Original research·AI search·~14 min read

What actually gets cited by AI answer engines

I joined roughly 9,500 Bing Copilot citations back to the source code of three sites I own. One finding survived every check: demand decides. Page craft is the cover charge. Demand is the lever.

Paul Takisaki· Published June 21, 2026· First-party data, independently re-analyzed

---

01I set out to answer a dumber question

I didn't set out to prove that demand decides. I set out to answer something more practical: of all the pages I've shipped, which ones do AI answer engines actually cite, and what do those cited pages have in common that the ignored ones don't?

I had the receipts: Bing Webmaster's "AI Performance" exports, page by page and query by query, for three sites I own and operate, deliberately different animals: a live, constantly updating tier list for a popular online game; a large library of college financial-aid pages built from one template; and a set of guides for recording family stories. Three audiences, three content shapes, one set of exports. So I joined every cited URL back to the actual source code that produced it, then had two separate AI systems analyze the same exports without talking to each other.^[1] Where they agreed, I trusted it. Where they fought, I went back to the raw CSV and settled it. This isn't a roundup of best practices. It's what my own traffic did.

02The page that should have been invisible

Here's what surprised me most. My single most-cited page, the live tier list on the game site, pulled 5,179 citations. That one page is 80.9% of that site's citations and more than half of every page-level citation across all three sites. And it does everything the standard advice says will sink you: zero JSON-LD schema, content rendered entirely in client-side JavaScript so none of its ranked entries appears in the static HTML, and an <h1> that is a plain topic label instead of a phrased question.

By the playbook, that page should be invisible. It's the opposite. Bing executes the JavaScript, reads the rendered tier table, and cites it constantly. The plumbing everyone obsesses over (schema, server-side rendering, question headers) turned out to be a floor, not an engine. You need a page that can be read and lifted cleanly. Past that, the plumbing didn't move the needle.

The more useful comparison is the page that didn't make it. Same site, a gimmick page (a random build generator), with the same domain, the same crawler access, the same internal links, even the same schema, got 14 citations while a sibling page on the same site pulled 498. The difference isn't technical, as far as I can tell. It has no answer in it. No stat, no tier, no fact a model can quote. That's the line: cited pages carry an extractable answer; the ignored one is just a toy.

03The thing I actually changed my mind about: depth

The belief I most wanted to keep was that craftsmanship is the lever: write the longer, richer, more thorough page and the machines reward you. The college financial-aid site let me test it cleanly. It builds 751 pages off one identical template, so I could hold everything constant and ask: does writing a deeper page get it cited?

At first glance, yes. Group the pages into thirds by word count and the citation rate climbs: 6.8% to 10.8% to 12.7%. Looks like a dose-response. Write more, win more. Tidy.

It's a mirage. The moment I split the pages by whether the school has real search demand, the depth effect collapses. High-demand schools get cited 46.6% of the time; low-demand ones, 6.2%. And inside that high-demand group, citation rate actually goes down as the pages get longer: 58.3%, then 45.8%, then 36.0%. The deep pages were getting cited because deep pages happened to be the ones I'd written for prominent schools people search.

Citation rate by word-count tertile, same metric and the same scale on both panels (cited %). Pooled, depth looks decisive. Isolate the high-demand pages and it inverts: the longer ones get cited less.

Citation rate by word-count tertile, pooled versus high-demand-only

Word-count tertile	All 751 pages	High-demand pages only
Shortest third	6.8%	58.3%
Middle third	10.8%	45.8%
Longest third	12.7%	36.0%

When I put word count into a regression next to demand, it came out statistically meaningless (odds ratio 1.09, not significant) while demand was the one variable that mattered, more than four times over. The proof is brutal and specific: 84% of my 167 deepest pages got zero citations. One of my deepest pages runs 2,713 words with the full dataset filled in, on a school almost nobody searches, and it has never been cited once, while a thinner page on a school everyone searches gets cited constantly. I wrote a small book nobody asked for.

04What I was sure mattered, and didn't

This is the part that should make any "GEO checklist" vendor nervous. I pulled the on-page features everyone sells (quick-answer blocks, tables, FAQ schema, fresh data, source links, "last verified" dates) and compared how often the cited pages and the uncited pages had each one. The uncited pages score higher on nearly every single feature.

Across 163 cited and 712 uncited pages, the "best practices" don't separate the two. The uncited pile is the same strong template pointed at topics nobody searches.

