Recommendations and change drafts

• A title, description, priority (critical, high, medium, low) and category.
• The evidence: findings, pages or search data it is based on.
• Why it matters, how to tell it worked (a falsification check) and a leading indicator.
• Impact and effort estimates, and the analyst's confidence.

New recommendations are pending. Review them in a site's Recommendations tab or across all sites under Review in the sidebar. Approve, or reject with a reason; select several to approve or reject in bulk.

Decisions carry over: when a later audit produces the same advice (same category and evidence), it arrives already approved or rejected, and older pending copies are marked superseded. Your reasons also feed the analyst's [memory](/docs/ai-analyst#memory).

After each later audit, every approved recommendation is re-checked:

• Improving: fewer affected URLs than when it was recommended, or 28-day clicks on the affected pages up 10% or more.
• Worse: more affected URLs, or clicks down 10% or more.
• Flat: neither.

Open an approved recommendation and choose Draft the changes. The analyst reads up to 20 affected pages as crawled and proposes concrete edits: titles, descriptions, canonicals, robots directives or structured data, page by page.

• The current value always comes from the crawl, never from the model.
• Titles over 60 characters, descriptions outside 70–160 characters, JSON-LD that doesn't parse, and edits to pages the recommendation doesn't cite are flagged.
• Copy each value, or export the draft as a Markdown checklist or JSON.
• Mark the draft applied when the changes are live; or Draft again any time.