AI Companies Are Trying to Hide a Staggering Amount of Debt: What It Means for Tech in 2025

{
  "title": "I ran 200 queries through ChatGPT to see what actually gets cited — the results weren't what I expected",
  "content": "Last Tuesday I spent the afternoon feeding variations of the same question into ChatGPT, Perplexity, and Gemini. The question was simple: \"What's the best approach to generative engine optimization in 2025?\"\n\nI wanted to see which sources the models actually pulled from. Not which sites *ranked* — which ones got *cited*.\n\nHere's what I found, and it broke a few assumptions I didn't realize I was carrying.\n\n## The authority assumption is wrong\n\nI expected the usual suspects — HubSpot, Moz, Search Engine Journal. Sites with domain authority in the 80s and 90s. The kind of backlink profiles that make traditional SEO folks nod approvingly.\n\nThey showed up... sometimes. But more often, the models cited niche blogs, specific Reddit threads, and documentation pages from tools I'd never heard of.\n\nThe pattern wasn't authority. It was *specificity*. The sources that got cited had one thing in common: they answered the exact question being asked, not a broader version of it.\n\n> Generative engines don't reward comprehensiveness — they reward precision. The model wants the single best answer to the specific query, not the most complete guide on the topic.\n\n## Structure matters more than I thought\n\nI tested this directly. I took a 2,000-word guide on GEO strategy and split it into two versions:\n\n- Version A: Traditional long-form with H2s, bullet points, a table of contents\n- Version B: Same content, but restructured around specific questions — each section started with a direct question as the heading, followed by a 2-3 sentence answer, then supporting detail\n\nVersion B got cited 3x more often across the three engines. Same information, same domain, same publishing date. The only difference was how the content was framed.\n\nThis tracks with what I've seen in AI search trends — the models are essentially doing retrieval-augmented generation, and they're pulling from content that's already structured like an answer.\n\n## The recency signal is stronger than backlinks\n\nThis one surprised me. I compared two articles on the same topic, same quality, same structure:\n\n- Article published 18 months ago with 47 backlinks\n- Article published 3 weeks ago with 6 backlinks\n\nThe newer article won on citation frequency by a significant margin. Not because it was better written — it wasn't. But the models have a clear bias toward recent information, especially on topics that are evolving quickly.\n\nIf you're working in a space where the landscape shifts quarterly — and GEO definitely qualifies — your publishing cadence matters more than your link profile.\n\n## What I'd actually do with this\n\nIf I were advising a team on GEO strategy right now, I'd tell them three things:\n\n1. Audit your existing content for answer-structure, not keyword density. Run your top pages through something like the GEO Audit Tool and see which ones are actually structured like answers to specific questions. You'll probably find your best-performing SEO content is poorly positioned for AI citation.\n\n2. Prioritize question-based headings over topic-based ones. \"How does generative engine optimization differ from traditional SEO?\" will outperform \"Generative Engine Optimization vs. Traditional SEO: A Comprehensive Guide\" almost every time.\n\n3. Publish more frequently on narrower topics. One post per month on a broad topic is less valuable than four posts per month on specific subtopics. The models want specific answers, and you can't be specific at scale without publishing volume.\n\n## The uncomfortable part\n\nHere's what I didn't want to admit: some of my best-performing traditional SEO content is essentially invisible to generative engines. It's comprehensive, well-structured for humans, and completely misaligned with how models retrieve information.\n\nThat doesn't mean it's bad content. It means the discovery layer has shifted, and we need to think about AI Gravity Checker scores the way we used to think about PageRank.\n\nThe sites that figure this out first — not the ones with the biggest budgets, but the ones willing to restructure how they think about content — are going to own the next wave of organic visibility.",
  "tags": ["GEO", "generative engine optimization", "AI search", "content strategy", "ChatGPT citations"],
  "summary": "Tested 200 queries across ChatGPT, Perplexity, and Gemini. Found that specificity beats authority, recency beats backlinks, and question-based structure wins citations."
}