web-optimization

name	web-optimization
description	Audit or generate web content optimized for traditional search (SEO), AI generative answer engines like ChatGPT/Perplexity/Google AI Overviews (GEO), and answer engines / featured snippets / voice (AEO). Use when the user asks to improve a page's ranking or AI-citability, run an SEO/GEO/AEO audit of a URL or file, add JSON-LD schema, create an llms.txt, or write new content that is search- and LLM-friendly.
license	MIT
metadata	{"authors":"Staksoft (https://www.staksoft.com)","source":"https://www.staksoft.com/insights/seo/beyond-keywords-the-definitive-guide-to-generative-engine-optimization-geo-in-2026"}

Web Optimization (SEO · GEO · AEO)

This skill treats SEO, GEO, and AEO as three lenses over one shared body of web-optimization practice. It supports two workflows: Audit an existing page, and Generate new optimized content.

The three lenses (what each optimizes for)

Lens	Optimizes for	Load file
SEO	Google/Bing ranking (blue links)	references/seo.md
GEO	LLM synthesis (ChatGPT, Perplexity, AI Overviews)	references/geo.md
AEO	Answer engines, featured snippets, voice	references/aeo.md

Shared support files:

• references/schema.md — JSON-LD patterns (load whenever structured data is involved).
• references/scoring.md — the rubric and priority weighting used by both workflows.

Progressive disclosure: Do NOT read every reference up front. Read only the lens file(s) relevant to the request, plus scoring.md for audits and schema.md when structured data is in play.

Workflow A — Audit

Use when given a URL or a local content file and asked to evaluate/improve it.

1. Acquire the content.
   • URL: prefer python scripts/audit.py <url> for deterministic, objective checks (returns JSON). Also fetch the rendered content to judge quality.
   • Local file (.html/.md): read it directly; run audit.py --file <path>.
2. Apply the three lenses. Read references/seo.md, geo.md, aeo.md and evaluate the content against each. Fold in the JSON facts from audit.py.
3. Score & prioritize using references/scoring.md (P0 blocking → P3 nice-to-have).
4. Emit the report using assets/audit-report-template.md: every finding = issue · lens · impact · concrete fix. Lead with the prioritized action list.

Workflow B — Generate

Use when given a topic/brief + target keyword or user intent and asked to produce new content.

1. Clarify intent — target query, audience, and primary lens emphasis if any (default: optimize for all three).
2. Draft applying all three lenses: answer-first structure (AEO), keyword/ intent coverage and clean heading hierarchy (SEO), high information density with concrete measurable claims (GEO).
3. Attach structured data — pick the right JSON-LD from assets/schema-templates/ per references/schema.md (Article + FAQPage are the common pair).
4. Produce an llms.txt entry from assets/llms-txt-template.md.
5. Output the content + JSON-LD + llms.txt entry + a short "why this is optimized" rationale mapping choices back to the three lenses.

Workflow C — Fix (interactive step-by-step)

Use when the user says "fix [URL]", "fix the issues", or "apply the fixes" after an audit — or wants to be guided through improvements interactively.

1. Run or recall the audit. If an audit was just completed, use those findings. Otherwise run Workflow A first silently, then begin fixing.
2. Load references/fix-playbook.md. It maps every finding type to the minimum question(s) to ask and what to generate.
3. Work P0 → P1 → P2 → P3, one fix per turn:
   • State: what you're fixing and why (one line).
   • Ask: the minimum question(s) needed (often zero — derive from fetched content).
   • Generate: the complete, ready-to-paste output (JSON-LD block, file content, rewritten copy, HTML snippet, diff of tag changes).
   • Confirm: "Fix N of M done — move to the next one? (or say skip / stop)"
4. Maintain a session checklist (✅ done / ⏭ skipped / ⬜ pending) and show it at the top of each turn so the user always knows where they are.
5. On "stop" or "done": show the full checklist summary and remind the user of any skipped items.

Key principle: Never ask the user to write anything manually. Every output should be copy-paste ready. If you need a number or URL you can't derive, ask for that one thing — then generate everything else yourself.

Quick rules of thumb

• Information density beats word count. Replace hedges ("might be fast") with measurable claims ("sub-50ms p95 latency"). LLMs cite specifics.
• Answer the question in the first sentence, then elaborate — this serves snippets, voice, and LLM extraction simultaneously.
• Structure for two readers: a human operator and a natural-language parser. Tables, lists, and JSON-LD help the parser without hurting the human.
• Authority is distributed. Off-domain mentions (GitHub, Reddit, citations) matter for GEO trust, not just on-page factors.