claude-seo Review: We Ran Two Full Audits on Our Own Site
We ran claude-seo v2.4.0 twice against trustedskills.dev, once with a headless browser and once without. It found every real bug we know about. It also gave two different health scores on the same day, and one run put a made-up star rating into copy-paste schema.
We assess AI agent skills by installing them, running them against real work, and reading the source. We don't accept payment for reviews.
How we know
We ran this skill against real work. The runs are listed below, including how it installed — or why it didn't.
Both runs used the same command against the same live site on the same day. We changed one thing between them: whether the headless browser (Playwright with Chromium) was available. We then checked the biggest findings against the live site ourselves. The full reports and transcripts are kept on file.
Quick verdict
Recommendedclaude-seo is the most thorough SEO audit we have run in Claude Code. Across two runs on our own site it found every indexing bug we already knew about, plus a few we didn't, and wrote the results to disk as a report, an action plan and (with Playwright) a PDF. The catch is consistency. The same site scored 43 and then 36 on the same day, and one run skipped a data source it had. So treat the findings list as the product and the score as noise, and check any schema before you paste it. Each full audit used about US$7–9 of model time and took 20–25 minutes.
What we ran it on
The task
/claude-seo:seo-audit https://trustedskills.dev with Playwright and Chromium installed. The site is a Next.js registry of about 50,000 pages. We gave no other instructions.
What happened
Finished in 19 min 15 s. Model usage came to US$7.22 at API list prices. It launched 10 specialist sub-agents (technical, content, schema, sitemap, visual, performance, GEO, agentic, SXO, backlinks), and all 10 wrote their report. It saved a Markdown report, an action plan, an HTML report, a PDF and 10 screenshots to disk.
It classed the site as a "SaaS / developer directory" with programmatic pages, and crawled a stratified sample of 423 of the 50,262 sitemap URLs. It scored the site 43/100.
Its top findings were real. We checked each one on the live site: about 46,700 inner pages declare the homepage as their canonical; paginated listings declare a canonical that redirects to a 404; the install commands point at an npm scope with no packages in it; and 395 sitemap URLs return 404. It also pulled Search Console data and showed that impressions fell from about 270 a day in late June to 8–23 a day from 18 September.
One defect: the ready-to-paste JSON-LD for skill pages contained an aggregateRating of 4.6 from 12 reviews. The site has no ratings. A note under the block said to include it only once real ratings exist. Pasted as written, it is invented review markup.
Environment: Windows 11, Claude Code 2.1.282 run headless (`claude -p`), orchestrator on Opus 5.5 with sub-agents on Opus 5.5 and Sonnet 5. claude-seo loaded as a plugin from a fresh clone of v2.4.0; its Python runtime was set up with the bundled `claude-seo setup`. The Claude Code profile had Google Search Console, PageSpeed Insights and Ahrefs MCP servers connected, so both runs could reach paid-tier data.
The task
The same command against the same site, 2 hours later, with Chromium not installed. We wanted to see what the audit loses without a headless browser.
What happened
Finished in 25 min 32 s. Model usage came to US$8.85 at API list prices. The same 10 sub-agents all wrote their report. The screenshot step failed with Error: playwright required, and the visual report said so at the top. It then gave a markup-only review and marked every item as not confirmed visually. There were no screenshots and no PDF.
It classed the site as a "Directory / registry (developer tools)". It crawled 441 URLs and scored the site 36/100, 7 points lower than the first run.
The critical findings matched the first run: homepage canonicals, 404 pagination canonicals, the empty npm scope, and the 395 dead sitemap URLs. It also found things the first run did not. The live review page said "We tested" in its meta description while the page itself said the tool was not run. The meta descriptions contained 70 "workflows workflows" errors. Its schema blocks used no fake rating, and it said "Do not fabricate ratings."
It also got some things wrong. It did not call Search Console, which was connected, and then reported "No Google Search Console data for this property." Its homepage lab test gave LCP 4.5 s and ranked that High. The first run measured 1.7–2.4 s on the same page.
