Keyword Cannibalization Checker Skill

Find the queries where your own pages compete with each other in Google, from your Search Console data: which pages split the impressions and clicks, which ones keep swapping places, which page pairs overlap across many queries, and which URL variants (trailing slash, www, tracking parameters) split signals.

python3 -m cannibal query-page.csv --brand 'widgetly' --csv reports
python3 -m cannibal bigquery-export.csv --search-type web --json > cannibal.json

Search Console query and page together (the UI export only gives them separately):

• Search Analytics API, Looker Studio or a Sheets add-on export with columns like query, page/url, clicks, impressions, position (comma, semicolon or tab separated; 1,234 and 1.234,5 numbers handled).
• BigQuery bulk export rows from searchdata_url_impression (query, url, impressions, clicks, sum_position, data_date, search_type, is_anonymized_query). Rows are aggregated, average position is computed as sum_position / impressions + 1 (Google's formula), anonymized queries are skipped and --search-type filters web/image/video.

• Cannibalized queries: two or more pages each with at least 10% of the query's impressions (--min-share) within the top 30 positions, graded HIGH (real clicks split between two top-10 pages, or the ranking page flipping 3+ times when daily data is present), MEDIUM and LOW.
• Competing page pairs: pages that share the most queries, with example queries and advice (merge and 301, or sharpen titles and internal links).
• URL variants: the same page under different URLs splitting a query.
• CSV (--csv DIR): cannibalization.csv and page-pairs.csv.

skills/keyword-cannibalization/SKILL.md tells Claude Code (or any agent that reads skills) how to get the right data, run the check and recommend fixes. Copy it into .claude/skills/; the kit is also a Claude Code plugin (.claude-plugin/plugin.json).

Run ./demo.sh for the example (9 queries, 6 cannibalized, one ranking flip) and ./run_tests.sh for the tests. Python 3.9+, no dependencies, works offline.

It sees Search Console numbers, not intent: two pages can rightly rank for one query. Review both pages and the live results before merging anything. Queries Google anonymizes aren't in the data.

Pairs with the SEO Log File Analyzer and SEO Audit MCP Server. License: LICENSE.txt.