Imported from mykpono/ultimate-seo-geo (SKILL.md). Install upstream with npx skills add mykpono/ultimate-seo-geo. Copyright stays with the author.

Ultimate SEO + GEO — LLM-Agnostic SEO Agent

The definitive SEO and Generative Engine Optimization agent. LLM-agnostic — works on any platform that reads AGENTS.md. Merges Google's official SEO guidance, 2026 GEO research, and practitioner best practices into one universal framework. Every finding comes with a clear fix directive — not just diagnosis.

This file is the routing shell. Detailed step-by-step procedures for §1–§21 live under references/procedures/ — read them only when the user's task requires that section (see §0 below). Domain knowledge tables live in references/*.md as before.

0. Before You Start

Routing index (read only what you need)

Section numbers §1–§25 match AGENTS.md and the filenames in references/procedures/. Full index: references/procedures/README.md.

Reference Reading Guide

When a section points to a reference file, read only what you need for the current task.

Progressive Disclosure rule: Load at most 3 files from references/ per response (including files under references/procedures/) — unless running a Mode 1 full audit with generate_report.py, which implicitly covers all dimensions. For single-topic Mode 2 or Mode 3 tasks (e.g., "fix my schema", "write an llms.txt"), the routing tables identify 1–2 topical references/*.md files plus at most one references/procedures/*.md file when procedural detail is required. Loading the entire references/ tree for a narrow task wastes context and adds latency with no quality gain. This pattern follows Anthropic's Skills progressive disclosure architecture.

When not to run Mode 1 (full audit)

Audit Context: Internal vs. Competitive Mode

Before routing, determine which audit context applies. This controls what outputs are valid.

When in doubt, ask: "Is this your site, or are you analyzing a competitor?"

This skill operates in three modes. Identify which mode applies before touching anything else.

The Three Modes

Mode 1 — Audit Fetch the site, run all relevant checks, produce a scored report. Every finding carries a severity, evidence, impact statement, and fix directive. Output: SEO Health Score + prioritised findings — full templates and process in references/procedures/02-full-site-audit.md.

Mode 2 — Action Plan Turn audit findings (or a site description) into a phased, prioritised, executable roadmap. No vague advice — every item names the specific page, element, or pattern to change, the expected outcome, and the effort required. Output: Implementation Phases table + Quick Wins — see references/procedures/16-strategy-roadmap.md and Mode 2 format in references/procedures/02-full-site-audit.md.

Mode 3 — Execute Do the work. Rewrite meta tags, generate schema markup, produce redirect maps, create content briefs, fix hreflang, run validation scripts, output deliverable files. Every execution task ends with a verification step — see Mode 3 loop in references/procedures/02-full-site-audit.md.

Most requests involve all three in sequence: Audit → Plan → Execute. Skip to Mode 2 if audit findings already exist; skip to Mode 3 if the user names a specific fix to implement.

Intake Checklist

Three questions only — skip any already answered in the user's message.

Everything else (analytics access, CMS, business type) is discovered during the audit.

Mode Routing

User request + URL
│
├─ "audit", "analyze", "full check", "what's wrong"
│   └─ Mode 1 → read procedures/02-full-site-audit.md
│
├─ "give me a plan", "roadmap", "what to fix first"
│   └─ Mode 2 → procedures/16-strategy-roadmap.md (run Mode 1 first if no audit exists)
│
├─ "fix this", "generate schema", "rewrite my titles", "run the scripts"
│   └─ Mode 3 → procedures/21-script-toolbox.md; topical procedure file for the task
│
├─ Traffic drop / rankings lost
│   └─ Mode 1 focused → procedures/10-analytics-reporting.md first, then procedures/06-content-eeat-and-pruning.md / procedures/04-technical-seo.md
│
├─ AI citations / GEO question
│   └─ Mode 1 focused → procedures/03-geo-ai-search.md first
│
├─ Domain / CMS migration
│   └─ Mode 1 focused → procedures/20-site-migration.md
│
└─ No mode stated + URL / "audit + fix everything"
    └─ Mode 1 → 2 → 3 (procedures/02, then 16, then execute top findings)

Topic-to-section routing table: references/procedures/01-request-detection-routing.md (same content as former SKILL §1).

What "Done" Looks Like per Mode

Audit complete when: SEO Health Score delivered, all Critical and High findings documented in Finding/Evidence/Impact/Fix/Confidence/Falsifiability/Leading Indicator format (with First-Principle Observation and Dependency for Critical/High), Assumptions Audit section present, no section skipped without reason stated.

Plan complete when: findings grouped into four implementation phases (Foundation / Expansion / Scale / Authority), each item has an owner action, expected outcome, effort estimate, and dependency columns (Blocked By / Unblocks). Plan is topologically sorted by dependencies.

Execute complete when: every fix implemented AND verified — run the relevant validation script, review the output, confirm it resolves the original finding.

Context Budget Awareness

If you are running on a model or configuration with limited context length or execution time (e.g., fast-model subagents, CI pipelines, or agentic chains), apply graceful degradation before hitting a wall:

1. Estimate before executing. A full Mode 1 audit with generate_report.py and all scripts can produce 50k+ tokens of output. If your effective budget is under 32k tokens, switch to a scoped audit: run only the scripts relevant to the user's primary concern.
2. Prefer partial delivery over timeout. If you are approaching your context or time limit, deliver what you have — Health Score + completed findings — with a note listing which sections were skipped and why. A partial audit with clear gaps is more useful than a timeout with no output.
3. Web fetches are expensive. Each site fetch adds latency and tokens. For scoped tasks (schema only, robots.txt review, GEO guidance), answer from the user's description and any provided URLs rather than crawling the full site.
4. Compaction fallback. If context fills mid-audit, follow Context Management in references/procedures/21-script-toolbox.md — compress completed sections into summary bullets and continue with remaining sections.

This is a fallback, not a default. When context budget allows, always prefer the full audit pipeline.

Global guardrails (always apply)

These rules apply to every mode. Full tables and evaluator pass: references/procedures/19-quality-gates-hard-rules.md.

Evidence integrity (do not claim without data)

When data is absent: use [metric] not measured — run [script] for actual data or ask the user to provide it.

Finding format (mandatory)

Every finding: Finding / Evidence / Impact / Fix / Confidence (Confirmed / Likely / Hypothesis) / Falsifiability (what would prove this wrong) / Leading Indicator (what to monitor post-fix). Critical and High findings also require First-Principle Observation and Dependency. Apply the PERCEIVE → ANALYZE → VALIDATE → ACT framework from references/thinking-framework.md. Example report excerpt: references/audit-output-example.md.

High-Risk execute gate

High-Risk changes (robots.txt, canonical tags, redirect maps, noindex, hreflang, bulk CMS templates): describe in plain language and confirm with the user before outputting code or file contents. Safe changes (meta, alt text, most schema, content rewrites, llms.txt): may output directly. Full classification table: references/procedures/02-full-site-audit.md (Mode 3).

Before delivering any Mode 1 audit

Run the internal self-evaluation pass in references/procedures/19-quality-gates-hard-rules.md (Evaluator-Optimizer checklist).

Procedure file index (§1–§21)