page-optimization

OpenGSC Page Optimization

Goal

Take one page from "it used to do better" to a rewritten draft the user can publish, with evidence for every change and a mechanical check that the rewrite did not invent a fact.

Required inputs

• The site (domain). Call list_sites if unknown.
• Optionally the page. If the user has not named one, find it — that is step 1.

OpenGSC MCP tools

• get_content_decay: pages trending down, with per-bucket history. The usual starting point.
• get_ctr_benchmark: top-10 queries whose real CTR trails the benchmark for their position.
• get_optimization_brief: the main call. Everything known about one URL — its queries, striking-distance keywords, CTR gaps, six-month trend, cannibalization conflicts, technical audit issues, and the live page as Markdown.
• analyze_text: deterministic verification of your draft. No model is called; the answer is the same every time.
• get_generations: what has already been written for this keyword, so a "refresh" does not become a second competing page.
• fetch_page_content: a competitor's page as clean Markdown, when the brief shows you are being outranked and you need to see by what.
• start_rewrite_job / start_generation_job: PAID and asynchronous. They return a job id, not text; poll get_generation_job. Only on explicit user request — see Guardrails.

Workflow

1. Find the page. get_content_decay for the site (or get_ctr_benchmark if the user's complaint is "impressions are fine, nobody clicks"). Pick the highest-impact candidate and confirm it with the user.
2. Diagnose before writing. Call get_optimization_brief with that URL and read what kind of problem it actually is:
   • Position 4–20 with real impressions → the content is not competitive enough. Rewrite.
   • Top-10 position but a large negative ctrGap → the ranking is fine, the snippet is not. Change the title and meta description; leave the body alone.
   • cannibalization rows with meaningful impressions → another of the user's own pages is competing. Consolidation first; rewriting either page in isolation makes it worse.
   • audit issues like noindex, a canonical mismatch, or a 4xx → fix the technical problem. No amount of prose outranks a noindex.
3. Check for existing work. get_generations filtered by the keyword. Extend what exists rather than producing a near-duplicate.
4. Write it yourself. You have the brief, the current body, the target queries and the competitors' gaps in context. Keep every number, price and product name from the original unless the brief gives you a reason to change it.
5. Verify. analyze_text with your draft as text and the original body as source. Read three things:
   • factDrift.severity — danger means values appear in your draft that are not in the source. Fix them before showing the user. This is not advisory.
   • structure.ok — false means you dropped or added headings. Usually accidental.
   • uniquenessPercent — very high on a refresh means you rewrote rather than updated, which throws away whatever was already ranking.
6. Present the draft with the evidence for each substantive change.