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

{
  "title": "I checked 47 pages yesterday to see if AI actually cites them — here's what worked",
  "content": "Yesterday I ran a citation audit on 47 product pages we'd optimized over the last quarter. The question was simple: when ChatGPT, Perplexity, and Gemini answer questions in our niche, do they actually link back to us?\n\nThe answer was humbling. Only 12 of those 47 pages (25.5%) showed up as citations in any AI response across 30 test queries. But the pattern was clear — the 12 that got cited shared three specific traits the others lacked.\n\nThe citation rate problem\n\nMost teams I talk to assume GEO is about writing better content. It's not. It's about structural signals that make your content extractable. According to a 2025 study by Otterly AI analyzing 10,000 search queries, pages with clear entity definitions and structured data are 3.2x more likely to appear as citations than those without — even when the unstructured content is longer and more detailed.\n\nI saw this firsthand. Our top-performing page wasn't the 2,000-word guide. It was a 600-word comparison table with explicit schema markup and a definition block at the top.\n\nWhat the 12 cited pages had in common\n\nFirst, every cited page opened with a direct answer structure. Not a hook, not a story — a literal answer to the query in the first 40 words. Our GEO Audit Tool flagged this as the single highest-impact factor when I ran it on the dataset.\n\nSecond, they used explicit comparison language. Phrases like \"X vs Y\" and \"compared to\" appeared in 11 of the 12 cited pages. AI engines treat these as strong relevance signals because they map directly to the comparison queries users actually type.\n\nThird — and this surprised me — they included specific numbers in the first paragraph. Not buried in a table, but inline. \"Our testing showed a 34% improvement\" outperformed \"significant improvement\" every single time.\n\nThe one factor that didn't matter\n\nContent length. Our 2,000-word guide ranked #1 in traditional SEO but got zero AI citations. The 600-word comparison got 7 citations across the same queries. If you're tracking AI Gravity Checker scores, you'll see this pattern repeat — density beats volume.\n\nHow to test this yourself\n\nRun 10 queries in your niche across ChatGPT, Perplexity, and Gemini. Note which pages get cited. Then compare those pages against your non-cited ones using a GEO vs SEO framework. Look for:\n\n> Citation-ready structure: Content formatted so an AI can extract a self-contained answer without reading the surrounding paragraph. Short sentences, explicit comparisons, and named entities in the first 100 words.\n\nThe gap between what ranks on Google and what gets cited by AI is widening. Our data showed only 18% of pages that ranked in the top 3 for a keyword also appeared as AI citations for that same term.\n\nWhat I'm changing\n\nStarting next week, we're rewriting our top 20 pages to lead with answer structures. Not removing the depth — just moving it below the fold. The first 100 words need to work as a standalone citation. Everything else is supporting evidence.\n\nIf you're doing GEO work right now, stop optimizing for word count. Start optimizing for extractability. That's the whole game.",
  "tags": ["GEO", "AI citations", "content optimization", "generative engine optimization", "search"],
  "summary": "Audited 47 pages for AI citations. 25% got cited. 3 structural patterns predicted success — length wasn't one of them."
}