Characteristics of pages that AI cites

In pages that AI cites, the first sentence is the answer and the headings are split into question-sized units. Their schema.org structured data doesn't diverge from the body, they cite authoritative sources within sentences, and they put a short answer and a long body together on one page. When ChatGPT, Perplexity, and Google AI Overviews assemble an answer, the pages they pick up first are determined by this form rather than by length.

These five items are the common skeleton found by reverse-analyzing Korean and English pages that were carried into AI answers almost verbatim. Of those, we kept only the items where PION, in its agency work, confirmed that the rate at which pages are attached as sources in AI answers actually moves. Rather than changing everything at once, decide which item to fix first for each page type and fix them in turn; that brings results faster.

1. Pages where the first sentence is itself the answer

If the first sentence of the intro or of each H2 is a complete answer to the question, AI takes that sentence almost verbatim. In AI search, the user's decision is made inside the answer screen, so "whether your sentence made it into that answer" determines visibility more than whether the page was indexed.

A sentence whose meaning overlaps with the user's question is more likely to be excerpted when it sits near the top of the screen. Search Engine Land calls this approach "answer-first" writing.

On a page targeting "when should I apply retinol," write the first sentence as "Apply retinol at night, on dry skin after cleansing." and put the background explanation below it. If you open with "the history of retinol goes back to the 1970s," the answer gets pushed three paragraphs down and AI can't find that sentence.

Here is how to fix it.

• Open the first sentence of each H2 in an assertive "X is ~" form.
• Use the name of the concept as the subject instead of pronouns such as "this" or "that."
• Within the first 200 characters of the intro, take the question a user would actually type, as it is, and answer it.

Leading with the conclusion feels awkward in interview or storytelling content. Exclude such pages from AEO and run them for other purposes.

2. What kind of headings does AI excerpt easily?

Pages with shallow heading levels written as questions get excerpted often. Instead of reading a page from start to finish, AI divides it into chunks of meaning by heading and takes only the body under the heading closest to the user's question. The information architecture principles that Moz has long emphasized as search fundamentals have become citation signals as they are.

Good headings follow word order similar to what people type into a search box. If you're targeting "best place for whitening treatments in Gangnam," write an H2 along the lines of "how to choose the best place for whitening treatments in Gangnam." Keyword-list headings like "whitening treatment Gangnam recommendation list" are steadily losing power in AI matching.

Use only one H1, stop at three or four H2s, and answer only one question per heading. If you cram two questions under one heading, the excerpt unit blurs, and AI, not knowing what to take, sometimes skips that passage entirely.

3. schema.org structured data that matches the body text

schema.org JSON-LD organizes page content in a machine-readable form and must match the text visible on the page. Structured data is a device that writes the same content once more in a machine-readable form. In its guide to AI features, Google states that there is no special schema you need to add to appear in AI Overviews and recommends keeping structured data consistent with the visible text on the page. Moz's explanation of schema markup also says that schema helps search engines understand the meaning and context of your content.

Korean sites often miss the following.

• FAQ markup that doesn't match the body. If the question and answer text inside FAQPage differs from the body visible on screen, that violates Google's spam policies.
• Inconsistent company names. If the name, alternate names, and url in the Organization schema differ across the site, social profiles, and press coverage, AI can't group them as the same company.
• Missing article metadata. Empty author and publication date fields in BlogPosting.

These three items are also what PION checks first when it takes on an account and opens a site for the first time. A wrong schema is worse than none, so even if it passes the Rich Results Test, a person separately checks whether it actually matches the body.

4. Authoritative sources woven into sentences

Pages that put a source and a link next to a claim get excerpted as more trustworthy evidence. An LLM reads "claim + source + link" as one set as a signal of factuality. This follows the same logic as the evidence-backed content principle Semrush emphasizes.

Put citations inside sentences, as in "according to X," "X's 2026 report says," or "X recommends ~." Once or twice per section is enough, and you can reuse the same source in each section with different wording. For links, leave the outlet name as plain text and wrap only the title of the material; that makes it easy for both people and AI to recognize.

The caliber of sources matters more for citation than the number of links. AI doesn't count a page loaded only with anonymous blog or community links as a trust signal; the signal registers only with authoritative industry outlets or academic and public-institution domains.

5. How do you build a two-tier layout that works in favor of citation?

A two-tier layout, with a compressed answer of 50–300 characters on top and a 1,500–3,000-character background body below, works in favor of citation.

AI uses both tiers together when it builds an answer. The short answer goes straight into the response body, and the in-depth body is called up as background evidence when the user asks back "why?" With only one tier, the page serves only one purpose and loses that many chances to be cited.

Before publishing, check the short answer against the long body using the items below.

• Is there a complete answer to the question within the top 300 characters?
• Does the body below fill in at least 1,500 characters of background?
• Are the short answer and the long body consistent, using the same figures and the same terms?

The last item is the one most often missed. If a figure written as "3 times" above is rewritten as "two or three times" in the body below, AI doesn't know which figure to trust and rates both lower.