Geo Content Engineering

View saved version →

---
name: geo-content-engineering
description: >-
  Create or rebuild web content so generative engines (ChatGPT, Gemini,
  Perplexity, Google AI Overviews) extract, cite, and mention it. This is the
  discipline called GEO (Generative Engine Optimization), also known as AI
  visibility, LLM SEO, or answer engine optimization. Use this skill whenever the
  user wants to write, restructure, or optimize any web page, article, guide,
  landing page, pillar page, or FAQ for AI search visibility, citations, or brand
  mentions. Trigger it when they mention GEO, generative engine optimization, AI
  visibility, LLM SEO, answer engine optimization, semantic chunking, getting
  cited by ChatGPT or Perplexity, or making content "AI-readable"; and also when
  they ask softer versions like "help me write content that AI will recommend,"
  "why doesn't ChatGPT mention my brand," or "how do I show up in AI answers."
  Covers both new content creation and rebuilding existing pages, plus the
  Share-of-Mentions measurement loop.
---

# GEO Content Engineering

Build web content that generative engines will pull into their answers. The
goal is not ranking in a list of blue links. The goal is being the block of text
a model extracts, the source it cites, and the brand it names when a user asks a
question.

## Why this changes how you write

Generative systems do not read a page top to bottom. They split it into semantic
blocks (chunks), turn each block into a vector embedding, and for any given
prompt they retrieve the single block whose embedding is closest to the prompt.
That retrieved block is what gets quoted, summarized, or cited. One sharp,
self-contained paragraph often beats an authoritative domain, because the model
scores the block, not the whole site.

Everything in this skill follows from that mechanic. Full explanation in
`references/mechanics.md`. Read it once if the user is new to GEO.

Three layers have to be right, in this order. A perfect text on a page a bot
cannot read is invisible. So gate access first, then structure, then reputation.

1. **Access.** The bot must reach and render the content. See
   `references/crawler-access.md`.
2. **Content and structure.** The text must be chunkable, quotable, complete,
   and current. This is the bulk of the work below.
3. **Off-page grounding.** External mentions confirm the claims. See
   `references/offpage-mentions.md`.

## Pick the mode

**Mode A: New content.** The user is creating a page that does not exist yet.
Go straight to Build.

**Mode B: Rebuild an existing page.** The user has a live URL that underperforms
in AI answers. Do not touch the text until you have a baseline, or you will not
be able to prove the rebuild worked. Run these first:

1. Measure a baseline Share of Mentions for the target topic cluster before any
   edit. See `references/measurement.md`.
2. Check the access gate on the live URL. See `references/crawler-access.md`. If
   the bot cannot render the page, fix that first. Rewriting an unreadable page
   is wasted effort.

Then both modes converge on Build.

## Build

Use the page skeleton in `assets/page-structure.md` as the frame. Fill it by
applying six content rules. Each rule has a full reference file; read the file
when you need depth, not before.

1. **Semantic chunking.** One block, one claim. Answer first, then evidence.
   Every block has to make sense lifted out of the page on its own. Name the
   entity instead of using pronouns ("the Model X battery," not "it"). Phrase H2
   and H3 headings as the questions users actually ask.
   → `references/chunking.md`

2. **Dense, plain language.** Active voice, short sentences, no filler. Every
   wasted word blurs the block's embedding and makes it lose retrieval to a
   tighter competing block. When there is no blanket answer, state the condition
   directly ("if you need A, choose option 1") instead of hedging.
   → `references/language.md`

3. **Citation-worthiness.** Give the model something it can

Preview of the first 4,000 characters. Read full instructions →

Referenced files: 25 — file references are metadata, not package contents.