Terrence Tao's ChatGPT Conversation about the Jacobian Conjecture Counterexample: What It Means for AI Search in 2025

{
  "title": "I ran my top 50 pages through an AI search audit — here's what actually moved the needle",
  "content": "Last Tuesday I exported every page ranking in my top 50 for branded and mid-funnel terms, then checked which ones actually showed up in ChatGPT and Perplexity answers. The correlation was weaker than I expected.\n\nOnly 18 of those 50 pages appeared in any AI-generated response for the same queries. And 6 of them were pages I'd have ranked dead last for \"AI visibility\" based on traditional metrics.\n\nSo I reverse-engineered what the 18 had in common.\n\n## The pattern I found\n\nIt wasn't word count. It wasn't domain authority. It wasn't even backlink count.\n\nThe pages that AI engines cited shared three structural traits:\n\n1. They opened with a direct answer — not a hook, not a story, just the answer to the question in the first 40 words\n2. They used explicit comparison tables — \"X vs Y\" formatted as actual tables, not buried in paragraphs\n3. They cited specific data points — numbers, percentages, study names, dates\n\n> The pages that AI engines cited shared three structural traits:\n> 1. They opened with a direct answer in the first 40 words\n> 2. They used explicit comparison tables\n> 3. They cited specific data points with sources\n\nPages that ranked on page 1 of Google but missed AI answers? They tended to open with brand storytelling, bury the lead, or use vague language like \"significantly better\" without saying by how much.\n\n## What I changed\n\nI took my 5 worst-performing pages (by AI visibility) and made three edits:\n\n- Moved the core answer to the first sentence\n- Replaced one paragraph of prose with a comparison table\n- Added 2-3 specific data points with source attribution\n\nTwo weeks later, 3 of those 5 pages started appearing in ChatGPT responses for related queries. Not all queries — but enough to notice a pattern.\n\n## The uncomfortable part\n\nHere's what bothered me: my highest-traffic page, a 3,000-word guide that gets 12K visits/month, has never once been cited by an AI engine. Meanwhile, a 600-word FAQ page with a comparison table gets pulled constantly.\n\nI ran a GEO Audit Tool scan on both pages. The long guide scored 34/100 for AI visibility. The FAQ page scored 78. The difference wasn't content quality — it was structural formatting.\n\nIf you're trying to figure out where your content stands, the AI Gravity Checker gives you a quick read on which of your pages are likely to surface in generative answers.\n\n## What I'd actually recommend\n\nStop optimizing for word count. Start optimizing for extractability.\n\nAI engines don't read your page like a human does. They scan for discrete, quotable units. Give them:\n\n- A clear thesis in the first line\n- Structured data (tables, numbered lists)\n- Named sources with dates\n- Short paragraphs (2-3 sentences max)\n\nI wrote up a more detailed breakdown of GEO vs SEO differences if you want the full framework.\n\nBut honestly? Just take your top 10 pages, ask ChatGPT the questions they target, and see if your content shows up. That 10-minute test taught me more than any audit.",
  "tags": ["GEO", "AI search", "content optimization", "ChatGPT visibility", "search strategy"],
  "summary": "Audited 50 pages for AI search visibility. Only 18 showed up. The difference wasn't quality — it was structure: direct answers, tables, cited data."
}