linkrank

Internal PageRank (Link Score) & keyword-cannibalization auditor — free & open-source.

linkrank crawls a website, builds its internal link graph, computes internal PageRank on a 0–100 “Link Score” scale (the same idea behind Screaming Frog’s Link Score and OnCrawl’s InRank), and flags keyword cannibalization and orphan pages — all locally, with no paid APIs.

It answers questions that usually require paid tools:

• Which pages does my internal linking actually concentrate authority on — and which are starved?
• Which pages are orphans (zero internal inlinks) or buried too deep?
• Where is my site cannibalizing itself (one anchor → many targets, duplicate titles, near-duplicate content)?

Made by KlientLab. MIT licensed.

Install

pip install linkrank

Quick start

linkrank https://example.com

Outputs a machine-readable JSON and a self-contained HTML report into ./reports/:

=== Internal link audit: https://example.com/ ===
pages=253  internal_links=1653  orphans=39  deep(>=4)=39  kw_clusters=11

Top 10 by Link Score:
  100.0  in= 253  https://example.com/
   28.2  in=  37  https://example.com/category-a/
   ...
JSON: reports/internal-link-audit-example-com-2026-07-06.json
HTML: reports/internal-link-audit-example-com-2026-07-06.html

Options

linkrank <start_url> [options]
  --max-pages N          cap fully-crawled pages (default 2000)
  --depth N              max crawl depth from the start URL (default 10)
  --timeout S            per-request timeout, seconds (default 12)
  --workers N            concurrent requests (default 10)
  --sim-threshold F      content-similarity cutoff for cannibalization (default 0.80)
  --include-subdomains   treat subdomains as internal
  --out DIR              output directory (default ./reports)
  --json-only            skip the HTML report

What it reports

• Link Score (0–100) per page — log-scaled internal PageRank, comparable to Screaming Frog.
• Inlinks / outlinks, click-depth from the homepage, orphan pages (0 internal inlinks).
• Cannibalization, three signals:
  1. one anchor text → multiple target URLs;
  2. pages sharing the same primary keyword (normalized <title>/<h1>);
  3. near-duplicate content (TF-IDF cosine ≥ threshold).

Methodology

• Internal PageRank — power iteration, damping factor α = 0.85 (Page & Brin, 1998). Each page passes its score divided by its number of followed out-links; rel="nofollow" links evaporate (counted in out-degree but transfer no score); dangling nodes are redistributed uniformly. The raw PageRank is reported both directly and as a 0–100 logarithmic Link Score (min–max on log), matching the way Screaming Frog Link Score and OnCrawl InRank present internal authority.
• Graph — nodes are 200-OK internal HTML URLs; a page’s canonical target is used as its node identity when that target is itself a valid page; redirects are collapsed onto their final URL.
• Cannibalization — anchor-to-many-targets from the crawl, primary-keyword clustering from <title>/<h1>, and TF-IDF cosine similarity of body text.

References: Page, Brin, Motwani & Winograd, The PageRank Citation Ranking (1998); Screaming Frog Link Score; OnCrawl InRank.

Requirements

Python ≥ 3.10. Dependencies (installed automatically): requests, selectolax, networkx, scikit-learn.

Notes & limits

• Respects robots.txt; skips non-HTML, mailto:/tel:/anchors and asset URLs.
• Deterministic output (stable sort) — runs are diffable over time.
• Enforces a hard page cap and per-request timeout — it will not hang on large/slow sites.
• Static-HTML crawler: links injected purely client-side by JavaScript may not be seen. Server-rendered <a href> links (the vast majority) are captured.

Contributing

Issues and PRs welcome. Ideas on the roadmap: JS rendering flag, GSC query×page cannibalization signal, interactive graph export.

License

MIT © 2026 KlientLab.