On-page feature	Cited pages	Uncited pages
Quick-answer block	65.6%	99.0%
Data tables	58.3%	94.9%
FAQ schema	66.3%	95.9%
Fresh / dynamic data	76.1%	94.9%
Source links	81.6%	97.5%
"Last verified" date	83.4%	95.2%

Read that table the right way: these features are a floor, not a selector. Ship below it and you risk being fetched and ignored. Clear it and you're merely eligible. Demand is what chooses. And the convergence with the outside research is hard to ignore: a controlled Ahrefs study of 1,885 pages found adding JSON-LD schema produced no meaningful citation uplift,^[4] and SE Ranking's analysis of nearly 300,000 domains found llms.txt had zero correlation with citations.^[5] My first-party data and their cross-site data point the same direction: the plumbing is hygiene, not the win.

05A few other things the data made me face

Every one of these sites is only capturing about a quarter of the citations available on queries it already shows up for. Even on the queries where my pages already appear, most of the citation pool is going to someone else. Closing that gap on pages I already rank for is often a bigger prize than building new ones.

And speed matters more than I gave it credit for. My fast site started getting cited on the first day of reporting; my slowest site took 64 days to land its first citation. I can't cleanly pull cause from this: the 64 days is measured from the day Bing began reporting, so it tangles a page's age with its indexing speed. But the fast site pings Bing on every publish (IndexNow) and the slow one doesn't, and that's a cheap thing to copy.

06Then I tested two engines I had no data for

Everything above is Bing Copilot. After I'd written it, I ran the same kind of test on Perplexity and ChatGPT Search, 25 queries against the same three sites, and one result stopped me cold.

On Bing, one mid-tier aid page has never been cited, not once. It's one of the hundreds of dead pages in the pile. But Perplexity cited it at position 4 and ChatGPT cited it at position 5, independently, off the exact same template that produces all my Bing zeros. And the kicker: my heaviest Bing page in that set, at 201 citations and indexed clean on Google, was cited by neither of them.

So the two systems are almost the inverse of Bing. Bing rewards the school people actually search; the chat engines reward the template that bothers to spell out the aid tiers when the school's own site is vague about them. The sharpest single number: all four of ChatGPT's citations to that site are pages Bing cited zero times, while my biggest Bing page in that set was cited by neither chat engine.^[2]

07The playbook I actually use now

I turned the findings into a decision stack I run before building anything. The one rule that overrides the rest: Gate 1 selects citations on Bing; Gates 2–4 only keep you eligible. Clearing them earns nothing on its own, because the uncited pages clear them too. If Gate 1 fails, stop. Don't build.

08What to stop worrying about

Each of these correlates with citations, and each got cut as non-causal once I controlled for demand:

• JSON-LD schema as a driver. My top two pages have none, for 5,677 citations between them. Keep shipping schema for traditional SEO. Just don't credit it for the AI win.^[4]
• Server-side rendering as a requirement. My most-cited page is 100% client-side and earns 5,179 citations. Bing renders it. "Bing can read CSR" is true; "you must SSR" is false.
• Raw depth and word count. Odds ratio 1.09, not significant. It only ever correlated because deep pages happened to be the ones written for high-demand schools.
• Question-shaped H2s. Zero percent of my cited templates use them. It's not a thing my cited pages do.
• Splitting hubs into per-entity pages when demand is aggregate, and sitemap priority: one of my pages carried priority 0.9 and earned zero.
• llms.txt as a citation lever. Present on only 3.1% of my cited pages, and SE Ranking found no correlation across ~300,000 domains.^[5]

The flip side, and the reason this very report exists: the highest-leverage content lever is original data. The Princeton GEO paper found that adding real statistics measurably lifts how often generative engines cite you.^[3] So the most citable thing I can publish about getting cited isn't more advice. It's my own numbers.

09How I know this (and what it can't tell you)

• Data sourceBing Webmaster Tools "AI Performance" exports: daily, per-page, and grounding-query reports.
• SitesThree content sites I own and operate: a live game tier list, a single-template library of college financial-aid pages, and a set of family-story guides.
• Window62–180 days per site, through June 17, 2026.
• Volume~9,500 citations by Bing's daily total; 9,019 at the page level (the two exports don't perfectly reconcile).
• MethodEach cited URL joined to its source code, then re-analyzed independently by two AI systems; a 751-page single-template regression isolated depth from demand.
• Cross-engineA 25-query probe on Perplexity and ChatGPT Search. Directional, not a second dataset.

10FAQ

Demand decides. Build for the question people are actually asking, make the answer easy to lift, and stop polishing pages the search box has never heard of. If you want help applying this to a site you own, that's the work I do.

End of report