geo_ai_search_optimization v1.0.0

name: "geo_ai_search_optimization" description: "Get cited by AI search answers (ChatGPT, Perplexity, Google AI Overviews, Gemini). Use when someone asks about generative engine optimization (GEO), AI search visibility, getting AI answers to cite a site, llms.txt, AI crawler access in robots.txt, or measuring AI brand mentions and citations."

GEO: Getting Cited by AI Search

Purpose

Make a site's content the kind AI answer engines actually cite: citable structure, evidence-dense writing, bot access that lets citation crawlers in, and a measurement loop that proves what changed. Grounded in the Princeton GEO study (Aggarwal et al., KDD 2024, arXiv 2311.09735), which is the only large controlled experiment on this topic. Everything else is practitioner observation, and this skill says which is which.

Workflow

1. Measure the baseline first. Pick 20 to 40 real questions buyers ask (no brand name in the prompt). Run each on ChatGPT, Perplexity, Gemini, and Google (AI Overviews or AI Mode) in fresh sessions. For each run log: prompt, engine, date, brand mentioned (yes/no), brand cited (yes/no + URL), position among cited sources, and how the brand is described. Feed the log to bin/geo_audit.py for a baseline report. See references/measurement.md. Do not change anything until this baseline exists.

1. Open the crawl layer. In robots.txt, explicitly allow the citation/search crawlers and user fetchers: GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, anthropic-ai, Claude-Web, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User, Google-Extended, GoogleOther, Applebot-Extended, Amazonbot, FacebookBot. Blocking training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot) is a separate policy choice and does NOT remove you from AI answers. Serve the full content in the initial HTML response. AI crawlers do not execute JavaScript, so client-rendered pages are invisible to them. Then add an llms.txt at the domain root (format in references/bot-access.md). Be honest about it: llms.txt is cheap insurance for developer tooling, not a ranking factor. No major answer engine is known to consume it.

1. Rewrite the money pages with the Princeton methods. Apply the three proven methods first, in this order (details and worked rewrites in references/princeton-geo-methods.md and references/rewrite-examples.md):

• Cite sources: name the institution, study, or report behind each claim. "Research shows" cites nothing.
• Add quotations: a direct, attributable quote from a named expert with their title.
• Add statistics: specific numbers with sources. "9.3 days down to 3.7 days" beats "much faster". Then: lead with the direct answer in the first 30% of the page, use short paragraphs and question-shaped headings, keep the tone authoritative and the language plain. Never stuff keywords; the study found it hurts.

1. Tune per engine. ChatGPT Search citations skew to Bing-indexed pages, statistical claims with named sources, long-form how-tos with step structure, and primary vendor docs. Perplexity favors recent clean-structured posts even from low-authority domains, comparison and review content, Reddit and Hacker News threads with specific reasoning, and primary research over syndication. Google AI Overviews cite from the Googlebot index, so standard index health applies. If Perplexity cites a site but ChatGPT does not, check Bing indexation first.

1. Re-measure on the same prompts. Repeat the baseline prompt set after 4 to 12 weeks (live-search engines can reflect changes within weeks; training-data effects take quarters). Same prompts, same engines, same logging. Report the delta in mention rate, citation rate, and share of voice. One check is a snapshot; the trend is the result.

Output Contract

• A baseline audit report (from bin/geo_audit.py): mention rate, citation rate, average citation position, share of voice, and which competitors and domains are cited most.
• A robots.txt delta and an llms.txt draft that matches the llmstxt.org grammar (H1 title, blockquote summary, H2 sections, - Title: description links, ## Optional last).
• Rewritten page content with the three Princeton methods applied: every factual claim either cites a named source or is framed as opinion; stats are specific and sourced; keyword density unchanged or lower.
• A re-measurement plan: same prompt set, same engines, scheduled 4 to 12 weeks out, with the delta defined before the rerun.

Operating Rules

1. No confident unverified claims. Every factual claim in output must be verified via search or a primary source. The Princeton numbers are reported lifts from one 2024 study, never guarantees. Say so when they come up.
2. Never invent numbers, dates, quotes, or sources. Unverifiable gets cut or framed as opinion. No exceptions.
3. Measure before and after. Never ship GEO rewrites without a baseline and a re-measurement plan. Without the loop there is no optimization, just edits.
4. Bot access is binary. A page is either fully readable in the first HTML response or invisible. No "mostly server-rendered" and no relying on llms.txt as a substitute for crawlable content.
5. Small brands win narrow questions. Don't rewrite everything. Win the specific buyer questions first; broad category pages follow once the trend moves.