AI search optimization: how to write passages AI can quote

AI search optimization at the page level comes down to one habit. Write every section so it still makes sense when ChatGPT, Perplexity or Google's AI Mode quotes it on its own. To optimize content for AI search, put the question in the heading, answer it in the first sentence, name the subject instead of writing "it", and back each claim with a number, a date and a source. Use tables for comparisons, and skip writing a separate version for AI.

This guide covers the writing, with a before and after rewrite for each habit. Writing only helps on pages AI search can already fetch and index, and our AI search engine optimization playbook covers that side.

How does AI search choose what to quote?

AI search engines pull the passages that answer each part of a question, often from several sites, and write one answer from them. Microsoft's Bing team wrote in October 2025 that AI assistants break pages into smaller pieces and assemble the most useful ones into a single response. Google's guide to its generative AI features says its systems can tell when a page covers several topics and show people the relevant piece.

So the unit you write for is the section, meaning a heading and the paragraphs under it. The model may read your passage without your intro or the section above it. Every habit below helps a section hold up when it is read alone.

Write headings that match the question

A good heading uses the words people type when they want that section's answer. Microsoft's guide makes the same point by swapping a vague "Learn More" heading for a question about what makes one dishwasher quieter than others. Acme, a made-up e-bike brand, runs through most examples here.

Before:  <h2>Battery</h2>
After:   <h2>How far can an Acme Trail e-bike go on one charge?</h2>

Before:  <h2>Our policy</h2>
After:   <h2>How do I return an Acme bike?</h2>

Not every heading needs a question mark. "Acme Trail vs Acme City" is already what a buyer searches before choosing, so it works as is. The test is whether the heading would make sense typed into a search box. "Overview", "Details" and "Our approach" fail it.

To find the wording, start from questions people already ask. AI Question Finder lists the questions Google AI Overviews or ChatGPT answer about a topic, with monthly searches and the sources each answer lists.

Answer in the first sentence under each heading

The first sentence under a heading should answer it, before any background. If an AI answer keeps one sentence from your section, you want it to be that one.

Heading: How far can an Acme Trail e-bike go on one charge?

Before:  Range is one of the most common questions we get from riders, and
         it's easy to see why. Nobody wants to run out of power halfway up a
         hill. The honest answer is that it depends on a lot of factors, so
         in this section we'll walk through everything that affects it.

After:   An Acme Trail e-bike goes up to 90 km on one charge in Eco mode and
         about 45 km in Turbo. Temperatures below 0°C cut range by roughly a
         fifth, and steep climbs cut it further.

The before version spends 52 words and never gives a number. The after version answers in its first 21 words, then adds the conditions that change the answer. Background can still follow.

Watch for openers like "In this section", "Let's look at" and "Have you ever wondered". They usually mean the answer sits two sentences further down. Our guide to answer engine optimization walks through a longer rewrite of a definition paragraph.

Name the subject instead of writing "it" or "this"

Name the subject in the first sentence of every section, even when the heading and the paragraph above already named it. A quoted passage arrives without the text that told the reader what "it" was.

Heading: Which charger works with my Acme battery?

Before:  It is compatible with all of them, but this depends on the model
         year. As mentioned above, older ones need an adapter.

After:   The Acme 4A charger works with every Acme e-bike battery made since
         2022. Batteries made from 2019 to 2021 also charge on the 4A with
         the Acme C-1 adapter, which costs $25.

The before version leans on four things outside the passage. "It", "them" and "this" point to earlier sentences, and "as mentioned above" points to an earlier section. Quoted alone, the passage answers nothing.

Writers learn to avoid repeating a name. In a passage that may be read alone, repeating it is the job. The same goes for "the tool" and "our platform". Write "Google Search Console" the first time it appears in a section, not "the console".

Back claims with numbers, dates and sources

A claim AI search can use says how much, when, and who says so. A vague claim gives a model nothing to repeat, and plenty of other pages make the same vague claim.

Before:  AI search is growing fast, and more people use it every day.

After:   Google's AI Mode had passed 1 billion monthly active users by July
         2026, Sundar Pichai said in Alphabet's Q2 2026 earnings remarks.

On the published page, link the source inside the sentence, here Google's post of those remarks, so readers can check it.

Research points the same way. In the GEO paper by Aggarwal and colleagues, published at KDD 2024, adding quotations, statistics and citations to sources were among the most effective changes, lifting a page's visibility in generated answers by up to 40%. Keyword stuffing did not help.

Dates matter as much as numbers. Write "in 2025" or "as of October 2026", never "this year", "currently" or "recently". A quoted sentence travels without your page's publish date, so "last year" inside it could mean any year. When a price or version changes, update the number and its date together. The Content Freshness Checker finds missing or conflicting dates on a page and flags dated claims worth a review.

Put comparisons in tables

Use an HTML table whenever a section compares options on the same attributes. Microsoft's guide names comparison tables, numbered steps and lists as formats that break details into clean pieces an AI answer can reuse. A prose comparison makes the reader hold six numbers in mind to find the one they need.

Before:  The Trail has the longest range at 90 km, while the City only gets
         70 km but charges faster, in 3 hours against the Trail's 5. The
         Cargo sits between them at 80 km and 4 hours.

The same facts as a table:

Use real <table> markup with a header row. A grid of <div> elements styled to look like a table loses the link between each value and its column label, and a table saved as an image has no text to quote at all. Keep a one-line conclusion in prose above it, such as "The Acme City charges fastest and the Acme Trail goes furthest", so the section still answers a question if only its text gets quoted.

Does AI search optimization mean writing separately for AI?

No. Google's guide says you don't need to write in a specific way for generative AI search, and that there is "no requirement to break your content into tiny pieces" for its systems to understand it. The same guide says optimizing for generative AI search is still SEO, because those features are built on Google's core ranking and quality systems.

That can read like a contradiction of Microsoft's advice on modular sections. It isn't one. Google is warning against fragmenting pages, and Microsoft is asking for sections that make sense on their own. One long page with clear headings and answer-first sections satisfies both, and it reads better on a phone.

Skip the tactics built only for bots. Google says you don't need AI text files or Markdown versions of your pages to appear in Search, and that Search ignores llms.txt. It also warns that making separate pages for every variation of a query, mainly to sway rankings or AI answers, breaks its scaled content abuse policy. Write one page per topic, for people, and make each section quotable.

Check which passages on your page AI can quote

Citability Analyzer runs these checks on one URL. It splits the page's main content into sections at each h1, h2 and h3. For every section it reports whether the first sentence leads with a direct answer, whether the passage depends on text outside it, such as a leading "It is" or an "as mentioned above", and how many statistics, attributed quotes and outbound sources it contains. It also flags a page with no subheadings, few question headings or sections over 400 words.

You get a 0 to 100 score for English pages with at least 150 words of prose, each section's lead sentence, and the passages worth reviewing with the reason for each. It reads the HTML your server returns, so text that only JavaScript renders isn't checked. It doesn't rewrite anything, and the score is an editing aid, not a prediction that any engine will cite the page. Each run costs 5 credits.