Structuring content for AI: the heading structure that gets quoted

Why does the heading decide what gets quoted?

Because the heading is the label a machine attaches to everything under it. The HTML Standard's section on headings defines the h1 to h6 elements as the headings of the sections they begin, and the order of those elements as an outline of the page. An assistant that lifts a passage as the answer to a question is likely to read the nearest heading above it as the statement of what the passage is about. If the heading says "Our Services", the passage is about services in general. If it says "Do you repair water heaters the same day?", the passage is the answer to that question.

That is the whole reason a question heading works. It is not a trick the assistants reward; it is the heading hierarchy doing what the standard says it does. A page whose headings, read alone, form a list of the questions its visitors ask has told a machine where every answer is before it reads a paragraph.

The hierarchy has to be real. One h1, which is the page's subject. An h2 for each section, and a section is one question or one claim. An h3 only for a point inside a section, never as a way to make text bigger. A heading that skips a level, or a bold paragraph styled to look like a heading, breaks the outline, and a machine reading the outline then binds the passage to the wrong question or to none.

What should go directly under the heading?

The answer, in the first sentence, and the whole of it within 40 to 80 words. That length is a passage an assistant can lift without cutting, and a reader can take in without scrolling. The detail, the exceptions and the reasons come after it, in their own paragraphs. A section that opens with three sentences of context before it answers has put the quotable part where a machine extracting the passage under the heading may never reach it.

The passage has to stand alone. No "as mentioned above", no "this" whose referent is in the previous section, no pronoun standing in for the business name. Read it with the rest of the page covered; if it still says something true and complete, it is an answer. If it needs the paragraph before it, it is a continuation, and continuations are not quoted.

Specifics make a passage worth lifting. A number with its source, a name, a condition ("within the city limits, before noon on a weekday"). This is the mechanism the one academic study of the subject measured: Aggarwal and colleagues, in the paper that named generative engine optimization, found that adding statistics, quotations and citations to a source raised its visibility in generated answers by up to about 40% on their benchmark. The gain came from making the passage checkable, not from making it longer. Citability is that property: self-contained, specific, attributable.

Which HTML elements matter, and which do nothing?

A fetcher reads the HTML a request returns and nothing else. The elements below are what it can and cannot make of the page; the third column is the common way each one is misused.

On the script row: Google describes rendering JavaScript as a separate and deferred step in its JavaScript SEO documentation, and no other assistant's crawler documentation promises a rendering step at all. Words that matter go in the HTML.

Should I write a separate version of the page for AI?

No. Google's page on AI features and your website lists no additional requirements for AI Overviews or AI Mode, no special optimizations and no new files or markup; it sends site owners back to ordinary SEO and people-first content. The structure described here is for people first. A heading that is the question is easier for a person to scan; an answer in the first sentence is what a person on a phone wanted anyway.

The same rule holds for structured data. Google's introduction to structured data says to mark up only what the visitor can see on the page. FAQPage markup describes question-and-answer pairs that are already visible; it does not create them, and Google stopped showing the FAQ rich result in Search on May 7, 2026. Use the markup to make the pairs easy to isolate, and expect the visible text to do the work.

So there is one page, written for the reader, structured so that a machine reading the outline finds the same things the reader does. Anything written for the machine alone is a second version of the truth to keep in step with the first, and the audit below found 43.1% of sites already disagreeing with themselves across just three pages.

What the crawl found

In our audit of 9,483 small business websites (40 trades, 90 metros, September 2026), 50.7% of sites had at least one heading phrased as a question and did not answer it in a form a machine could isolate: the question was there, and nothing under it could be isolated as the answer. 61.5% had no question-and-answer content of any kind.

Length was not the problem. The median homepage ran to 880 words, and only 3.9% were under 150. The pages said plenty; what they said was not in a form a machine could attribute, quote or verify. 3.0% rendered their content only in JavaScript, so a crawler that runs no scripts received an empty shell, and 2.1% kept the real content, the price list or the service menu, in a PDF.

The fix on most of those sites is not new writing. It is moving the sentence that answers the question up to the line under the heading, turning the heading into the question, and putting the words that are in images and files into the HTML. An afternoon per page, for the pages that matter.

One section, before and after

The section below is a composite, not a real site: a heading that names a category, a line of adjectives, and the facts kept in a picture.

<h2>Our Services</h2>
<div class="hero-text">
  Quality workmanship you can count on. Big job or small,
  we handle it all with a smile. Call today!
</div>
<img src="services-menu.png" alt="">

The same section, restructured. Nothing has been added that the business did not already know; the heading has become the question a caller asks, the first sentence answers it, the conditions follow, and the words are in a paragraph rather than a picture.

<h2>Do you repair water heaters the same day?</h2>
<p>Yes, inside the city limits, if you call before noon on a weekday.
The truck carries parts for the common tank models; a tankless unit
usually needs a second visit for the part. The visit fee is stated on
the prices page and comes off the repair if you go ahead.</p>

Read the second version with everything else covered. It names the service, the area, the condition, the exception and where the price is. A machine binding it to the heading above has a question and its answer. A person on a phone has what they called to find out. That is the entire technique, and it repeats for every question the front desk answers in a week.