# Research first, write second: a page AI engines can cite | Agility Writer Blog > 180 facts in a ledger before one sentence was written. The two-pass research method behind a Wikipedia-style page, and why most AI content skips it. URL: https://agilitywriter.ai/blog/write-for-ai-citation-research-first/ Last-Modified: 2026-09-26 Author: Adam Yong Case Studies # Research first, write second: a page AI engines can cite 180 facts in a ledger before one sentence was written. The two-pass research method behind a Wikipedia-style page, and why most AI content skips it. Adam Yong · Founder & CEO, Agility Writer · September 26, 2026 · 6 min read ![AI search interface showing an answer with numbered citations linking back to source pages](/images/featured/stylized-ai-search-interface-mockup-with-citation-.webp) 180 facts. That’s what sat in a JSON file before I let anything write a sentence of aiseocourse.wiki [https://aiseocourse.wiki/ →](https://aiseocourse.wiki/) . Each fact had a value, a source URL, the date I checked it, and a confidence flag: primary, secondary, or unverified. 156 of them made it onto the two finished pages. 24 didn’t, because I couldn’t confirm them. Most AI-written content does this backwards. Prompt, draft, then go hunting for a citation to hang on each claim. The draft decides what the facts are, and the facts get bent to fit. That’s why so much of it reads fine and cites nothing you’d trust. There’s data behind the complaint now. OtterlyAI’s AI Citations Report [https://otterly.ai/blog/the-ai-citations-report-2026/ →](https://otterly.ai/blog/the-ai-citations-report-2026/) , published February 2026 off more than a million citations across ChatGPT, Perplexity and Google AI Overviews, found that chunked, quotable, schema-tagged pages get 3 to 5 times more citations than pages that aren’t. The report doesn’t publish the methodology behind that multiplier, so treat it as a direction rather than a precise number. This post is the method I used instead. It’s not complicated. It’s just slow in the right place. ## Step one: build the ledger, not the outline The first session had one rule: no prose. Fetch every primary source, extract every number, date, name, price and quote into a ledger, and stop. For a page about AI SEO courses, “every primary source” meant the seller’s own pages, the Skool listings, the founder’s site, conference speaker pages, podcast transcripts, and Google’s own blog posts for the AI Overviews timeline. Then 13 search queries across the course names, the category terms, and the founder’s name, fetching the sales page of every course that turned up. The ledger schema was small: id, entity, fact, value, date\_verified, source\_url, confidence. That last field is the one people skip. “Primary” means I read it on the seller’s own page. “Secondary” means a reviewer or a news site said it. “Unverified” means I saw it on a third-party site and couldn’t find where they got it. Unverified facts don’t get published. They get a line in a notes file. ## Step two: run it twice, with different tools I ran the research pass in Claude Code first. Then I ran a separate pass in ChatGPT’s deep research mode, with a prompt that asked for the same ledger shape and no article prose. Then Claude Code read the ChatGPT dossier as a secondary source and checked each fact against the primary URL it cited before importing it. This caught things. One pass found a launch date of 17 August 2026 on the official site. The other found the founder saying 10 August on his own blog. Both went in the ledger as a recorded conflict, and the page says both [https://aiseocourse.wiki/ai-seo-rainmakers →](https://aiseocourse.wiki/ai-seo-rainmakers) , with sources. A single-pass draft would have picked one and looked confident. It also caught the competitor. A site called aiseo-course.wiki, with a hyphen, launched the same day as the contest keyword was revealed, and several of its numbers appeared in my ChatGPT dossier as if they were primary. They weren’t. They were that site’s numbers, and two of them didn’t match the Skool listing on the day I checked. ## Step three: outline from fact IDs Before writing, every section of the outline got a list of fact IDs. “Pricing model: F012, F013, F014, F015, F016 to F021.” If a section had fewer than three IDs, it got flagged as thin, and the flag said what research would fill it. That’s the point where you find out whether you have a page or a pamphlet. My category article had six criticism subsections because the ledger had the material. My entity article’s first draft had nine, because the ledger had that much too, and I cut it to four after publishing because a reference page isn’t