A published article is a hypothesis, not a result. Search Console turns it into evidence: impressions show which queries a page can compete for, average position shows how close it is to the visible results, and click-through rate shows whether the title and description earn the click when the page is shown.
Pick a window and keep it fixed
Search data is noisy at short range. A 28-day window compared with the previous 28 days is long enough to survive a holiday dip and short enough to catch a regression before it compounds across a whole cluster. Changing the window between reviews makes every comparison ambiguous, so the window is a decision to make once, not per report.
Act on bands, not single days
Pages sitting between positions four and thirty are usually the cheapest wins. They already rank, so a clearer title, a tighter introduction, or one genuinely missing section can move them into a stronger position. Pages outside the top hundred rarely respond to the same treatment; they need a different angle or a different page, not another rewrite.
Indexation deserves the same discipline. If a large share of submitted pages are not indexed, publishing more pages will not fix the underlying problem. Slowing down and reviewing the pages that were skipped is usually the faster route back to growth.
Rollback is part of measurement
If an optimisation coincides with a three-position drop across the next comparison window, revert it and keep the original. That is not a failure of the process; it is the process working. Recording what changed, when it changed, and what happened afterwards makes every later change cheaper to evaluate, and makes it possible to tell a bad edit from normal fluctuation.
A loop only closes when evidence changes the next action. Reports that are read but never acted on are expensive documentation. A short review that ends with two or three named pages, a specific edit for each, and a date to check them again is worth more than a complete dashboard nobody uses.