# How to write content for AI search: answer first, then prove it

> In short: Write one page per question. Answer it in the first two or three sentences, in the words a buyer would use, then back the answer with specifics a model can repeat: who it's for, prices, numbers, a comparison table, dated sources. Write for a person first; Google says you don't need to write in a special way for AI, and clear pages serve both.

Guide from Keats, published 2026-11-13, updated 2026-11-13. Canonical: https://getkeats.com/guides/answer-first-content

Assistants read fast and quote selectively. When one reads your page while answering a question, it's looking for the sentences that answer that question. If they're in the first lines, plainly stated, you're easy to use. If they're under a hero banner, three paragraphs of scene-setting and a slogan, you're easy to skip.

None of this is a trick for machines. Google's own guidance is blunt: "You don't need to write in a specific way just for generative AI search", and "make pages for your audience". The pages assistants quote most are the ones a hurried person would find useful too.

## One page, one question

Start by deciding the question the page answers, in the buyer's words. Not "Prepco's methodology", but "how long does it take to prepare for the exam while working full time?". A page that tries to answer ten questions answers each of them less directly than a page written for one.

This doesn't mean thin pages. A page can be long and thorough; it just needs one job. Related questions get their own pages, linked to each other.

## The structure

1. **The answer, in two or three sentences.** First thing on the page, under the title. It should make sense quoted on its own.
2. **The evidence.** The specifics that make the answer credible and repeatable: numbers, prices, a table, your own data, a dated source.
3. **The detail.** Steps, edge cases, examples, caveats. Use headings that are themselves questions or plain statements.
4. **What to do next.** One link onward, not a wall of calls to action.

Here's the opening of a Prepco page written that way:

> **How long does exam prep take if you work full time?** Most working adults who pass prepare for 10 to 14 weeks at 8 to 10 hours a week. Prepco's evening plan is built around that: three weeknight sessions and one weekend practice exam.

The first sentence answers the question. The second gives a figure (Prepco's own, from its course data, so it would cite that). The third says where Prepco fits, without adjectives.

## Specifics are what get repeated

Assistants repeat facts, not tone. Compare:

| Vague | Specific |
|---|---|
| "Affordable plans for every learner" | "Two plans, monthly or a one-off fee for the whole course, with both prices on the page" |
| "Trusted by thousands" | "12,400 students in 2025; 86% passed on their first attempt (our data, method linked)" |
| "Works with your tools" | "Exports to Google Calendar and Notion" |
| "Award-winning support" | "Replies within four hours on weekdays" |

(The Prepco figures are made up for this example. On your own pages, use real ones, dated, with how they were measured.)

The 2023 paper that named generative engine optimization tested rewriting methods on a research benchmark and reported visibility gains of "up to 40%" in generative engine answers, with what worked depending on the subject. Its broad lesson matches the table: content that gives a model something concrete to say is easier to use than content that asserts.

## Tables, lists and headings

- **Use a table for any comparison** across more than two things. It's the fastest way for a reader, human or not, to extract a fact.
- **Use numbered lists for steps** and bulleted lists for options.
- **Make headings informative.** "Pricing" is fine; "How much does Prepco cost?" is better on a page about that question.
- **Keep facts in text.** A price inside an image or a carousel a crawler can't run is invisible.

Google also says there's "no requirement to break your content into tiny pieces for AI". Write normal paragraphs; just put the important one first.

## Keep it current, and say when

Date your facts. A page that says "as of November 2026, plans start at..." is more trustworthy than one that doesn't, and it's easier for you to spot when it's stale. Update the page when the facts change, rather than publishing a new one beside it.

## What to skip

- **Hidden text and prompt-like instructions** aimed at assistants. They're spam, and they'll be treated that way.
- **Keyword-stuffed FAQ blocks** of questions nobody asks.
- **Rewritten common knowledge.** Google warns against content that "could easily be produced by a generative AI model"; an assistant has no reason to cite a page that says what it already knows.
- **Separate "AI versions" of pages.** One good page serves both readers.

## Where the page should live

Put it where your buyers and the assistants would look: a product page for product facts, a guide for how-to questions, a comparison page for "X vs Y". For software companies the help centre is often cited more than the blog, which our guide on [AEO for B2B SaaS](https://getkeats.com/guides/aeo-for-b2b-saas) covers. Structured data can describe the page, but it's not what gets you quoted; our guide to [schema markup and llms.txt](https://getkeats.com/guides/schema-markup-ai-search) separates what's proven from what's hoped. And for where a new page fits in the order of work, see [getting named in ChatGPT](https://getkeats.com/guides/get-named-in-chatgpt).


## Your agent does this

Ask your agent: "Take this week's top Keats card and draft the page from its brief." It calls `get_brief`, which lists the target questions, who is named instead, the pages the assistants cited and how the page should be written, and drafts it answer first for you to review.

## Sources

- [Google Search Central: Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide), read 2026-11-13
- [Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735)](https://arxiv.org/abs/2311.09735), read 2026-11-13