Environment: Windows 11, Claude Code 2.1.282 run headless (`claude -p`), orchestrator on Opus 5.5 with sub-agents on Opus 5.5 and Sonnet 5. claude-seo loaded as a plugin from a fresh clone of v2.4.0; its Python runtime was set up with the bundled `claude-seo setup`. The Claude Code profile had Google Search Console, PageSpeed Insights and Ahrefs MCP servers connected, so both runs could reach paid-tier data. Chromium was left out of this run's data directory, so the browser could not start.
What we set out to learn
Our earlier source review left five questions open. This is what the two runs showed.
| Question | What happened |
|---|---|
| Does site-type detection work? | Yes, near enough. It classed our registry as "SaaS / developer directory" in one run and "Directory / registry" in the other. Both runs used the right checks for a programmatic site, such as sampling the sitemap and looking for scaled, templated content. |
| Is the schema output valid, and about the right page? | All eight JSON-LD blocks across both runs parsed as valid JSON and described the page they were written for. One block in the first run had an invented 4.6-star rating, so the answer is "valid, not always honest". |
| How much does Playwright change the result? | On our server-rendered site, only the visual part changed. Both runs fetched the same raw HTML (mode_used: raw), so the technical, content and schema findings were the same. Without Chromium it lost the screenshots, the PDF, the mobile fold and tap-target checks, and one overflow bug. |
| Does the parallel run finish? | Yes. 10 sub-agents in each run, all 10 reported, with no truncation. The runs took 19 min 15 s and 25 min 32 s. |
| Is the Critical → Low order right? | The top was right: the canonical bugs and the npm scope come first in both runs. The middle was noisier. Both ranked missing security headers High, which is not a ranking fix. One run ranked 395 dead URLs (0.8% of the sitemap) Critical, and a single lab LCP reading High. |
What's included in v2.4.0
The plugin has grown a lot since its first release. v2.4.0, released 23 September 2026, ships 26 skills and 19 agents. It also includes a Python runtime for rendering, crawling and reports, and optional extensions for Google APIs, Ahrefs, Semrush, DataForSEO and others.
The full audit (/seo-audit) launched 10 specialists on our site: technical, content, schema, sitemap, visual, performance, GEO, agentic, SXO and backlinks. Other commands cover single pages, local SEO, e-commerce, hreflang, topic clusters, drift monitoring and content briefs.
None of the external data sources are required. Our runs had Search Console, PageSpeed Insights and Ahrefs connected, so the backlink and traffic findings above depend on that.
How to install
Claude Code plugin (recommended)
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seo
The first run sets up a Python environment and downloads Chromium for the visual checks. On our Windows machine that setup took 13 min 44 s and finished without errors.
Manual install
The README tells you to clone the repo and read install.sh or install.ps1 before you run it, rather than piping a remote script into a shell. We agree.
The plugin works without any API keys. For traffic, indexing and backlink data, connect Search Console, PageSpeed Insights or Ahrefs. The setup wizard is /seo google setup.
What it does well
🎯 It found the bugs that matter, in both runs
Both runs put the site-wide canonical bug first, and both found the 404 pagination canonicals, the empty npm scope behind the install commands, and 395 dead sitemap URLs. We confirmed all four on the live site. For a 50,000-page site, those are the fixes an experienced SEO would make first, and the tool ranked them first.
🧑🤝🧑 Real parallel specialists
Each run launched 10 specialist sub-agents, and all 10 finished and wrote a report. The specialists checked each other's work, too. For example, the schema agent re-fetched a page when the crawl data under-reported its markup. The parallel run did not truncate or stall in either run.
📁 Reports you can hand over
Every run wrote FULL-AUDIT-REPORT.md, ACTION-PLAN.md, an HTML report, raw data and one findings file per specialist into a folder. The Playwright run also made a PDF and 10 desktop and mobile screenshots. You don't copy results out of the chat.
🔌 Uses your connected data when it remembers to
In the first run it queried Search Console, URL Inspection, PageSpeed Insights and Ahrefs on its own. That turned a static audit into a diagnosis: it tied a fall in impressions to the canonical bug and to pages Google had crawled but not indexed.
