# How to write a page an AI answer will quote

> Extraction rewards one specific shape: a question heading, then a self-contained answer in 40 to 60 words. Most pages bury the answer in paragraph four.

Source: https://armature.tech/library/write-content-an-ai-answer-will-quote
Published: 2026-09-03
Publisher: Armature, Inc. (https://armature.tech)

---

There is one pattern that does most of the work, and it takes ten minutes per page.

## The pattern

> Put the question a person would actually type as the heading. Immediately below it, answer that question directly in 40 to 60 words, in a paragraph that makes sense with no reference to anything before it.

That is it. Everything else on this page is detail.

## Why it works

An extraction system is looking for a self-contained statement that answers a question without needing the rest of the page. If it finds one, it uses it and cites you. If it does not, it takes something from a page that has one, or writes its own summary and cites nobody.

Most pages make it fail. They open with context, then background, then a story, then the answer in paragraph four. A parser takes the first plausible thing it finds, and paragraph one is rarely it.

## Writing the answer paragraph

**Self-contained.** No "as we saw above", no "this", no "it" pointing at an earlier noun. Somebody should be able to lift the paragraph out and have it still make sense.

**Direct.** The first sentence answers. Qualifications come after.

**Specific.** One real number, dated if it can be. "Fast" is not extractable. "Returns in under 40 milliseconds at the 95th percentile, measured in September 2026" is.

**Complete but short.** 40 to 60 words. If it needs a second paragraph, it is not one answer.

Bad:

> Search relevance is a complex topic with many facets. Different teams approach it differently depending on their needs and constraints. In this article we will explore several approaches and discuss the trade-offs involved.

Good:

> Typo tolerance means the search returns the right result when the query is misspelled. It needs an index that stores character sequences rather than whole words, usually a trigram index. Postgres supports it through the `pg_trgm` extension. Dedicated search engines enable it by default.

## Writing the heading

Use the question. Not the topic.

| Write this | Not this |
| --- | --- |
| How much does it cost? | Pricing |
| How does it work? | Architecture |
| Can I self-host it? | Deployment options |
| What happens when a job fails? | Error handling |
| Why is my search missing results? | Relevance tuning |

Topic headings are a habit from print. Question headings match what gets typed.

## The other extractable shapes

**Tables.** Models extract tables reliably. Any comparison, any set of options with attributes, any group of numbers should be a table rather than prose. A five-row table communicates more and survives extraction better than four paragraphs.

**Numbered steps.** For anything procedural. One action per step. No step that contains two actions.

**Definition sentences.** "X is Y that does Z." Boring and highly quotable.

**Short lists with bold leads.** Each item starting with the term in bold, then the explanation.

## The markup

Add JSON-LD for what the page actually is: `FAQPage` if there is a visible FAQ, `HowTo` for a procedure, `Article` otherwise, plus `BreadcrumbList` and `Organization`.

One rule that matters more than the markup itself: **the schema must describe what a reader can see.** Marking up an FAQ that is not on the page is a guidelines violation and it can cost a whole site its rich results. The markup is a label, not a claim.

## The things that stop extraction entirely

Any one of these is enough.

- The content is injected by JavaScript. Many crawlers do not run it.
- The page is behind a login or a form.
- The answer is only in an image or a diagram.
- The page has no dates, and a competitor's has last month's.
- The page is a PDF. Better than nothing, worse than HTML.

Check the first one with one command:

```bash
curl -s https://example.com/your-page | head -c 2000
```

If your content is not in that output, nothing else on this page matters.

## What does not work

- Keyword density. Extraction does not count keywords.
- Word count for its own sake. A 3,000 word page that never states the answer loses to a 600 word page that does.
- Writing "in 2026" into every heading.
- Self-serving comparison pages. Discounted by models and by readers.
- Hidden text of any kind. It does nothing for extraction and it is a penalty.

## The ten-minute pass

For any page you already have:

1. Find the question it answers. Make that the heading.
2. Write the 40 to 60 word answer. Put it directly under the heading.
3. Turn the longest comparison into a table.
4. Add one specific number with a date.
5. Put a visible published and updated date on the page.
6. Check the content appears in `curl`.

Six steps, ten minutes, and it works on every page you have already written.

## Common questions

### How do you write content that AI answers will quote?

Put a question a person would type as the heading, then answer it directly in 40 to 60 words immediately below, in a paragraph that makes sense on its own with no reference to earlier text.

### How long should the quotable answer be?

About 40 to 60 words. Long enough to be complete, short enough to be lifted whole. If it needs a second paragraph to make sense, it is not extractable.

### Does schema markup make content more quotable?

It removes ambiguity rather than adding ranking. FAQPage, HowTo and Article markup tell a parser what it is already looking at, so it does not have to guess and get it wrong.

### What stops a page from being quoted?

A buried answer, pronouns pointing at earlier paragraphs, no specific numbers, content injected by JavaScript, and gated access. Any one of them is enough.

## Read next

- [Answer engine optimization, explained](https://armature.tech/library/answer-engine-optimization) (Markdown: https://armature.tech/library/answer-engine-optimization.md)
- [Technical content marketing when the reader is a machine](https://armature.tech/library/technical-content-marketing-for-ai-search) (Markdown: https://armature.tech/library/technical-content-marketing-for-ai-search.md)
- [Documentation for coding agents](https://armature.tech/library/documentation-for-coding-agents) (Markdown: https://armature.tech/library/documentation-for-coding-agents.md)
- [Generative engine optimization (GEO)](https://armature.tech/library/generative-engine-optimization) (Markdown: https://armature.tech/library/generative-engine-optimization.md)

---

Armature helps software products get discovered and used by coding agents.
Service: https://armature.tech/discoverability · Results: https://armature.tech/leaderboards/sectors · Contact: contact@armature.tech