Best Claude SEO Skills: 5 Packs Tested on a Real Site

On 22 October I typed "claude seo skills" into Google. The first result was a GitHub repo with 19,600 stars. Two searches later I found a Reddit thread warning that the same skill quietly adds its own promo footer to your audit.

That's the whole problem in two searches. There are hundreds of these packs now, seoskills.sh alone lists 423, and nearly everything written about them is a list. Nobody runs them side by side.

So I did. I planted 20 SEO problems on a copy of a small site, wrote down the answers first, then gave the same job to 5 skill packs and to plain Claude Code with no skill at all. When that turned out too easy, I built a second copy with 20 harder problems and ran it again.

The short version: plain Claude Code found 19 or 20 of the 20 easy problems on its own. The skills only pulled ahead on the hard set, where marketingskills and SE Ranking's seo-skills caught all 20 and plain Claude caught 18. Every audit cost between $0.26 and $0.57.

I run Claude Code every day for my own SEO and Google Ads work. None of these packs pays me, and I'm not selling one. The only skill I give away is my own seo-skyscraper-research skill, free at the bottom of this post.

Best Claude SEO skills: the verdict table

Every audit ran on Claude Sonnet 5.5 with only that pack installed. Two graders (Claude Opus and Codex) scored each audit blind against an answer key I wrote before the first run. The easy set ran with the full evidence, and also again with only the URL; the scores below are the first URL-only run of each pack.

The hard-set column is one run per pack. The 18s and the 16 were scored the same by both graders; the two 20s were scored by Codex only.

Which one should you install?

Short answers, from the test:

• Best overall: marketingskills (seo-audit). It caught all 20 on both sets, it's free, it's prompt-first, and it runs anywhere skills run. It's also the most installed SEO skill out there.
• Best if you pay for SE Ranking: SE Ranking's seo-skills. It matched marketingskills even without its data connection. With the MCP connected, it pulls real rankings and backlinks, which the others can't.
• Best free, no install: Hawk Academy. Single .md files. Download one, drop it in a folder, done. Nothing runs on your machine.
• Skip if you want a checklist: open-seo. It's built to pick the 1 to 3 changes that matter most, so it misses small things on purpose. Good for a quick "what should I do first", not for a full audit.
• Not on Claude Code? Only the prompt-only packs move cleanly to Claude.ai. See where each pack works.

And the honest one: on a small site, plain Claude Code is already good. If your site has 20 pages and you're not sure, try a plain "audit this site" first. A skill earns its place on the stuff Claude skips, like HTTP headers and mobile tags.

What a Claude SEO skill is

A skill is a folder with a SKILL.md file in it. The top of the file says what the skill is for. The rest is the steps Claude should follow. It can also carry scripts, templates and reference files, which Claude only reads when it needs them.

Claude loads a skill in two ways: it matches your request to the skill's description, or you call it by name, like /seo-audit. In my test, every pack loaded its skill when I named it.

There are two kinds, and the difference matters more than the brand:

A skill isn't the same as an MCP server. MCP gives Claude data and tools (your Search Console, a rank tracker). A skill gives Claude a process. The good packs use both: the skill says what to check, and the MCP fetches the numbers.

Anthropic's skills documentation covers the format.

Where each pack works and how to install it

Any chat, by copy and paste: a skill is a text file, so the prompt-only ones (marketingskills' seo-audit, open-seo, Hawk Academy) also work pasted into ChatGPT, Gemini or a Claude.ai chat. claude-seo and SE Ranking lean on scripts and an MCP, so pasting gives you only their checklist.

Claude.ai in the browser: go to Customize > Skills, click +, then Create skill > Upload a skill, and pick a ZIP of the skill folder. It works on every plan, Free included, but only with code execution on (Settings > Capabilities > Code execution and file creation). Prompt-only skills work fine there. Scripts run too, in Claude's sandbox, but network access depends on your settings, and an MCP the skill expects isn't there unless you add it yourself.

To update: run the same install command again, or /plugin marketplace update <name> for plugins. To remove: delete the skill's folder, or /plugin uninstall.

How I tested them

I didn't want another list, so I built a lab.

1. A fake site with real problems. A small coffee roaster site (17 pages, then 23 in the second set) with 20 SEO problems planted on purpose. For example, a noindex left on the wholesale page, a canonical pointing at the wrong page, a broken link, a redirect chain, and a page that only exists in the sitemap.
2. The answers first. I wrote the answer key before the first run and saved its fingerprint (a sha256 hash), so I couldn't move the goalposts later.
3. The same job for everyone. Same prompt ("do a full SEO audit and write it to audit.md"), same model (Claude Sonnet 5.5), a fresh folder for every run, and only that pack's skills installed. Plus plain Claude Code with no skill, as the baseline.

