What to run and when

What we are solving

I do not ship audit tooling with this atlas, and I am not going to. The good tools already exist, and my own rewrite of one would be a worse copy that nobody maintains, me included.

What I did not have was the map. Which tool answers which question, in what order, and what to skip once the symptom says where to look.

That map is the /atlas: commands at the bottom of the table, and none of them is an audit. /atlas:start calls the third-party audits and orders their findings along the route. The rest do what no audit covers: a content plan, a weekly slice of numbers, a draft read against the style guide.

Every third-party row below is a package sitting on my own machine. The one exception is a paid service: I read the licence column off the source repository, and for that service off what the money actually buys.

Steps

Tool	What it does	When to reach for it	Where it comes from	Licence
seo-audit	Technical SEO audit — crawlability, indexing, on-page, page speed	Search returns nothing for the site, or a page fell out of the index	Skill, coreyhaines31/marketingskills	MIT
geo	Entry point over the GEO skills, routes a URL to the right one	You want one command and do not yet know which check you need	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-audit	Full GEO audit through parallel subagents, one composite score	A baseline before you change anything, and the same run a month later	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-citability	Scores passages for how likely a model is to quote them	The pages are written and assistants still cite somebody else	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-crawlers	Reads robots.txt, meta tags and headers into an access map per agent	Before any citability work, and after any change at the edge	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-llmstxt	Validates an existing llms.txt or drafts one from the site structure	You are publishing the file, or its links have gone stale	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-schema	Detects and validates JSON-LD, generates the missing blocks	The entity markup is absent, or a redesign may have broken it	Skill, zubair-trabzada/geo-seo-claude	MIT
geo-report	Collects the audit results into one document for a reader	Somebody else has to act on the findings	Skill, zubair-trabzada/geo-seo-claude	MIT
Playwright MCP	Drives a real browser in your own profile, with your live sessions	Search Console, a store console, any panel with no usable API	npm @playwright/mcp, microsoft/playwright-mcp	Apache-2.0
gws	One authenticated command line over the Google Workspace APIs	Keeping the weekly series in a Sheet, pulling an export out of Drive	npm @googleworkspace/cli, googleworkspace/cli	Apache-2.0
Semrush	Domain Overview over Google data: positions, backlinks, an authority score, an AI Search panel	You want positions and links for a domain you cannot measure from the inside	Hosted service, semrush.com	Paid subscription, not a skill
/atlas:start	Walks a project through the route, calling the audit skills at the right points	You have a live URL and no idea which problem to fix first	This repo, plugin/	MIT
/atlas:content-plan	Turns a project description and a query export into clusters, one page each	You need to know what to write, and in which order	This repo, plugin/	MIT
/atlas:report	A weekly slice of Search Console and Analytics: what moved, what to do	The weekly snapshot, once the property is verified and a token exists	This repo, plugin/	MIT
/atlas:voice	Brings a draft in line with STYLE.md before the linter sees it	Before opening a pull request against this repository	This repo, plugin/	MIT

Where these skills come from, and where their licence lives

The skills CLI takes an owner and a repo name, so installing the third-party ones is two lines:

npx skills add -g coreyhaines31/marketingskills
npx skills add -g zubair-trabzada/geo-seo-claude

The -g flag is not optional. Without it the skill lands in the current project, and one directory over it is gone. The package itself sits in ~/.agents/skills/ with a symlink in ~/.claude/skills/, so look in both.

The licence does not come down with them. An installed SKILL.md carries no licence line, no author and no repository name, and provenance survives in one place only — the lock file the installer wrote:

cat ~/.agents/.skill-lock.json

Take the repository name out of there. Then go and read the licence in the repository itself.

Which of them to run first when nothing is in the index

seo-audit. It walks fetch, block and canonical in one pass, which is faster than checking seven things by hand, and much louder when one of them is broken.

On the GEO side the order matters more than the choice. Run geo-crawlers before geo-citability, every time. Access decides whether the writing gets read at all.

I once ran them the other way round, and I got a cheerful citability report about a page no agent had fetched in weeks.

What a composite score is actually good for

geo-audit for the baseline, a focused skill for the fix: a composite number moves when anything moves, so on its own it tells you little.

Run the same audit again after the work and read the difference, because that difference is the part I actually look at.

When a browser is the only way into a panel

Playwright MCP earns its place where there is no API at all: a panel behind a login, a console with no export button.

Drive your own profile, so the authenticated sessions are already there — a clean profile shows you a sign-in screen and nothing else.

What gws covers, and what it does not

gws --help

The list it prints comes from your own installed build, and that list is the answer, not what a guide says. It covers Workspace.

Search Console and Analytics are not in it. The failure below is me finding that out the slow way.

How the four /atlas: commands get installed

Two lines in Claude Code — add this repository as a marketplace, then install the plugin:

/plugin marketplace add Smolevich/pet-project-atlas
/plugin install atlas

The auditing stays with the third-party skills above, and that is deliberate. /atlas:start checks nothing itself: it decides which third-party skill to run and in what order to read what they return.

/atlas:report takes one of two routes into Search Console, and says which before it pulls anything.

The cheap one is the browser you are already logged into: Playwright MCP opens the Performance report in your own profile, with no Cloud project and no token. The other is the REST API, which costs a project with the API enabled, a token scoped to webmasters.readonly and a quota project. Without the last one Search Console answers 403.

Take the browser for a report you run by hand on a Monday. Take the API when it has to run unattended.

The other three do their own work, but none of it is auditing: /atlas:content-plan builds clusters, /atlas:report calls the Search Console API, /atlas:voice reads a draft against STYLE.md. That is cheaper to keep alive than a second audit engine.

What did not work

• Reaching for gws to pull Search Console numbers. Its help offers an <api>:<version> form for unlisted APIs. I tried searchconsole:v1 and analyticsdata:v1beta; both came back as unknown services. Those numbers come out of their own APIs, or the console by hand.
• Looking for the licence inside the installed skill. The licence column above had to be filled from somewhere, so I went digging through the installed folders. They hold a SKILL.md and scripts: no licence, no author, no repository name. The terms live in the source repository, whose name survives only in the lock file.
• Treating an audit score as the work. A composite number moves when anything moves. What actually changed my behaviour was one line of a report. The agent that got a 403. The passage nobody could quote.
• Running citability checks while the crawler was blocked. The report scored my prose happily. No agent had read the page in weeks. I was grading a document nobody had fetched.
• Driving a clean browser profile for panels behind a login. Every run started at a sign-in screen. The profile with the live session is the only one that reaches the panel.
• Believing this list is stable. Third-party skills change owners, licences and behaviour between updates. Check the repository on the day you install, and again before you copy anything out of it.

Verify

• cat ~/.agents/.skill-lock.json and read the sourceUrl for each skill you installed. That is the repository whose licence applies to you.
• Open that repository and read its LICENSE file. If there is none, the terms are unknown. Unknown means do not redistribute.
• Run gws --help and find your service in the printed list. An absent service is absent, whatever a guide says.
• Run seo-audit against a page you already repaired. If the report still flags it, the fix did not reach production.
• Run geo-crawlers from a network that is not yours. An access map built from inside your own network hides the edge rules.

What the paid row does and does not measure is its own page: what a paid rank tracker measures.

Which of these to run for the symptom you actually have is the Tools index.