🤖 Checks for AI search and agents
Both runs tested each AI crawler (GPTBot, ClaudeBot, PerplexityBot and others) for a 200 response, checked llms.txt, and scored how easy the site is for AI agents to use. Few other SEO skills cover this ground.
Where it stops
🎲 The score moves between runs
The same site scored 43/100 at 04:31 UTC and 36/100 at 06:58 UTC. The category scores moved by up to 13 points. Homepage LCP was 1.7–2.4 s in one run and 4.5 s in the other. Only the 4.5 s run ranked it as a High issue. Don't track progress by the health score from single runs.
🕳️ It can skip a data source and say the data doesn't exist
Search Console was connected for both runs. The first run used it. The second run didn't call it, and its report said "No Google Search Console data for this property". If your audit says a source is missing, check that it really is.
⚠️ Check schema before you paste it
The first run's ready-to-paste skill-page JSON-LD contained a 4.6-star aggregateRating from 12 reviews that don't exist. Its URLs had a trailing slash that redirects. A note under the block said not to use placeholder ratings, but the block itself was ready to copy. Invented review markup breaks Google's structured-data rules.
💸 A full audit isn't cheap
Each run used about US$7–9 of model time at API list prices (US$7.22 and US$8.85) and took 19–26 minutes on a 50,000-page site. On a Claude subscription, that comes out of your usage limits. The README compares the tool against agency fees, but it doesn't give a per-audit cost.
🌐 Without a browser, you lose the visual audit only
Our site is server-rendered, so both runs read the same HTML. Without Chromium, the audit lost the screenshots, the mobile fold check, tap-target sizes, the PDF and a 58 px horizontal overflow on mobile. The technical, content and schema findings were the same. On a JavaScript-rendered site the gap is probably larger. We didn't test that.
🔒 Claude Code only
It ships as a Claude Code plugin with a Python runtime, 26 skills and 19 agents. The SKILL.md files hold useful knowledge, but the scripts, agents and launcher need Claude Code.
Who should use it
🏢 Agencies doing onboarding audits
One command gives a report folder you can hand over, with an action plan sorted by severity. With Search Console and Ahrefs connected, it explained why traffic had fallen, not just what was misconfigured.
🧱 Large programmatic sites
It knew to sample a 50,000-URL sitemap instead of crawling it, and it pulled out template-level bugs that affect tens of thousands of pages at once.
🤖 Teams targeting AI search
The GEO and agent-readiness checks go further than any other SEO skill we have run.
Probably not for you if…
You want a stable number to track week to week: the health score moved 7 points between two runs on the same day. You want a quick check. A narrower skill like the marketingskills seo-audit found the same two indexing bugs in about nine minutes. You're not on Claude Code.
Frequently asked questions
Does claude-seo actually crawl my website?
Yes. On our site it read the sitemap index (50,262 URLs), then fetched a stratified sample of 423 to 441 pages. It also fetched the homepage, robots.txt, llms.txt and redirect chains live. With Playwright installed, it also opened pages in a headless browser for screenshots and layout checks.
Do I need Playwright?
For the technical, content and schema checks on a server-rendered site, no. Our two runs found the same issues. You need it for screenshots, the PDF report, and mobile layout checks such as the fold and tap-target sizes. Without it, the visual report says it could not capture anything, and it marks its findings as unconfirmed.
Do I need Ahrefs or Search Console?
No, but they change what you learn. With Search Console connected, the first run could explain why traffic fell (impressions down from about 270 to under 25 a day, with key pages crawled but not indexed). Without that data, you get a list of what's wrong but not how much it's costing you.
How long does a full audit take, and what does it cost?
On a 50,000-page site, 19 and 26 minutes, using US$7.22 and US$8.85 of model time at API list prices. A small site should cost less. We have only measured a large one.
Can I trust the health score?
Not as a precise number. Two runs on the same site on the same day scored 43 and 36. The list of Critical findings was the same both times, and it was correct. Use the findings, not the score.
Can I use it with Cursor or Codex?
Not as shipped. It's a Claude Code plugin with its own agents, scripts and launcher. The SKILL.md files contain useful knowledge you could adapt, but the audit workflow we tested runs only in Claude Code.