4. Two conditions. First with a crawl export, the raw HTML and a Search Console file. Then with only the URL, so the skill had to find everything itself.
5. A harder second set. The first set was too easy (more on that below), so I fixed all 20 problems and planted 20 harder ones: a noindex sent only in the HTTP header, a 302 that should be a 301, a wrong hreflang code, content that only appears with JavaScript, a schema price that doesn't match the page, and so on.
6. Blind grading by two graders. Claude Opus and Codex each scored every audit against the key without knowing which pack wrote it. They agreed on almost every score.

Rules I kept the same for every pack: no web search, and no running the packs' own scripts, so the test is about the skill's instructions. 25 audits in total.

What I can't tell you from this: which pack is best for local SEO, e-commerce, big sites or link building. One small site can't test those.

Best Claude Code SEO skills, rated

marketingskills (coreyhaines31): best overall

• What it is: a big marketing pack (52 skills, 56K GitHub stars, MIT). The SEO part is seo-audit, plus ai-seo, schema and programmatic-seo.
• Score: 20 of 20 on the easy set. 20 of 20 on the hard set, one of only two packs to get there.
• Cost: $0.36 to $0.54 per audit.
• Good: a real checklist. It caught the header-only noindex and the missing viewport tag that plain Claude missed.
• Watch out: it's a big install if you only want SEO. You can install one skill with --skill seo-audit.

SE Ranking seo-skills: best if you have SE Ranking

• What it is: 32 skills from the SE Ranking team (MIT), built around their MCP data server.
• Score: 20 of 20 on both sets, without the MCP connected.
• Cost: $0.40 to $0.50 per audit.
• Good: very thorough technical checks. Its hard-set audit listed 50 findings, the most of any run.
• Watch out: the skills are written for SE Ranking's data tools, so the real value needs an account. The install.sh is designed to be piped from curl into bash; read it first.

claude-seo (AgriciDaniel): the famous one

• What it is: a full plugin: 26 skills, 19 agents and 170 scripts (19.7K stars, MIT, v2.4.2).
• Score: 20 of 20 on the easy set. 18 of 20 on the hard set, the same as plain Claude: it missed the header noindex and the viewport tag too.
• Cost: $0.39 to $0.47 per audit.
• Good: clear priorities and a sensible roadmap at the end.
• Watch out: it was the only pack whose audits suggested restructuring content "for AI answers". The Reddit PSA is real: the main /seo skill still tells Claude to add a "Built by agricidaniel, join the AI Marketing Hub community" footer with two Skool links after major commands. In my one /seo audit run the instruction loaded but Claude didn't print it. Check your output before it goes to a client.

open-seo (every-app): a prioritiser, not an auditor

• What it is: skills for the OpenSEO app (24.5K stars, MIT). Prompt-only.
• Score: 17 of 20 on the easy set, 16 of 20 on the hard set, the lowest.
• Cost: $0.32 to $0.57 per audit.
• Good: it doesn't drown you. It leads with the one to three changes that matter most and explains them to a non-expert.
• Watch out: by design it skips small problems. In one run it missed a missing H1 and then said every page had one. Its skill also says to stop if OpenSEO isn't connected; it didn't stop, it just carried on.

Hawk Academy (StudioHawk): best free download

• What it is: 56 free single-file skills from an SEO agency. I used 7: technical-seo-audit, cannibalization-detector, striking-distance-finder, schema-markup-auditor, internal-linking, title-tag-optimizer and screaming-frog-analyser.
• Score: 20 of 20 on the easy set. 18 of 20 on the hard set: it missed the noindex sent only as an HTTP header, and the product page linked only through a button.
• Cost: $0.26 to $0.39 per audit, the cheapest.
• Good: nothing runs on your machine, and you can read each skill in a few minutes.
• Watch out: you have to collect them one page at a time.

Do Claude Code SEO skills give instructions that Google says to ignore?

Google updated its guide to optimizing for generative AI features in Search in July 2026. In short: Search doesn't use llms.txt, there's no special schema for AI, and you don't need to chunk or rewrite content for AI.

What the packs said:

That's better than I expected. A year of GEO hype, and none of these packs pushed llms.txt.

Can SEO skills improve my website's GEO?

Not by much, because of what the GEO hype leaves out. When people ask an AI about you, or about the service you sell, it rarely cites your own website. That's even more true for service queries, which usually get more impressions and more competition. Ask for "the best marketing agency in Bali" and the answer mostly comes from listings and directories, and most of those are paid.

