrections) to tell “how to”, “what is”, comparison and other searches apart by itself, so fewer Claude calls are needed. It is used only when it is accurate enough on your spot checks.
• Priority: looks at recommendations you already carried out and how many clicks they gained, and proposes better scoring weights. Proposals are never applied automatically; you decide.
• Related articles: mixes how similar two articles are, which search queries they share and which pages readers actually visit together, and picks the best six for each article. Clicks on the related block (with consent) measure whether it works.

If a model is not better, too new or more than 14 days old, the site quietly uses the rule, and every number says which one made it.