AI tools almost never quote a whole page. They pull a few short passages that answer the question, and most of those passages come from near the top. CXL found 55% of Google AI Overview citations came from the first 30% of a page, and Kevin Indig found 44.2% of ChatGPT citations did too. If your best answer sits halfway down the page after a long warm-up, AI tools are much less likely to use it.
I spent years as an editor before I ever touched SEO, and the old newspaper rule still applies. Put the most important information first. What’s changed is who’s reading. Today a machine skims your page in milliseconds, scores each passage against a question, and keeps only the best few sentences. Every ContentMender page and post now opens with a bolded answer for exactly that reason.
This guide covers how AI tools pick passages, what the research says about page position and structure, and a practical way to restructure your existing pages so they get quoted.

How Do AI Tools Pick Which Passage to Quote?
They break your page into small chunks, score each chunk against the question, and keep only the highest-scoring sentences. Your page competes sentence by sentence, not as a whole.
Dan Petrovic’s team at DEJAN reverse-engineered Google’s Gemini grounding process by studying the raw data its API returns. Here’s what they found across 7,060 queries and 883,262 snippets (DEJAN, March 2026).
| Finding | What the data shows | Source |
|---|---|---|
| Exact sentences, not summaries | Google pulls exact sentences from your page for grounding rather than rewriting them first. The average chunk is about 15.5 words. | DEJAN, 883,262 snippets |
| A fixed word budget | Each query gets a grounding budget of roughly 2,000 words total (median 1,929), shared across all sources. | DEJAN, 7,060 queries |
| Rank decides your share | The top source gets a median of 531 words (28%). The fifth source gets 266 words (13%). | DEJAN, 2026 |
| Only a third of a page survives | On average about a third of a page’s content makes it into grounding. Pages under 1,000 words had 61% coverage. Pages over 3,000 words had 13%. | DEJAN, 2026 |
| The cap is about 540 words | Grounding plateaus at roughly 540 words per page, no matter how long the page is. | DEJAN, 2026 |
That last finding changes how I think about long content. A 4,000-word guide doesn’t get more space in Gemini’s answer than a 2,000-word guide. It just has more sentences competing for the same 540 words. Length still helps you rank and cover more questions, but every section has to earn its place.
DEJAN also found what gets left out. Navigation, footers, time-sensitive promotions, and unrelated content were excluded. What got in were core service statements, specific details, process steps, and pricing.
Why Does the Top of the Page Matter So Much?
Because that’s where AI tools find most of what they quote. Two independent studies, one on Google and one on ChatGPT, found the same steep drop after the first third of a page.
| Finding | What the data shows | Source |
|---|---|---|
| Google AI Overviews | 55% of cited passages came from the first 30% of the page. The 10% to 20% zone alone produced 27 of 100 citations. | CXL, 100 AI Overview citations, Mar 2026 |
| Bottom of the page | The bottom 40% of pages produced 21% of citations. | CXL, 2026 |
| ChatGPT | 44.2% of 18,012 verified ChatGPT citations came from the first 30% of a document. | Kevin Indig, via CXL |
| Answer near the top | A review of 54 AI citation studies scored placing the answer near the top 8.8 out of 10, one of the strongest content factors. | Zyppy, May 2026 |
| Question-first writing | Content that posed a question and answered it immediately was cited about twice as often (18% vs. 8.9%). | Kevin Indig, via Ahrefs |
CXL described the tension well. Long content written for engagement and storytelling works against how AI tools pick passages. Their recommendation, which I agree with, isn’t to write shorter articles. It’s to answer earlier, then build the supporting case underneath.
There’s an important exception. CXL found several citations from the bottom of pages came from FAQ sections, because each question and answer stands on its own. That’s why our posts end with six click-to-open FAQs. They give AI tools a second set of clean, quotable answers at the bottom.
What Makes a Passage Easy for AI to Quote?
A quotable passage answers one question completely, uses specific facts, and makes sense without the paragraphs around it.
| Trait | Weak version | Quotable version |
|---|---|---|
| Self-contained | “This approach works better for most of them.” | “Quarterly pest control plans prevent more infestations than one-time treatments for most homeowners.” |
| Specific | “Google looks at a lot of things to rank local businesses.” | “Google says local results are based mainly on relevance, distance, and prominence.” |
| Direct | “There are many factors to consider when pricing a roof.” | “Most roof replacements in our area cost between [your low price] and [your high price], depending on size and materials.” |
| Sourced | “Studies show reviews matter.” | “BrightLocal’s 2026 survey found 97% of consumers read reviews for local businesses.” |
The bracketed example is a template, so fill it with your real prices. Never publish a number you can’t back up.
These traits line up with the strongest content factors in Zyppy’s review of 54 studies, including factual specificity (8.3), explicit phrasing (8.1), citing sources (8.0), and self-contained passages (8.0) (Zyppy). The original academic GEO research from Princeton and IIT Delhi found similar methods could boost visibility in AI answers by up to 40%, with results varying by subject (Aggarwal et al., KDD 2024).
A simple test I use with writers is to cover everything except one paragraph. If a stranger could read that paragraph alone and know exactly what it means and who said it, it’s quotable.
Does Page Structure Change How Often You Get Cited?
