October 16, 2026 ·

Your H2 structure decides which passage gets quoted

Your H2 structure decides which passage gets quoted illustration for CiteTrack AI

⚡ Key takeaways

  • → An AI answer engine splits your page into chunks (a heading plus the paragraph under it), scores each chunk against the prompt, and quotes the single best-matching chunk. The H2 above a passage is the strongest signal a retrieval system uses to decide what question that chunk answers.
  • → Only a small share of ChatGPT citations point to URLs on Google's first page, which proves citation logic (passage-level scoring) is different from ranking logic (whole-page scoring). Pages that never crack page one still get quoted daily.
  • → Question-phrased H2s that mirror the exact prompt a buyer types have the strongest correlation with citation, per HubSpot's analysis of 14 million AI citations. Keep headings under 10 words.
  • → The paragraph directly under an H2 should state the answer in sentence one and run 40 to 75 words, the band engines quote most often. Cited passages also lean on plain, definitive wording rather than hedged phrasing.
  • → Over-optimizing backfires: all-question headings destroy the page's readable spine, forced short answers amputate topics that need depth, and definitive phrasing on uncertain claims spends trust you can't rebuild.

How an H2 assigns your passage to a query

An H2 maps a passage to a query by giving retrieval a labeled unit to score. When an AI answer engine runs a search, it doesn't grade your whole page against the prompt. It chops the page into chunks, and the heading text above each chunk becomes the strongest signal for what question that chunk answers.

A passage (or chunk) is a heading plus the paragraph or two directly under it. It's the unit the system stores, retrieves, and quotes. Think of it as a card, not a chapter.

This is why an AI "quotes a passage" instead of citing the whole page. When ChatGPT or Perplexity answers a question, it found one chunk that answered the specific prompt and lifted that. The rest of your page might as well not exist for that query. A different prompt could pull a different chunk from the same page, or none at all.

The mechanic behind this is retrieval-augmented generation: the engine retrieves relevant chunks first, then writes an answer from them. During retrieval, the H2 acts as the chunk's headline. It carries most of the meaning the system uses to decide whether this passage matches the query.

A heading like "Overview of pricing models" labels a chunk that answers almost nothing. "How much does citation tracking cost per month" labels a chunk a real prompt can find. The passage below the second heading is the one that gets quoted. For the full retrieval pipeline covering all five stages, see how AI decides what to cite.

Why the H2 above a passage decides the match, not the whole page

An answer engine never quotes "your page." It quotes one chunk, and the H2 sitting above that chunk is the label the retriever reads to decide what the chunk is about.

Here is the mechanic. Retrieval-augmented systems chunk a page into small sections, usually one heading plus the text under it. Each chunk gets an embedding, a numeric representation of its meaning. The prompt gets one too. The engine compares them and grabs the closest match. Your H2 text carries heavy weight in that comparison because it names the topic of everything below it.

This is why a genuinely good passage under a vague heading loses. Say you wrote a sharp 50-word explanation of how AI referral attribution works, but you parked it under an H2 that reads "Getting started." The retriever sees a chunk labeled "getting started" and scores it low against a prompt like "how does AI referral attribution work." The answer is right there in your prose. The label told the system to skip it.

Question-based headings win because they mirror the shape of the prompt. HubSpot's analysis of 14 million AI citations found question-phrased headings had the strongest correlation with getting cited.

The clearest proof that citation logic differs from ranking logic is the overlap gap. Only a small minority of the URLs ChatGPT cites also appear on Google's first page for the same prompt. Most of what ChatGPT quotes is not what Google ranks first. Google's algorithm scores a whole page against a query. An answer engine scores individual passages, and the H2 is what turns a passage into a scorable, findable unit.

That gap is the opportunity. Pages that never crack page one still get quoted daily in AI answers, purely on the strength of their passage-to-prompt structure.

Write each H2 as the question your buyer actually prompts

CiteTrack citability scoring showing per-page AEO and GEO scores with heading structure analysis
CiteTrack's Chrome extension scores heading structure and shows which H2s match real buyer prompts and which don't.

An H2 works best when it repeats the question a real person types into ChatGPT, not the keyword your team argues about in a spreadsheet. Question-based headings have the strongest correlation with getting cited (HubSpot, 14 million AI citations). The retrieval system matches the prompt against your heading, then quotes the passage beneath it.

Keep the heading under 10 words. Longer dilutes the match. Mirror the buyer's actual words. Someone shifting from Google to Perplexity asks "how do I know if ChatGPT is citing my site," not "AI citation monitoring solutions."

Weak H2 (keyword guess)Strong H2 (real prompt)
AI visibility solutions overviewHow do I check if AI is citing my brand?
Citation tracking benefitsDoes AI visibility tracking affect revenue?
llms.txt best practicesWhat does an llms.txt file actually do?
Comparing AEO platformsHow is CiteTrack different from Profound?

To find these questions, work from query fan-out instead of guessing. Ask ChatGPT or Perplexity your seed query, then note every follow-up it suggests or answers on its own. Each distinct follow-up is a candidate H2.

  1. Prompt each engine (ChatGPT, Gemini, Perplexity, Claude, Grok) with your core question.
  2. Copy the sub-questions they raise inside the answer or in follow-ups.
  3. Pull real phrasings from your support tickets, sales calls, and Search Console queries.
  4. Cluster near-duplicates, then write one H2 per cluster in the buyer's exact words.

AirOps found pages with structured, question-based headings and FAQ sections saw a 2.8x citation lift over unstructured pages. One caution: the heading only earns the citation if the paragraph under it delivers a tight, direct answer.

