EmptyOS GEO Audit

Wrapper around the upstream geo-seo-claude skill that points it at the right place: the publish app's built static output for a specific site. Output lives in the vault so future audits can diff against past ones.

When to use

User intent	Example phrasing
One-shot audit of the default site	"GEO audit", "audit my site for AI search"
Audit a specific site	"audit the portfolio site", "geo check on binbian.net"
Pre-publish gate	"is this post going to be cited by ChatGPT?"
Periodic check	"weekly GEO check" (run from a scheduler job)

Don't use when the user wants traditional SEO (keyword density, backlinks) — geo-seo-claude upstream handles both, but the trigger phrases above are about AI-engine citability specifically. The skill itself will still produce a hybrid report.

Prerequisites

This skill wraps geo-seo-claude (https://github.com/zubair-trabzada/geo-seo-claude). Install it once via your normal Claude Code skill mechanism (clone into .claude/skills/ or however you manage personal skills). If it's not installed when this skill runs, the wrapper falls back to reporting the install steps and writing a minimal placeholder audit so the audit folder still has a dated entry to diff against next time.

The publish app must have built the target site at least once. Output lives at data/apps/publish/sites/<site-id>/site/.

What it does

1. Resolve the target site. Read data/apps/publish/sites.json. If the user named a site, match by id or by name substring. Otherwise pick the site whose id == "default", or the first site if no default.

2. Confirm the built output exists. Check data/apps/publish/sites/<id>/site/index.html. If missing, prompt the user to run a build first via the publish app (POST /publish/api/build) and stop.

3. Invoke the upstream geo-seo-claude skill against data/apps/publish/sites/<id>/site/. Pass any site metadata that helps it score brand authority correctly (site domain, author, social_links — from sites.json).

4. Write the audit report to the vault:

   {vault}/30_Resources/EmptyOS/publish/audits/YYYY-MM-DD-geo-<site-id>.md
   

   Use the existing daily date format (no time component) so a same-day re-run overwrites — audits are point-in-time snapshots, not streams.

5. Include a ## Diff vs previous section when an earlier audit exists for the same site. Highlight: score change, new issues, resolved issues. Keep this section short (≤8 bullets) — the user reads it first.

6. Report back in chat with: site audited, audit file path, score (if upstream returns one), top 3 highest-leverage fixes. Don't dump the whole report into chat — link to the vault file.

Output convention

{vault}/30_Resources/EmptyOS/publish/audits/
├── 2026-05-23-geo-default.md
├── 2026-05-23-geo-portfolio.md
├── 2026-05-16-geo-default.md       # last week's
└── 2026-04-30-geo-default.md       # last month's

Frontmatter on each audit:

---
tags: [audit, geo, publish]
site_id: default
site_domain: binbian.net
audited: 2026-05-23
upstream_skill: geo-seo-claude
score: 78           # optional, only if upstream returned one
---

This lets future audits be queried via VaultIndex and the vault graph picks them up under the publish project.

What this skill is NOT

• Not a runner for the publish build itself. It audits the output of a build; it never triggers one. Build via the publish app's own UI / API.
• Not an auto-fixer. It reports findings; the user (or a follow-up skill / agent) decides which fixes to apply. This matches EmptyOS's "with you, not for you" north star — schema markup changes to your published site shouldn't happen without you seeing them.
• Not a recurring auto-runner. If the user wants periodic audits, schedule it via the scheduler app pointing at this skill's invocation. Don't bake cron into the skill itself.

Cost note

geo-seo-claude makes LLM calls (it grades pages, generates schema markup, classifies content for AI-engine compatibility). Cost scales with the site's page count. Surface this when the user invokes the skill against a site with >50 pages — give them a rough cost estimate and let them confirm before running.

Cross-references

• apps/public/standard/publish/ — owns the build that this skill audits. The [provides.publish] site config in data/apps/publish/sites.json is the source of truth.
• .claude/skills/eos-screenshot/SKILL.md — sibling skill for visual capture of the same sites; share the privacy/branding posture (.eos-personal + .eos-branding).
• Upstream geo-seo-claude — does the actual GEO scoring + recommendation work. This skill is purely the EmptyOS-context binding around it.