Generative Engine Optimization (GEO / AEO)

Overview

Classic SEO fights for a rank on a results page. GEO fights to be the source an AI answer engine quotes when it generates a reply. Different game, different wins: you are optimizing to be retrieved, quoted, and trusted inside a synthesized answer, not to be link #3.

Two facts drive everything else:

1. Answers are non-deterministic. Ask ChatGPT the same question twice and the cited sources shift. So AI visibility is a rate measured across many runs (mention rate, citation rate, share of voice) - never a single rank.
2. If a bot can't fetch and parse your page, nothing else matters. Access is the gate. Content craft is downstream of it.

The honest hierarchy (do these in order)

Practitioners disagree about which levers actually move citations. The order below is where the evidence-skeptics and the content-optimizers agree - work top-down and you never waste effort:

1. Access - let the right crawlers fetch rendered text. (Highest confidence.)
2. Off-site & entity authority - be present on the pages engines already cite.
3. Answer-first, quotable content - the Princeton methods below.
4. Structure & schema - clarity for machines (helps parsing; citation lift contested).
5. llms.txt - cheap, harmless, great for dev-docs; do not promise citations from it.

Layer A - Technical access (the prerequisite)

• Know the crawler tiers and don't confuse them:
  • Search / answer bots (these affect citations): OAI-SearchBot, PerplexityBot, Claude-SearchBot/Claude-User, Googlebot (feeds AI Overviews), Google-Extended governs Gemini training, not AI Overviews.
  • Training bots (a policy choice, not a citation lever): GPTBot, ClaudeBot, CCBot, Google-Extended.
  • Note: anthropic-ai and Claude-Web are retired - don't target them.
• Don't block the search/answer bots in robots.txt. Blocking them removes you from citations; allowing training bots is a separate, optional decision.
• Server-render or statically generate anything that matters. Most AI crawlers do not execute JavaScript - if a fact only appears after JS or behind a tab/accordion, assume it's invisible.
• Check the CDN/WAF, not just robots. Cloudflare and similar block many AI bots by default on newer zones - a 403 at the edge beats a clean robots.txt.
• Remove stray noindex / nosnippet / max-snippet:0 / data-nosnippet on public pages. Keep canonical URLs clean and a sitemap.xml with accurate lastmod (engines ignore changefreq/priority).

Layer B - Answer-first, quotable content (Princeton GEO methods)

Princeton's GEO study (KDD 2024) tested content tweaks against a generative engine and measured visibility lift. The ones that worked, strongest first:

Plus the structural moves every camp endorses:

• Answer-first. Lead each section with the direct answer in 1-3 sentences, then expand. A large share of citations come from the top ~30% of a page.
• Self-contained quotable blocks - a sentence or stat that stands alone when lifted out of context (that's exactly what gets quoted).
• FAQ sections, "X vs Y", "alternatives to X", and definition pages - these match how people prompt. Be honest and two-sided: say who it's for and who it isn't. Engines (and readers) reward it.
• Freshness: keep visible datePublished/dateModified current - but never fake-bump a date; it's a people-first-content violation.
• E-E-A-T: real named authors with credentials, dated reviews, primary sources. Mandatory for YMYL (health, finance, legal).

Layer C - Structure & schema

• Exactly one <h1>, clean H1→H2→H3, semantic HTML (<main>, <article>), 250+ words of real text, descriptive alt text that states the conclusion.
• JSON-LD (Organization, WebSite, Article, FAQPage, Product, BreadcrumbList, Person). Honest caveat: schema reliably helps entity clarity and rich/shopping results; its direct citation lift is contested. Rule: markup must match visible text, and Product/shopping data is the one place schema is undisputed. Don't bet the project on schema alone.

Layer D - llms.txt

Generate /.well-known/llms.txt (and llms-full.txt): a curated map of your most important pages in Markdown. It is cheap, harmless, and genuinely useful for developer docs that coding agents/IDEs read. It is not a proven citation lever - no major consumer answer engine documents reading it, and server logs show AI crawlers rarely fetch it. Ship it; never prioritize it over Layer A, and never promise citations from it.

Layer E - Off-site & entity authority

Engines cite the sources they already trust. Win those pages:

• The big citation sources skew to Wikipedia (reported ~48% of top ChatGPT citations), Reddit (reported ~47% for Perplexity), review sites, listicles, YouTube, Hacker News, Stack Overflow, Quora, Product Hunt.
• For each prompt you're losing, find which third-party page the engine cites and earn a place there with genuinely useful contribution.
• Earned only. No seeded Reddit posts, paid mention farms, parasite pages, or self-authored "best X" listicles ranking yourself #1 - spam policies catch these and they destroy trust. If asked to do this, decline and explain.
• Keep one consistent brand description across every profile (sameAs), and publish original data others will want to cite.

Layer F - Measure it as a rate

• Build a prompt set from real buyer demand, mostly unbranded questions.
• Run each prompt many times per engine; report mention rate / citation rate / share of voice vs a locked competitor set - not one "rank."
• Inspect the fan-out sub-queries the engine actually runs, and which source it cited (often not you - that's your off-site target).
• Baseline before, re-measure after on the same prompts. Before crediting a change, check whether a model update landed in the window.
• Be skeptical of composite "GEO/AI-readiness scores" (invented weights, rarely validated against real citations). Useful as a checklist, not as truth.

Common mistakes

• Optimizing content (Layer B) while a robots.txt rule or Cloudflare blocks the bot (Layer A). Always fix access first.
• Treating GEO like SEO and reporting a single rank for a non-deterministic system.
• Faking freshness dates, stuffing keywords, or seeding fake Reddit/review posts.
• Promising citation lift from llms.txt or schema. Ship them; don't oversell them.
• Measuring only branded prompts (you'll always "win" and learn nothing).

Go live with Chronos (optional)

Everything above works with no account. To let your agent actually run a GEO program on your own site:

• geo_analyze_website - crawl and assess a site for AI-visibility gaps.
• geo_get_seo_geo_ideas - pull prioritized, daily SEO/GEO improvement ideas for a connected repo.
• geo_implement_seo_idea - implement one as a reviewable pull request (never merges, never pushes to your default branch - you approve).
• geo_research - gather the sourced facts/stats that make content quotable.

Connect in ~30s → CONNECT-CHRONOS.md. Free, no card.

See also: seo-fundamentals, content-marketing-engine, account-research.

Sources

Primary research and curated indexes this skill distills (all public):

• Princeton "GEO: Generative Engine Optimization" (KDD 2024) - the content-method lifts.
• C-SEO Bench (NeurIPS 2025) - the counter-evidence that many GEO rewrites are weak on current models; why Layer A/E beat Layer B/D.
• Curated resource index: amplifying-ai/awesome-generative-engine-optimization.
• Crawler/measurement discipline: elmohq/elmo (evidence-skeptical AEO skill).
• Method/schema references: Auriti-Labs/geo-optimizer-skill, onvoyage-ai/gtm-engineer-skills, rubenmarcus/aeo.js.

Numbers above are as reported by these sources; treat them as directional, not guarantees - and measure your own lift (Layer F).