How Can Publishers Test a Claude SEO or GEO Skill Found on GitHub?

A useful test of a Claude SEO or GEO skill is not whether it produces a polished audit. It is whether, on a task whose answer you already know, it identifies the right passages, explains what it actually accessed and proposes an edit you can review. Start with one small publisher task rather than a site-wide score.

Identify what you would install

Find the repository and pin the version you intend to test: record a release or commit identifier, the installation method and the Claude product in which you plan to run it. Read the README, skill instructions, manifest and any scripts. Is the package only instructions for Claude, does it execute code, or does it need tools available in your Claude session? The answer changes what you must review and what the skill can plausibly observe.

Do not infer the contents of one package from another package’s landing page. For example, Claude SEO Skill advertises a GitHub ZIP that users add to Claude and then invoke in a conversation. A separate assessment of seo-geo-claude-skills describes installation through skills.sh or Claude Code’s plugin marketplace. Those are advertised installation paths for different offerings, not evidence that either setup worked in your environment.

Even inventories of the same package can differ. That assessment describes 20 skills, five commands and 24 diagnostic scripts, while an undated tool listing describes 20 skills and nine commands. The assessment says five commands were consolidated from an earlier nine; without a version on the listing, inspect the commands and scripts in the version you would install. Record that inventory instead of treating either count as a timeless specification.

Before installation, inspect executable files, requested permissions, network access and any destinations to which page content or other data could be sent. Use a disposable environment and material you are comfortable sharing with the tools available to the run. Do not give a site-audit helper access to unpublished drafts merely to find out whether it can read a public FAQ.

Give it a discrepancy you can check

Set up a hypothetical test on a site you control. The page at publisher.example/guides/returns says, “You can return an eligible item within 30 days of delivery.” Its linked FAQ says, “You can return an eligible item within 14 days of delivery.” These are deliberately conflicting sample passages, not a reported finding on a live site.

Ask the skill: “Review this returns guide and its linked FAQ. Identify any conflicting return periods, quote the exact passage and URL for each, and propose a prioritized correction. Say which pages you could not access.” Give it the URLs, not pasted passages, if the feature you are testing is page access. Otherwise, a correct answer could come entirely from text you supplied.

This matters particularly for a crawl claim. Claude SEO Skill advertises a workflow in which Claude fetches the homepage, robots.txt and sitemap.xml, then crawls selected sections. Treat that as a claim to test in the product and mode you use, not as a record of requests made during your run. Save the available tool output showing requested URLs, results and retrieved passages. If you can inspect access records for the controlled site, check for the relevant requests there too. A generated audit that names a page is not, by itself, a fetch record; a request record alone does not show that the skill used the correct passage.

Grade the answer, not the audit’s appearance

The central question is whether the output quotes 30 days from the guide and 14 days from the FAQ with the correct URLs. It should flag the conflict and recommend resolving which policy is authoritative before changing the other page. A usable edit might be: “If the approved policy is 30 days, change the FAQ’s 14-day sentence to 30 days, then check the surrounding eligibility conditions.” If the approved policy is 14 days, the guide needs the corresponding change instead. The skill cannot determine company policy merely by finding two pages.

Note separately whether it labels each passage as retrieved, provided by the user or unverified. If it could not reach the FAQ, an honest inability to assess the discrepancy is better than a confident invented quotation. If it claims to have crawled the site, compare that statement with the available tool output and site access records; keep the conclusion no stronger than those records allow.

Finally, treat a GEO or “citability” score as a prompt for questions: What pages, criteria and observations produced it? Claude SEO Skill lists checks including citation-oriented formatting and entity clarity, and displays an example audit with SEO, GEO and AEO scores. Those advertised checks and displayed scores do not establish that the reviewed pages earned AI citations. Formatting advice can still be a worthwhile editorial suggestion when it helps readers locate a clear answer; it is not a forecast of mentions, referrals or conversions. For the broader distinction between a tool’s report and an observed outcome, see [what AEO and GEO tools can actually measure](What Can AEO and GEO Tools Actually Measure?).

Keep the skill if this run produces traceable, appropriately qualified work for access requirements you accept. Revise its instructions or reject it if it invents findings, obscures where passages came from or needs permissions out of proportion to a two-page review.