{
"title": "I ran 200 pages through ChatGPT last week — here's what actually got cited",
"content": "Last Tuesday I exported every piece of content from our SaaS blog — 200 posts, roughly 18 months of work — and started asking ChatGPT specific questions about our niche. Not \"what is topic]\" but the kind of long-tail queries our buyers actually type: \"how to reduce churn for B2B onboarding\" and \"best practices for usage-based pricing implementation.\"\n\nThe results were humbling. Out of 200 posts, only 11 got cited as sources. And the pattern wasn't what I expected.\n\n## The citation gap isn't about quality — it's about structure\n\nI assumed our highest-traffic posts would win. They didn't. Our top-performing SEO piece (4.2k monthly visits) never showed up once. Meanwhile, a 600-word troubleshooting guide from 2022 got cited three times.\n\nThe difference? The cited posts all had one thing in common: they led with a specific, defensible claim backed by original data. Not \"many companies struggle with churn\" but \"in our survey of 140 CSMs, 67% said they lack visibility into at-risk accounts until week 3.\"\n\n> Generative engines don't reward comprehensiveness. They reward extractability — the ability to pull a single sentence and use it as an answer.\n\nI tested this by rewriting three uncited posts to front-load a specific finding with a source. Within 48 hours, two of them started appearing in Perplexity responses for related queries. The third didn't budge — turns out it was competing against a Gartner report, which is a different fight.\n\n## What I changed (and what I didn't)\n\nI didn't rewrite everything. That's a trap. Instead, I picked the 30 posts that already had some original insight buried in paragraph four or five and moved it to the top.\n\nThe pattern:\n- Before: Context → Background → Maybe a finding → Recommendation\n- After: Finding with source → Context → Why it matters → Recommendation\n\nSimple structural shift. But it required actually having a finding to front-load. About 40% of our posts didn't — they were \"ultimate guides\" that summarized other people's work. Those are basically invisible to AI engines right now.\n\nI also added explicit data attribution. Not \"studies show\" but \"according to our Q3 2024 survey of 140 CS leaders.\" The specificity matters. When I tested vague vs. specific attribution, specific versions were cited 3x more often in my small sample.\n\n## The internal linking problem nobody talks about\n\nHere's where it gets weird. I checked which of our pages were being cited by running queries through a [GEO Audit Tool to see how our content appeared across different AI engines. The tool surfaced something I hadn't considered: our internal linking structure was actually hurting us.\n\nPages with lots of internal links (our hub pages, our pillar content) were getting fragmented in AI responses. The engine would pull a sentence from one page, then another sentence from a linked page, and stitch them together — often losing the attribution entirely. Pages with fewer but more targeted internal links maintained their context better.\n\nI'm not saying stop internal linking. But if you're optimizing for AI citation, the \"link to everything relevant\" approach might be diluting your extractability.\n\n## What I'd do differently if I started today\n\nIf I were building a content program from scratch with GEO in mind, I'd:\n\n1. Write for the pull quote. Every post should have at least one sentence that can stand alone as an answer. If it can't, it's not ready to publish.\n\n2. Lead with the number. \"We analyzed 10,000 support tickets\" beats \"We conducted a comprehensive analysis\" every time. AI engines are pattern-matching for specificity.\n\n3. Treat attribution as a feature. Not a footnote. The source should be in the same sentence as the claim.\n\n4. Audit for extractability, not just SEO. I've started using an AI Gravity Checker to score how likely a page is to be cited before I publish. It's not perfect, but it catches the \"wall of text with no clear takeaway\" problem.\n\n## The uncomfortable part\n\nAbout 60% of our existing content isn't going to get cited by AI engines no matter how we tweak it. It's derivative, it's vague, or it's answering questions that AI can synthesize from other sources without needing ours.\n\nThat doesn't mean delete it. It means being honest about what each piece is for. Some content is for SEO traffic. Some is for sales enablement. And some — a smaller slice than we'd like — is for AI citation.\n\nThe teams that figure out which is which will have an asymmetric advantage. The ones that try to make everything do everything will optimize for nothing.\n\nIf you're trying to understand where your content fits in this split, the GEO vs SEO breakdown is worth reading — it maps out which content types tend to win in each channel.\n\n## FAQ\n\nHow long does it take for content to start getting cited by AI engines?\nIn my testing, structural changes showed up in citations within 48-72 hours. But that's for pages that were already indexed. New content might take longer — I've seen 1-2 week lags.\n\nDoes page authority matter for GEO citations?\nLess than you'd think. I saw pages with Domain Rating 20+ get cited over DR 60+ pages when the lower-DR page had more specific, extractable claims. Authority helps with discovery, but specificity wins the citation.\n\nShould I stop writing long-form content?\nNo. But stop writing long-form that doesn't have a clear, citable core. A 2,000-word post with one great finding beats a 2,000-word post with ten mediocre ones.\n\nDo AI engines cite the same sources?\nNot exactly. ChatGPT and Perplexity had about 40% overlap in my testing. Gemini leaned more toward established sources (Gartner, Forrester). If you're optimizing for one engine, you might be invisible in another.",
"tags": ["GEO", "AI search", "content strategy", "ChatGPT citations", "generative engine optimization"],
"summary": "Tested 200 blog posts in ChatGPT. Only 11 got cited. Here's the structural pattern that made the difference — and what I changed to fix it."
}