Which Free Claude SEO Skill Is Actually Best? I Benchmarked 8 on My Own Site

TL;DR — There are dozens of free SEO skills for Claude Code now and no honest comparison of them. So I ran 8 of them — plus vanilla Claude Code as a baseline — against a real SaaS site with real Google Search Console data, and scored what they actually did, not what their READMEs claim. GitHub stars turned out to be a terrible predictor of quality. The skills that won grounded their advice in the live site and the real data; the ones that lost hallucinated problems that didn't exist. Full disclosure: one of the entrants, SEOAgent, is my product — so the methodology is all here and you can reproduce it yourself.

Why I built this benchmark

Search "SEO skill" on GitHub and you'll hit a wall of them — claude-seo, Agentic-SEO-Skill, seo-geo-claude-skills, a dozen more, most published in the last few months. Every one promises keyword research, technical audits, schema, GEO. None can tell you which is any good, because "good" for an SEO skill isn't a feature list — it's whether the advice it gives is correct for your actual site.

So I built a benchmark that measures exactly that: I pointed each skill at a real site with real Search Console data and checked every claim it made against reality.

How the test worked

The control site (and the twist that makes it a real test)

The control site is usebaxter.com, another product of mine — so I have the real Google Search Console data to check every claim against.

Here's the twist, and it's the whole point. Baxter is mid-pivot: the live site is now a family/household AI assistant, but 100% of its historical search traffic is from its previous life as an email-management tool. That mismatch is deliberate, and it's the perfect intelligence test. A skill that just extrapolates your Search Console history will confidently build you a strategy for the old business. A skill with real judgment reads the live product, infers where you're actually going, and treats the legacy rankings as authority to migrate — not a target to chase.

The rules

• Same prompt for every entrant: "Increase organic traffic for this SaaS."
• Each skill runs on its own documented method — no coaching, no hints about the pivot. Figuring that out is the test.
• Every factual claim is verified against the live site and the real GSC export (28 days: ~18,305 impressions, ~27 clicks, dozens of blog posts stuck at positions 50–80).
• Free skills only. Anything requiring a paid Ahrefs / Semrush / SE Ranking key for its core flow was excluded.

I scored five things: did it infer the real business direction, did it verify against the live site, did it produce a migration plan for the legacy content, did it ship real artifacts, and how much did it hallucinate.

The results at a glance

#	Skill	Score /100	Migration plan	Hallucination	Free?	GitHub stars
1	SEOAgent	90	✅ Concrete	0%	Free	— (npm)
2	SEO/GEO (aaron-he-zhu)	82	❌	2%	Free	20
3	Hawk Academy	78	✅ Concrete	15%	Free	1
4	Agentic-SEO-Skill	68	❌	5%	Free	713
5	Claude SEO (mangollc)	58	❌	18%	Free¹	36
6	42 SEO Commands (lionkiii)	48	❌	8%	Freemium	18
7	Claude SEO (AgriciDaniel)	48²	~ Partial	42%	Free	10,215
8	Vanilla Claude Code	30	❌	5%	—	—
9	Distribb	10	❌	10%	Freemium³	87

¹ Needs a free Firecrawl key. ² AgriciDaniel's clean-run scorer errored; score is from the earlier un-coached run (same method, unchanged skill). ³ Free tier exists, but most of its flow is gated behind a paid API key — see below.

SEOAgent finished first, ahead of aaron-he-zhu's SEO/GEO skill and Hawk Academy. The margin matters as much as the ranking — I'll get to that in the caveats — but first, the finding that surprised me most.

Do GitHub stars predict quality?

Barely.

Skill	GitHub stars	Benchmark score
Claude SEO (AgriciDaniel)	10,215	48
Agentic-SEO-Skill	713	68
Distribb	87	10
Claude SEO (mangollc)	36	58
SEO/GEO (aaron-he-zhu)	20	82
42 SEO Commands (lionkiii)	18	48
Hawk Academy	1	78

The most-starred skill in the field, AgriciDaniel/claude-seo with 10,215 stars, landed mid-pack. The two best open-source finishers had 20 stars and 1 star. There's essentially no correlation between how popular a skill is and how well it performs on a real site — stars measure marketing and README quality, not whether the advice is correct.

If you've been picking skills by star count, stop.

What separated the winners from the losers

Run uncoached, the field split cleanly on two capabilities almost nobody had: verifying against the live site, and producing a real migration plan. Everything else — keyword lists, schema snippets, generic "improve your titles" advice — was table stakes every skill (and vanilla Claude Code) could do.

