The SEO Skill Benchmark
Which free SEO skill for Claude Code actually gives correct advice? seo-skill-bench runs every skill for real — headless, three times, one identical prompt — against a fixture site with 10 planted defects and 5 hallucination traps, then scores what each skill actually did against a machine-readable answer key. No vibes, no star counts: every transcript, score, and adjudication is public.
| # | Skill | Composite | Detection | Trap avoidance | Judgment | Execution | Median time | Cost/run | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | SEOAgent ↗ | 84.9 | 81% | 100% | 9.0/10 | 50% | 13 min | $4.46 | |
| 2 | claude-seo-skills (lhitches) ↗★ 1 | 76.0 | 52% | 100% | 8.0/10 | 100% | 7 min | $2.20 | |
| 3 | Agentic SEO Skill ↗★ 713 | 75.4 | 71% | 73% | 9.0/10 | 100% | 10 min | $2.81 | |
| 4 | Vanilla baseline (no skill) | 73.0 | 71% | 100% | 6.0/10 | 100% | 7 min | $2.31 | |
| 5 | claude-seo-skill (mangollc) ↗★ 36 | 70.4 | 67% | 73% | 7.0/10 | 100% | 9 min | $3.08 | |
| 6 | claude-seo ↗★ 10,215 | 67.7 | 38% | 100% | 6.0/10 | 100% | 7 min | $1.70 | |
| 7 | claude-seo-skills (lionkiii) ↗★ 18 | 66.6 | 52% | 73% | 6.0/10 | 100% | 12 min | $2.96 | |
| 8 | Marketing Skills (Corey Haines) ↗★ 36,047 | 66.6 | 52% | 73% | 7.0/10 | 100% | 8 min | $1.87 | |
| 9 | SEO/GEO Claude Skills ↗★ 20 | 66.0 | 57% | 100% | 6.0/10 | 100% | 9 min | $2.16 | |
| 10 | Distribb Skill ↗★ 87 | 65.9 | 43% | 100% | 7.0/10 | 100% | 8 min | $2.33 |
Medians of 3 runs per skill on a pinned model (claude-sonnet-5), fixture pivot-saas. Composite = 40% defect detection + 25% trap avoidance + 25% blind judgment + 10% execution, pre-registered in RUBRIC.md before any runs. Observed fleet-to-fleet variance is roughly ±8 points — gaps inside that band are statistical ties. Stars are GitHub stars at test time. SEOAgent maintains this benchmark and enters its own skill; every trap hit is hand-adjudicated with receipts in ADJUDICATIONS.md.
How the benchmark works
Why you can trust a benchmark we maintain
You shouldn't trust it — you should check it. The rubric was pre-registered before any runs. Scoring is mechanical, against a manifest, with the scorer's self-tests in the repo. Every fleet is committed with full receipts: transcripts, scores, invalid runs, and hand-adjudications of every trap hit. That history includes the fleet where our own skill finished dead last — 9th of 9, below running Claude with no skill at all — and two scoring corrections that lowered our numbers, applied because they were correct. The benchmark caught a data-leak-class bug, a crawler auditing the wrong website, and a session-wedging hook in our product; fixing what it caught is why the skill now leads. Read the full story.
Reproduce it, or beat it
Everything is MIT-licensed: fixtures, manifests, harness, rubric, transcripts, adjudications. One command re-runs any entrant:
git clone https://github.com/aleclindz/seo-skill-bench.git && cd seo-skill-bench node harness/run.mjs --skill <id> --fixture pivot-saas --runs 3
Maintain a skill? Open an issue or PR adding it to skills.json — the next fleet picks it up and publishes your scores, good or bad. The top of this board is one good release away from changing hands. That pressure is the point.
Show your score
Every entrant gets a live badge that re-renders from the current leaderboard — drop it in your README: 
[](https://seoagent.com/seo-skill-benchmark)
Want the skill at the top of the board?
SEOAgent is free for Claude Code, Cursor, and Codex — origin-verified crawls, evidence-cited findings, mechanical hallucination guards, and the field's only real migration planner.
FAQ
What is the best SEO skill for Claude Code?
On the current seo-skill-bench leaderboard, SEOAgent scores highest on the composite (defect detection, hallucination-trap avoidance, blind-judged strategy, execution). The benchmark is open source and every run is reproducible, so the answer can change with any release — check the live leaderboard.
How does seo-skill-bench score a skill?
Each skill runs headlessly, three times, with one identical prompt against a fixture site containing 10 planted defects and 5 traps (things that are already correct — recommending a "fix" is an objectively scored hallucination). Composite = 40% defect detection + 25% trap avoidance + 25% blind judgment + 10% execution, pre-registered before any runs. The published score is the median of the runs.
Can I trust a benchmark maintained by one of the entrants?
Check the receipts instead of trusting anyone: every transcript, score, invalid run, and adjudication is committed to the public repo — including the fleet where SEOAgent finished dead last, and scoring corrections that cut against it. The rubric is pre-registered, scoring is mechanical against a manifest, and one command re-runs any entrant.
How do I submit my SEO skill to the benchmark?
Open a pull request or issue on the seo-skill-bench GitHub repo adding your skill to skills.json (a GitHub repo with a SKILL.md, or an npm package). The next fleet run picks it up, and your scores — good or bad — are published with full receipts.