Thirteen sub-skills make Claude Code the audit console
GEO-SEO Claude Code installs a main skill plus 13 specialized sub-skills for citability, crawler access, llms.txt, brand mentions, platform checks, schema, technical SEO, content, reports, prospects, proposals, and comparisons. Five agent files split a full audit across separate areas before Claude combines the results. That is a useful arrangement for an agency already living in Claude Code because the instructions, scripts, JSON-LD templates, and output formats stay editable. It is much less attractive if your team does its SEO work in a browser dashboard.
The checked-out project was small: our sandbox counted 71 files, about 3,220 lines of source, and 0.6 MB before dependencies. Most of the product is method and orchestration rather than a crawler built from scratch. Python utilities fetch pages, score passages, scan brand mentions, generate llms.txt, and render PDF reports. Claude supplies the judgment and synthesis. That division makes the toolkit easy to inspect, but it also means prompt wording and model behavior can affect what reaches the final report.
What happened when we ran it
Our run installed 50 Python packages in 47 seconds and occupied 207 MB on disk. The build completed successfully in 5 seconds. Those numbers are modest for an AI-adjacent developer tool because the repository does not bundle a model or a web application. The documented installer creates an isolated environment under ~/.claude/skills/geo/.venv, so it avoids changing system Python packages. There is no Dockerfile, which matters if your normal review process expects a disposable image supplied by the project.
Pytest passed all 14 tests in 5 seconds, with 0 failures. Pip-audit also reported 0 known vulnerabilities in the installed dependency set. The checkout had one CI workflow and a tests directory. These results cover commit 5d068e9 in our fresh Debian container with 3 CPUs and 8 GB of RAM. They show that the packaged code builds and its available tests pass; they do not establish that any generated audit is factually correct for a live site.
English-only scoring can underrate translated pages
The scorer assigns 30% of its total to answer-quality patterns that open issue 80 shows are English-specific. Definition, early-answer, and citation checks look for English phrases, so a non-English page can lose much of that section even when its translated structure matches an English page. The same issue says FAQ answers inside dl and dd elements never reach the block scorer. Anyone auditing Spanish, French, or another language should treat the absolute score as unsuitable until those patterns and elements are handled.
Passing 14 tests does not cover every website shape the tool claims to assess. Issue 57 reports that the default requests-based fetcher missed hydrated content and JSON-LD on a JavaScript-heavy site. Issue 81 describes instructions that ask Claude's page tool for response headers it does not return, plus Microdata and RDFa checks that lack raw HTML input. These are painful failures because an audit may produce a confident negative finding instead of stopping with an error. A consultant must compare the report with rendered HTML and raw responses.
The one-command installer writes to a global Claude directory
The installer puts 13 related sub-skills and 5 agents under the user's ~/.claude tree. The README's shortest route pipes install.sh into Bash, then users invoke commands such as /geo audit or /geo quick inside Claude Code. Python, Git, and the Claude Code CLI are required. Open issue 36 asks for project and custom destinations because the fixed global location complicates pinned team setups, side-by-side versions, and CI jobs. That matters more in an agency than it does on one person's laptop.
Our 207 MB environment installed cleanly with Python 3.12, but the stated Python 3.8 requirement is misleading. Issue 69 traces current Pillow, lxml, and Playwright pins to a Python 3.10 floor and shows the installer accepting macOS system Python 3.9 before failing during dependency resolution. The report also says copied skill files can remain after that failure. Use Python 3.10 or newer, inspect the target directories before installation, and remember that uninstalling the skill does not delete ~/.geo-prospects runtime data.
A September 2026 push is active, while correctness work is open
GitHub recorded the last push on September 2, 2026, and listed 10,152 stars with 25 combined issues and pull requests. The repository has an MIT license and no latest GitHub release, so adopters need to pin a commit rather than assuming a release tag defines a tested package. Recent issue activity includes detailed reports about scorer language, missing fetch inputs, installer version checks, and unreliable audit findings. That activity is useful evidence of scrutiny, but the combined open count is not a defect count.
The 47-second install and complete 14-test pass make GEO-SEO Claude Code easy to evaluate without much sunk time. Its command set can give an experienced reviewer a consistent checklist and a serviceable report draft. It should not be the final authority on whether a client has an llms.txt file, schema, headers, or citation-ready content. Pin the commit, audit English sites first, and require a human to reproduce every negative or commercially important claim before the report leaves your team.