SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsgsc-analysis

gsc-analysis

Marketing & growth

View repository

Updated 10-16 05:54

About this skill

GSC analysis

Turn Search Console data into a diagnosis and a short list of actions ranked by business impact. Every number comes from GSC; every cause is labelled as a hypothesis unless the data proves it.

Preflight

Run the shared task preflight and tooling contract. For this skill:

  • Resolve the GSC property from the accessible list (sc-domain: and URL-prefix variants), and state which one you use.
  • Read .seo-context.md and any existing analyses in the project first, so you do not repeat them and so priorities follow the confirmed money pages, ICP, and conversions.
  • Confirm the question: health check, drop diagnosis, opportunities, a specific page or cluster. Default to a health check plus opportunities when the user just says "check the GSC".

Tooling & credentials

  • Auth mode: mcp
  • Requires: a GSC MCP whose search analytics tool accepts a property, an explicit start/end date, dimensions, filters, and paging. AminForou/mcp-gsc covers all of these (get_advanced_search_analytics: row_limit up to 25,000, start_row, start_date/end_date, site_url on every tool, plus batch_url_inspection).
  • MCPs that bind one default property or only accept "last N days" still work for the default property; say which analyses they prevent.
  • Script: scripts/gsc_rows.py, Python 3 standard library only.
  • Fallback: none for live data. A user-provided GSC UI export can be analyzed with the script, labelled as an export.
  • If missing: stop, ask the user to connect a GSC MCP, and continue after they confirm.

Data rules

Apply these before reading any number:

  1. Exclude the last 2-3 days (incomplete data) or request final data only.
  2. Impressions are not search volume, and anonymized queries mean query totals are lower than page or site totals.
  3. Weight positions by impressions whenever rows are summed. Never average position columns.
  4. Recompute CTR from clicks / impressions. MCPs return it as a percent or a fraction.
  5. Compare equal-length windows starting on the same weekday. Add year-over-year when the site is seasonal.
  6. Split brand from non-brand before judging SEO performance. Brand clicks hide non-brand losses.
  7. Check completeness. If a call returns exactly row_limit rows or has_more, page with start_row or narrow the filter. Never conclude from a truncated result.
  8. Regex filters (includingRegex, excludingRegex) are a Search Console API feature; some MCPs pass them through without documenting them. Test once; if refused, pull wider data and segment with the script.

Pulling data efficiently

  • Pull only the dimensions the question needs. date alone for trends; page or query for rankings; query,page only for cannibalization or page-level query mapping.
  • Large results are often saved to a file by the agent runtime. Do not read them into context: run the script on the file.
  • Save each pull to a scratch file (JSON as returned) when it will be reused, and name it by property, dimensions, and date range.

Workflow

  1. Trend first. Pull date for 16 months, then gsc_rows.py aggregate <file> --by month. Locate breaks before looking at pages.
  2. Run the recipes the question calls for from references/recipes.md:
    • drop diagnosis, content decay, quick wins, CTR gap, content gaps, cannibalization, intent and brand segmentation, page refresh inputs, GEO query signals, indexing checks.
  3. Tie findings to the business. Rank actions by proximity to conversion (money pages, BOFU queries, ICP topics) first, then by potential clicks. Say when a large traffic opportunity has low business value.
  4. Verify the key numbers in the final answer against the pulled data before presenting them.

Interpretation rules

  • Separate facts (numbers, dates, pages) from hypotheses (causes). Name the checks that would confirm a cause: site changes on that date, Google update dates, URL inspection, SERP inspection.
  • A page in positions 1-3 with CTR under 1% is not a title problem: the SERP already answers the query (featured snippet, AI Overview, tool). Recommend a reason to click (calculator, worked example, template, data) rather than a title rewrite.
  • Impressions down with position stable means Google shows the page for fewer queries or the SERP changed; it is not a content-quality ranking loss. Check the queries the page lost.
  • Cannibalization is a problem only when several pages hold a meaningful share of the same query's impressions and positions are unstable. Low-impression secondary URLs are normal.
  • Do not recommend creating new pages for queries an existing page already ranks for in positions 5-20; improve that page first.

Output

  • Lead with a 3-5 line summary: the current state, the main break or opportunity, and the first action.
  • Then one table per recipe used, with exact figures and the date range and property stated once.
  • End with prioritized actions: action, target page, evidence, expected impact, and whether it is a quick win or a project.
  • Write in the user's language.
  • Ask before saving a file. Present the analysis in chat. Then ask whether to save it and where; suggest a dated name such as YYYY-MM-DD-gsc-<topic>.md in the project's client or SEO folder. Do not write files without that answer.

Related skills

  • Refreshing one page from its queries: subkeyword-injector.
  • Link fixes for underlinked targets or consolidated pages: internal-linking.
  • Reviewing a page flagged here: seo-roast.
  • A reusable interactive dashboard from GSC data: seo-audit-report.