Skill	Inferred intent	Verified crawl	Migration plan	Shipped artifacts	Low hallucination
SEOAgent	✅	✅	✅	✅	✅
SEO/GEO (aaron-he-zhu)	✅	✅	❌	❌	✅
Hawk Academy	✅	❌	✅	❌	❌
Agentic-SEO-Skill	✅	✅	❌	✅	✅
Claude SEO (mangollc)	✅	❌	❌	✅	❌
42 SEO Commands (lionkiii)	✅	❌	❌	✅	❌
Claude SEO (AgriciDaniel)	✅	❌	~	✅	❌
Distribb	❌	❌	❌	❌	❌

"Verified crawl" means it checked the live site and didn't recommend adding tags or schema it already serves. "Migration plan" means a real per-asset decision — harvest, redirect, or sunset — for the legacy content. Of nine entrants on their own methods, exactly two produced one. That scarcity is the whole story of the ranking.

The trust test: hallucination

This is where the popular skills fell apart. When you check every claim against the live HTML and the real GSC data, the failure mode isn't vagueness — it's confident, specific, wrong.

Skill	Hallucination rate	The costly miss
SEOAgent	0%	Recommended nothing that already existed live
SEO/GEO (aaron-he-zhu)	2%	—
Agentic-SEO-Skill	5%	A couple of unverifiable source-line claims
42 SEO Commands (lionkiii)	8%	Claimed the blog has 0 posts — it has 76, earning 18,305 impressions
Distribb	10%	Halted before it could inspect the site
Hawk Academy	15%	Re-recommended schema the site already serves
Claude SEO (mangollc)	18%	Claimed the homepage is client-side-rendered with no copy — it's server-rendered
Claude SEO (AgriciDaniel)	42%	Audited a stale copy of the repo; flagged "fixes" that are already live

Two well-known skills let a hallucination drive their #1 recommendation. One insisted the homepage had no content; another declared the blog empty. Both were verifiably false against the live site — and both were the skill's headline fix. This is the entire case for grounding an SEO agent in real data: plausible-but-wrong is the default failure mode, and it's invisible unless something checks.

How each skill performed

SEOAgent (1st)

The only entrant to combine a verified live crawl, a concrete per-URL migration plan (76 legacy blog URLs sorted into harvest / redirect / sunset), correct intent inference, and shipped artifacts in one run — with zero hallucinations. Disclosure: it's my product.

SEO/GEO by aaron-he-zhu (2nd)

The trust leader. Near-zero hallucination, the sharpest read of the pivot, exact reconciliation against GSC. It just didn't ship a concrete migration plan or artifacts.

Hawk Academy (3rd)

One of only two skills to produce a real migration plan from its own method. Docked for re-recommending already-live schema.

Agentic-SEO-Skill (4th)

An honest, well-verified technical audit — but it skipped GSC entirely, so its strategy floated free of the site's real demand, and it built no migration plan.

The rest

mangollc and lionkiii both shipped useful artifacts but let a hallucination anchor their top fix. AgriciDaniel — the 10,215-star flagship — audited a stale repo and hit a 42% hallucination rate. Distribb halted at its own API-key gate before it inspected anything.

Full disclosure and the honest caveats

SEOAgent is my product, so take the #1 with the appropriate salt — here's the unvarnished version:

• SEOAgent did not dominate every axis. aaron-he-zhu's skill matched or slightly beat it on pure verification and hallucination — it was the trust leader. SEOAgent won on breadth: it was the only entrant to do all five things well in one run. The margin is a real ~8 points, not a blowout. If a competitor bolted a migration planner onto its already-strong verification, this would be a tie.
• It's one run per skill, on one site. Directionally solid, not a lab result. Treat it as a starting point.

I'd rather publish that than a fake landslide. A comparison you can't poke holes in is the only kind worth reading.

How to reproduce this

The point of a benchmark is that you don't have to trust me. Point any of these skills at your own site with your own Search Console data, ask it to grow your traffic, and check its claims against your live pages. The skills that hold up are the ones that looked before they spoke.

The bottom line

Most free Claude SEO skills give the same generic advice; a few give confidently wrong advice; and the ones worth using are the ones that read your live site and your real data before saying anything. If you want the one that finished first: SEOAgent ships as a free skill for Claude Code, Cursor, and Codex — it audits your site against a live crawl, plans content migrations from your real GSC data, and writes every change as a commit you approve. Start with the free skill. For the wider landscape, see our roundup of the best SEO tools for Claude Code.