Google's AI Overviews are the odd one out, and not because they're smarter. Where there's little competition, they'll cite a company's own page praising its own brand. So before you rewrite your headings for AI, check where the answers for your queries actually come from.

Giving Claude real data

A skill without data guesses. On a real site, the audit gets much better when Claude can see your numbers. Three ways:

• Search Console. Export the Performance report (queries and pages) as a CSV and give it to Claude. It's free and it's the best data you have. In my lab, the Search Console file is how the packs spotted two posts competing for the same keyword.
• Keyword Planner. For demand. I used it to pick this topic: "claude seo skills" gets 1K to 10K searches a month, and "claude skills" is up 99,900% year on year (from almost nothing).
• Paid data through MCP. SE Ranking, DataForSEO and OpenSEO connect through MCP and give Claude live rankings and backlinks. It's worth it when you audit sites you don't own, or need competitor data.

People on Reddit say the same. One site owner with 4.53 million impressions in 3 months got the biggest win from making the title tag, H1 and og:title identical. One page went from 0.16% to 1.04% click-through.

The easy way: install the Claude in Chrome extension and connect it to the same Claude account you work in. Claude can then use your open Search Console and Keyword Planner tabs, read the numbers and even export the reports it needs by itself.

Is it safe to install Claude SEO skills?

Mostly, if you read what you install. I looked at every pack before installing it:

None of them had hidden shell commands in the skill text (in Claude Code, a ! right before a command in backticks runs it as soon as the skill loads). None of the scripts I grepped call anywhere unexpected.

The bigger risk is the ecosystem, not these five. Snyk's ToxicSkills study (February 2026) found a critical issue in about 1 in 7 of the nearly 4,000 skills it checked from public directories. Datadog showed in May 2026 that a skill's shell commands can run before the model even reads it.

Reading the files isn't always enough, though. In June 2026 the security firm AIR published a fake skill to test this. Called brand-landingpage, it told the agent to read the "docs" on a domain AIR owned. At first that domain sent visitors to Google's real docs, so every scanner passed it. Once the skill had spread, AIR changed the page to tell agents to download and run a script. In AIR's words:

26,000 agents were affected, including corporate accounts' agents. We could have had full control of every one of their agents, their private conversations, and every internal system they could reach.

Their script only collected email addresses. The skill file itself never changed, so reading it or pinning a version wouldn't have caught it.

My rules:

• Read the SKILL.md and every script before installing. If there are 170 scripts, at least read the ones the skill calls.
• Install per project, not globally, while you try a pack.
• Pin a version, and read the diff when you update.
• Be extra careful with third-party skills. A safe skill today is no promise of a safe skill tomorrow: the author can change it, or the pages it links to, at any time. Be most wary of skills that tell Claude to follow a link for their instructions.
• Never send an audit to a client without reading it. That's how the claude-seo footer got out.

One more from Reddit, because it's the scariest: the same site owner let Claude "improve" their top articles, and it replaced three of them with 200 to 300 word stubs. They got them back from the Wayback Machine. Let a skill suggest changes, not make them, and compare word counts before and after.

Build a small SEO skill yourself

The best skill is often a short one you write for your own process. Here's the research process behind this post, written down as a skill. It checks demand, reads the SERP and the competitors, finds the gaps, and writes the outline.

Step 4 needed a workaround: Reddit. The Claude in Chrome extension won't open reddit.com, and Reddit blocks curl and WebFetch too. So Claude reads Reddit on my Android tablet instead, over ADB (Android's USB debugging tool). It opens a thread in the Reddit app with adb shell am start and the thread's URL, dumps the text on screen with uiautomator dump, scrolls and reads the next screen. That's how it read the five full threads quoted in this post, comments included, not just Google's snippets of them. It's slow, but it's the real thing.

Four things that make a skill work:

1. A description that says when to use it, and when not to. "Do NOT use for technical audits" stops it firing on the wrong job. A marketer on Reddit gave the same tip: keep your skills' descriptions from overlapping.
2. One job per skill. A skill that audits, writes and reports does all three badly.
3. A fixed output. Name the files and the sections. Then every run looks the same and you can compare them.
4. A check at the end. Mine checks that every number has a source and a date.

What Claude SEO skills can't do

• Links and outreach. A skill can find prospects. It can't build relationships.
• Strategy. It finds SEO problems, not where your customers drop out. It doesn't know your business, your margins or what your boss wants.
• Anything without data. No Search Console and no crawl means guesses about traffic and rankings.
• Find what's not in the HTML. In my hard set, most runs missed what only shows in HTTP headers. JavaScript-heavy sites are worse.
• Write like you. For that you need a voice file, not an SEO skill.

Which SEO job do you still do by hand that you'd hand to a skill tomorrow?

Common Questions