Yes. Clear headings, sections of a reasonable length, and visible text all make it easier for AI tools to find and lift the right passage.
| Finding | What the data shows | Source |
|---|---|---|
| Section length | Pages with 120 to 180 words between headings received about 70% more ChatGPT citations than pages with sections under 50 words. | SE Ranking, 129,000 domains |
| FAQ sections | FAQ sections in the main content nearly doubled the chances of being cited by ChatGPT. | SE Ranking, 2025 |
| AI-ready structure | Headings, sections, and tables that organize content clearly scored 8.6 out of 10 across 54 studies. | Zyppy, 2026 |
| Visible content wins | AI tools that fetch pages live read the visible HTML. Facts hidden in code or behind scripts were often missed. | searchVIU test |
| Content visibility | Information in visible HTML text rather than hidden tabs or scripts scored 7.6 out of 10. | Zyppy, 2026 |
One clarification, because it’s easy to overcorrect. Google’s own AI optimization guide says you can ignore tactics like “chunking” content specifically for AI (Google Search Central). I read that as a warning against chopping pages into choppy fragments. Clear sections with real headings are just good writing. Breaking a guide into 40-word scraps isn’t.
Zyppy made the same point. You don’t need to chunk your content into bite-sized pieces. You need clear structure with headings, sections, and tables.
Why Should One Page Answer the Follow-Up Questions Too?
Because AI tools search for several related questions at once, and a page that answers more of them has more chances to be picked. This is called query fan-out, and Google confirms its AI features use it.
Google says AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics to develop a response (Google Search Central). If someone asks “Should I repair or replace my water heater?” the AI may also look up typical lifespans, repair costs, replacement costs, and warning signs.
| Finding | What the data shows | Source |
|---|---|---|
| Fan-out rankings matter | Ranking for the related fan-out queries scored 9.3 out of 10 across 54 studies, nearly as high as ranking for the main query (9.4). | Zyppy, 2026 |
| Topic clusters help | Ranking for multiple related queries across a site scored 8.9 out of 10, because it raises the odds of being cited at least once. | Zyppy, 2026 |
| Format should fit intent | Matching the page format to the question, such as lists for “best” questions and steps for “how to” questions, scored 9.0. | Zyppy, 2026 |
In practice, that means each important page should answer the main question first, then the three to five follow-up questions a customer would ask next. A water heater page that covers lifespan, cost, warning signs, and repair-versus-replace gives the AI a reason to pull from you for every one of those sub-questions.
It also means your related pages should link to each other. A cluster of connected pages on one topic tells search engines and AI tools you cover that subject thoroughly. That’s how we organize client sites, with service pages, location pages, and supporting posts all linked together.
How Should You Structure a Page for AI Search?
Lead with the answer, organize the rest around the questions people actually ask, and finish with clean FAQs. This is the template we use on every ContentMender post.
| Page section | What goes there | Why it works |
|---|---|---|
| Opening paragraph | A bolded, two-to-four-sentence answer to the page’s main question | Sits in the first 30%, where most citations come from |
| Credibility line | Who’s writing and why they’re qualified | Supports trust signals for readers and AI tools |
| Question headings | H2s phrased the way people ask, like “How Much Does X Cost?” | Matches the sub-questions AI tools generate |
| First sentence under each heading | A direct answer to that heading’s question | Gives each section its own quotable lead |
| Tables and lists | Comparisons, steps, prices, and sourced data | Easy to extract, easy to verify |
| FAQs | Six questions, each answered in two or three sentences | Self-contained answers that can be cited from deep on the page |
You can see the pattern on this page. The answer comes first, every heading is a question, and the first sentence under each one answers it.
How Do You Fix Pages That Bury the Answer?
Find where each page actually answers its main question, then move that answer up. Most pages don’t need a rewrite, just reorganizing.
- Pick your top 20 pages by traffic or leads in Google Search Console.
- Write down each page’s main question in a customer’s words.
- Find the answer on the page. If it appears after the first third, that page goes on the fix list.
- Move the answer into the opening paragraph and bold it.
- Turn vague headings into questions, and put a direct answer in the first sentence under each.
- Rewrite passages that depend on context so they make sense alone. Replace “it,” “this,” and “they” with the actual names.
- Add specific, sourced facts where the page relies on general claims.
- Add or tighten an FAQ section with real questions from customers and sales calls.
Then track results. Ask ChatGPT, Gemini, and Google’s AI Mode your target questions each month and note which pages they cite.
If you’d like help, page restructuring is part of the on-page work in our Hyperfocused Lead Generation and Hyperlocal Lead Generation programs. For related reading, see how meta descriptions work in AI search and why updated pages get cited more in AI search. Or start with a free audit and we’ll show you which of your pages bury the answer.
FAQs
Where on a page do AI tools find what they quote?
Mostly near the top. CXL found 55% of Google AI Overview citations came from the first 30% of a page, and Kevin Indig found 44.2% of ChatGPT citations came from the first 30% of a document.
How much of my page does Google’s AI actually use?
DEJAN’s analysis of Gemini grounding found only about a third of a page’s content makes it into the AI’s working context on average, capped at roughly 540 words per page. Longer pages don’t get more space.
Should I write shorter content for AI search?
Not necessarily. Longer content can rank for more questions. The fix is to answer earlier, put the main answer at the top, and make every section start with a direct answer.
What makes a paragraph easy for AI to quote?
It answers one question completely, uses specific facts, names its sources, and makes sense on its own. Avoid vague references like “this” or “they” that only make sense with surrounding text.
Do FAQ sections help with AI citations?
They can. CXL found several AI Overview citations from the bottom of pages came from FAQ sections, and SE Ranking found FAQ sections nearly doubled ChatGPT citation chances in its data.
Should I chunk my content into small pieces for AI?
No. Google says you can ignore chunking tactics. Use clear headings and sections of reasonable length instead. SE Ranking found sections of 120 to 180 words outperformed very short sections.



