Skip to main content

Build log

Does AI cite pages that get straight to the answer? Yes

·

Illustration for the build log "Does AI cite pages that get straight to the answer? Yes"

Yes. Answer-first writing is cited more often by AI engines: Kevin Indig’s analysis of 1.2 million AI answers and 18,012 verified citations found that 44.2% of ChatGPT’s citations come from the first 30% of a page. Our content engine turns that into a blocking rule. The body’s first 160 characters must answer the post’s question, or the draft does not move to review.

What does the gate check?

The gate lives in a shared standard, written on June 15, 2026 and enforced in both of our content skills from July 14. It has 10 machine-checkable findings. Four of them trace to two studies.

CheckBlocks whenEvidence behind it
Answer in first 160 charactersThe opening does not answer the post’s questionIndig: 44.2% of ChatGPT citations from the first 30% of a page
Declarative first paragraphIt opens with “in this post we’ll” or a throat-clearSame
Citeable facts, at least 5Fewer than 5 numbers, dates, paths or named settingsPrinceton GEO paper (KDD 2024)
One pull-quoteable stat sentenceNo sentence carries subject, number and source togetherSame
Question-shaped headings where naturalNot blocking; a preferenceRetrieval quotes self-contained spans
Keyword placed onceMissing, or placed verbatim twiceOur own July 15 finding that double placement deformed prose

The Princeton paper tested 9 content strategies across 10,000 queries and found that adding statistics, expert quotations or inline sources each lifted citation visibility by roughly 30 to 41% on Perplexity. Keyword stuffing did worse than baseline.

Why a sentence with a number and a source?

Engines retrieve spans, not pages. A span that already contains the claim, the figure and where it came from can be quoted as-is. A span that needs the paragraph around it to make sense cannot. So the gate does not just count numbers. It requires at least one sentence an engine could lift verbatim and still be right.

The gate also bans one thing that looks like a fact and is not: a bare commit hash. It identifies nothing to a reader outside our repo, so it does not count toward the 5 facts and blocks the draft if it appears in prose.

How does it hold up on this queue’s drafts?

It holds up on structure, including against this post. We scored the 10 other drafts in this queue with Collimer, our own product, before scheduling them. On the first pass, 9 of the 10 scored 3 out of 3 on its answer-first signal, with the first answer sentence between 1% and 8% of the way into the page. The tenth put its first answer at 21% and scored 2 out of 3 until we rewrote the opening. Stat density was the weaker signal: 4 of the 10 maxed it on their final pass. That says the gate does what it claims, putting the answer early and the numbers in. It says nothing yet about whether those posts get cited more.

This post’s own first draft scored 0 out of 3 on answer-first. Its opening hedged (“according to the best published evidence we have”) before it answered, and Collimer found no direct answer sentence at all. The opening above is the rewrite.

What does this not prove?

It does not prove that answer-first posts get cited more. Three limits, stated plainly.

  • Both studies are secondhand. We did not run them. The Princeton paper is from 2024 and measured one engine. Indig’s figure is about ChatGPT.
  • Correlation in the Indig data. Pages that answer early may be cited more for other reasons, such as being better pages.
  • No before-and-after on our own site. We have not compared citation rates of our posts written before the gate with those written after it.

What the gate does guarantee is narrower: every post that clears it states its answer up front and carries at least one quotable, sourced number. A separate rule, covered in why every post here is scored by Collimer first, stops the stat requirement from turning into made-up stats.

What we’re still figuring out

  • Whether our own citation rate moved after July 14. We have the dates and have not run the comparison.
  • Whether 160 characters is the right cut. We have not tested other cutoffs.
  • Whether answer-first writing costs us human readers who wanted the story before the conclusion.

For agents: try this yourself

  • Test the first 160 characters. Given a page and the question it is named for, quote its first 160 characters and say whether they answer the question.
  • Find the quotable sentence. List every sentence on a page that has a subject, a number and a source together, and say which one an engine would most likely lift.

How this was made

Drafted by the Chronicler from repo commits, our own standards and two published guides, then edited and published by Chase.

See how the Chronicler works →

Try this with your own agent

2 prompts you can hand to your own agent (or run by hand) to work with what this post documents. Edit the bracketed parts for your context.

Test the first 160 characters

Here is a page: [URL or paste]. It is meant to answer: "[question]". Quote its first 160 characters of body text (after the title). Say whether those characters answer the question on their own. If not, rewrite the opening so they do, without hedging before the answer.

Find the quotable sentence

List every sentence on [page] that contains a subject, a number and a source together. For each, say whether it still reads correctly if lifted out of the page on its own. Pick the one an AI engine would most likely quote and say why.

More in Chronicle Build