an investigation. The ledger tells you what you can say. Editorial judgement tells you what you should. ## Step four: write, with every sentence carrying its ID The writing pass was the fast part. Each factual sentence got a footnote marker that maps back to a fact ID. The reference list, the sources page, and a public data.json file were all generated from the same ledger, so they can’t drift apart. Change a fact in the ledger, rebuild, and the footnote, the reference entry and the JSON record all update together. Two pages, 4,800 and 4,500 words, 43 and 16 references. Wikipedia layout, because that’s the shape retrieval systems already trust: definition first, infobox, numbered sources, neutral voice. This is the thinking behind the citation work we’re building into Agility Writer now. Not shipped yet, so don’t take this as a feature announcement. The design principle is the same one above: a claim is only worth publishing if a machine can follow it to a source in one hop, so the tooling checks where a page already gets cited and where the gaps are, rather than generating more confident-sounding prose. ## What this costs, honestly The research passes took most of a day between the two tools. The writing took two hours. The on-page pass, schema, sitemap, link checks, took another two. Call it a full working day for two pages. That’s absurd for a content mill. It’s cheap for a page that has to be the source of record on its topic. And the ledger is reusable. The directory site I built next, aiseocourse.directory [https://aiseocourse.directory/ →](https://aiseocourse.directory/) , pulls its comparison cards from the same facts. A third article for another of my sites used the same course records, and so did the free course I wrote for beginners [https://aiseocourse.net/ →](https://aiseocourse.net/) and a scored guide covering 45 programs [https://script.google.com/macros/s/AKfycbyu2jIT8Z_wQnd8GZDytGznMwzR_L0isgRrzH1XLj7K2LbWkXcaxiNMjV8V_d5HEGzZpQ/exec →](https://script.google.com/macros/s/AKfycbyu2jIT8Z_wQnd8GZDytGznMwzR_L0isgRrzH1XLj7K2LbWkXcaxiNMjV8V_d5HEGzZpQ/exec) . The research cost was paid once. ## The tell in most AI content Read any AI-drafted article and look at where the citations sit. If they cluster at the end of paragraphs as “sources” without a specific claim attached, the draft came first. If each number has its own marker, and the marker takes you to a page where that number actually appears, the research came first. This is also why citation and ranking are drifting apart. Ahrefs puts 28% of ChatGPT’s most-cited pages at zero organic visibility [https://ahrefs.com/blog/ai-seo-statistics →](https://ahrefs.com/blog/ai-seo-statistics) in Google. More than a quarter of what the model quotes is invisible to the search playbook most teams are still running. AI engines are getting better at telling the difference. Not perfectly, and not yet everywhere. But the pages I see cited in AI Overviews and Perplexity answers are increasingly the ones where a machine can follow a fact to its source in one hop. Build the ledger. Then write. The finished pages are at aiseocourse.wiki if you want to see what the output of this looks like. Every footnote goes somewhere real [https://aiseocourse.wiki/sources →](https://aiseocourse.wiki/sources) . ![Adam Yong](/images/misc/professional-headshot-of-adam-yong-founder-seo-ent.webp) Adam Yong Founder & CEO, Agility Writer SEO professional and consultant since 2011. 15+ years of practical SEO experience baked into Agility Writer's methodology. ✓ SEO professional since 2011 · Founder & CEO of Agility Writer ## More from the Case Studies category ![Traffic chart showing growth from zero to 50,000+ monthly organic visits over six months](/images/featured/case-study-from-zero-to-50000-organic-visits-in-si.webp) ### From Zero to 50,000+ Organic Visits in Six Months — A Marketing Agency Case Study A content marketing agency used Topical Map Helper, Bulk Mode, and G-Smart Optimizer to launch a single-cluster authority site that hit 50,000+ monthly organic visits in under six months. Here's the workflow, the numbers, and the lessons. [From Zero to 50,000+ Organic Visits in Six Months — A Marketing Agency Case Study →](/blog/from-zero-to-50000-organic-visits-in-six-months/) ![](/images/cta/modern-3d-abstract-saas-illustration-with-soft-blu.webp) Available now, start in under 2 minutes ## Ready to Ship Long-Form Content That Ranks? Start your $1 trial. 5 credits, real articles, cancel anytime. Try Agility Writer for $1 [https://app.agilitywriter.ai/register →](https://app.agilitywriter.ai/register) See Pricing [/pricing/ →](/pricing/) Cancel anytime Credits roll over Trustpilot 4.7/5