Content6 min read

Anatomy of a citable passage: how to write paragraphs AI engines quote

AI answers are assembled from paragraphs, not pages. The paragraph under your heading either stands on its own as an answer or it does not get used.

By Shimon Carroll, Founder, SEO for AI Agents · Published

A citable passage is a paragraph that answers one question completely on its own. It names the subject instead of saying "it", gives the answer in its first sentence, adds one or two specifics such as a number, a date, or a named source, and sits directly under a heading that states the question. AI engines retrieve and quote passages rather than whole pages, so a paragraph built this way can be lifted into an answer intact, while a paragraph that depends on the three before it cannot.

That is the whole idea. The rest of this guide is the evidence for it, a rewrite you can copy, and the mistakes that quietly make good pages unquotable.

Why the paragraph became the unit of SEO

Answer engines work in two steps: find candidate sources, then compose an answer from pieces of them. Google says its AI Overviews and AI Mode use a "query fan-out" technique, "issuing multiple related searches across subtopics and data sources", and then show supporting links to the pages they drew from. The piece that gets drawn from is a passage. The same is true of featured snippets, which Google says its systems choose automatically when they judge "a page would make a good featured snippet" and which you cannot request.

There is one hard eligibility rule underneath all of this. Google states that to appear as a supporting link in AI Overviews or AI Mode, a page "must be indexed and eligible to be shown in Google Search with a snippet." Snippet controls such as nosnippet and max-snippet therefore affect AI features too. A page that blocks snippets has opted out of being quoted.

What the research actually measured

The most rigorous public study of what makes content more visible inside generated answers is the GEO paper by a team led from Princeton University, published at the ACM KDD conference in 2024. They built a benchmark of 10,000 queries and tested nine ways of rewriting source content. Adding quotations from relevant sources raised the visibility of a source in generated answers by about 41 percent on their main measure, adding statistics by about 33 percent, citing sources by about 28 percent, and improving fluency by about 29 percent. Keyword stuffing, the classic SEO reflex, made visibility 8 percent worse.

Two findings from the paper matter for smaller sites. The gains were largest for sources that were ranked lower to begin with: citing sources lifted the visibility of fifth-ranked sources by 115 percent while lowering it by about 30 percent for sources already ranked first. And the effective methods all add verifiable substance, a quote, a number, a source, rather than rearranging words. The experiments ran mainly on a generative engine built for the study, so treat the percentages as directional, not as a forecast for any one assistant.

The five properties of a citable passage

  1. It answers in the first sentence. Not "In this section we will look at", but the answer itself. If the passage were cut off after one sentence, the reader should still have it.
  2. It names its subject. A retrieved passage arrives without the paragraphs around it, so "it", "this", and "the tool" become meaningless. Repeat the entity name.
  3. It carries one or two checkable specifics: a figure with its source and date, a named standard, a concrete example. This is the substance the GEO study found engines reward.
  4. It sits under a heading that states the question in the words a person would use. Headings are how a retrieval system finds the passage that answers a sub-question.
  5. It is complete but not bloated. Long enough to answer without the reader needing the next paragraph, short enough to be lifted whole. Our audit uses a published window of 134 to 167 words as a target; treat any such number as a guide, not a rule.

A before-and-after rewrite

Here is a typical opening from a service page, followed by the same content rewritten as a citable passage.

Welcome to Harbor Tax Partners! When it comes to your finances, you deserve a team that truly cares. In this guide we will walk you through everything you need to know about how we can help you with your taxes this season.

Before: nothing an engine can quote

Harbor Tax Partners is a CPA firm in Baltimore that prepares federal and Maryland returns for small businesses and their owners. Most engagements cost between $1,200 and $3,500, depending on entity type and the number of states you file in, and returns are delivered within 15 business days of receiving complete documents.

After: names the entity, answers, adds specifics

The rewrite names the business and what it is, answers the questions a buyer actually asks first (what, where, how much, how long), and contains specifics an engine can quote with confidence. The figures in this example are invented for illustration; on your own page they must be true and current, because a quoted wrong price is worse than no quote.

Mistakes that make good pages unquotable

  • Burying the answer under a long introduction, a hero slogan, or a table of contents. The first passage under the H1 is the most likely to be read.
  • Putting the answer inside an accordion, tab, or carousel that only renders after JavaScript runs. Several AI crawlers never see it; see which rendering modes AI crawlers can read.
  • Writing headings as slogans ("Peace of mind, delivered") instead of questions ("How long does a tax return take?").
  • Splitting one answer across a heading, a bullet list, and a callout box, so no single passage holds it.
  • Stuffing the target phrase into every sentence. The GEO study found it lowered visibility.
  • Leaving numbers unsourced or undated. A specific that cannot be checked is a liability, not an asset.

A checklist for your top 20 pages

  1. List the questions each page should answer, using real customer questions from sales calls, support tickets, and the People Also Ask results for the topic.
  2. Give each question its own heading, phrased as the question.
  3. Rewrite the first passage under each heading to answer in its first sentence and name the subject.
  4. Add one sourced, dated specific to each passage where you have one.
  5. Confirm the passages are in the raw HTML and that no snippet control blocks them.
  6. Re-check every quarter: specifics go stale, and a stale number is the fastest way to lose a citation.

How SEO for AI Agents measures this

The passage extractability check walks every paragraph on a page in order and records verifiable facts about each one: its exact word count, the heading above it, and whether it falls in the published target window. The result is a heatmap down the page, so you can see which paragraphs are cleanly quotable and which are fragments or walls of text. Every cell maps to a real paragraph you can recount yourself; read more in the glossary entry on passage extractability.

Whether rewriting actually earns citations is an empirical question, and AI answers vary from run to run; SparkToro found the same prompt almost never returns the same list twice. So the citation measurement asks each engine the same question several times and reports how often you were cited, with a confidence interval and the verbatim answers. Rewrite a page, re-measure, and judge the change by that rate rather than a single screenshot.

Keep reading

Sources