The answer-first paragraph that sits under every H2

The paragraph directly under an H2 is the one that gets pulled. Put the answer in the first sentence, then support it. AI systems retrieve the chunk that resolves the question, not the one that introduces it.

  1. State the answer in sentence one. Lead with the claim, number, or definition the H2 promises. This makes the passage stand alone when lifted out of context.
  2. Keep the whole passage between 40 and 75 words. Long enough to answer the question completely, short enough for an engine to lift whole without trimming.
  3. Use definitive language, not hedged. Write "X does Y" instead of "X may sometimes help with Y." Definitive phrasing was nearly twice as likely to be cited.
  4. Make the passage readable with zero surrounding context. If it needs the paragraph above to make sense, rewrite it so the subject is named inside the passage.
  5. Swap prose for a table when comparing values. A table gives an engine discrete values it can quote and attribute, where the same numbers in sentences arrive as one undifferentiated block.

The choice between a table and a paragraph comes down to shape. Comparing three tools across five features? Table. Explaining one mechanism or making one claim? A 40 to 75 word answer-first paragraph. For the broader editorial method, see how to write content ChatGPT will cite.

Before and after: rewriting one H2 section to win the citation

A section from a real SaaS help page about reducing customer churn. The H2 reads "Understanding churn." The answer sits three sentences deep.

Before:

"Understanding churn is essential for any subscription business. There are many factors that contribute to why customers leave, and teams often overlook the early signals. Broadly speaking, monthly churn measures the percentage of customers who cancel in a given month, calculated by dividing the number who left by the number you started with, then multiplying by 100. Getting a handle on this metric is the first step toward improvement."

The actual answer is buried in sentence three, wrapped in setup on both sides. The chunk spends its first 30 words on nothing quotable.

After — H2: "How do you calculate monthly churn rate?"

"Monthly churn rate is the percentage of customers who cancel in a given month. Divide the number of customers who left during the month by the number you had at the start, then multiply by 100. If you began with 500 customers and 20 canceled, your monthly churn rate is 4%."

Same facts. 52 words. What changed:

  • The H2 became the exact prompt. "Understanding churn" matches no prompt. "How do you calculate monthly churn rate?" matches the one people type.
  • The answer sits in sentence one. 44% of ChatGPT citations come from the first 30% of a page's content.
  • The passage hits the 40-75 word band. Short enough to quote whole, complete enough to stand alone.
  • A concrete number replaces the hedge. "500 customers, 20 canceled, 4%" gives Perplexity a quotable fact and ChatGPT a worked example.

Why heading order and the first 30% of the page change what gets read

Position on the page changes what gets read at all. Kevin Indig's analysis of 18,012 verified ChatGPT citations, published by Search Engine Land, found 44% came from the first 30% of a page's content. Where a passage sits matters almost as much as how it's written.

The mechanism is simple. Answer engines break a page into chunks, then score each chunk. Chunks near the top get scored first and carry a small positional weight. A great answer buried under 1,500 words of setup still competes, but it starts from behind.

Put your highest-intent question H2 near the top. The question a searcher actually types belongs in the first third of the page, with its answer directly under it.

Sequence matters too. In AirOps' analysis of 12,000+ URLs, 68.7% of ChatGPT-cited pages followed proper sequential heading structure (H2s in logical order with H3s nested underneath), versus just under a quarter of traditional Google page-one results. The structure that wins citations is not the same structure that ranks in classic search.

Two moves follow: lead the page with the question carrying the most buyer intent, and keep your H2s in an order a reader and a parser can follow.

How ChatGPT, Gemini, Perplexity, Claude, and Copilot treat the same H2 differently

CiteTrack Brand Intelligence dashboard showing what AI says about a brand across engines
Brand Voice shows which passages each engine pulls from your pages and how answers differ across ChatGPT, Perplexity, Gemini, Claude and Grok.

The same H2 gets read differently because each engine weights different signals when it decides what to lift.

EngineWhat it weightsWhat loses the citation
Google AI ModeSequential heading structure, visible step contentBroken hierarchy, relying on deprecated HowTo schema
ChatGPTProse clarity, definitive language, early placement, domain reputationHedged wording, answer buried below the first 30%
PerplexityA tight 40-75 word fact it can quote almost verbatimLong paragraphs, answer wrapped in setup
ClaudeNuanced, well-sourced prose from trusted domainsThin or promotional content without third-party corroboration
CopilotStructured data and recent content via Bing's indexPages not indexed by Bing or missing schema
GrokConcise factual claims, real-time X/Twitter corroborationOutdated content, no social proof

The section that satisfies all six follows the same shape. Write the H2 as the question your buyer prompts. Directly under it, put a 40-75 word answer-first paragraph in definitive language. Follow with supporting detail and steps as clean structure below. That pattern gives Perplexity its quotable line, ChatGPT its early clear prose, and Google its readable hierarchy.

One structure change pulls extra weight across every engine: convert repeated data into a table. A table is the one format all engines find easy to parse, quote, and attribute.

Semantic HTML checklist so your H2s actually count as headings

An AI system reads your headings from the HTML tag, not from how they look on screen. If your H2 is a styled div that mimics a heading, the retrieval layer sees body text, and the passage below it loses the heading-to-query link.

Failure modes worth catching:

  • Styled divs pretending to be headings. A designer or page builder sets a bold 24px div and it looks like an H2. To a parser, it's not one. Common in WordPress themes and block builders that let you set font size without changing the tag.
  • Skipped heading levels. Jumping from H1 to H3 breaks the hierarchy a model uses to understand which passage answers what. 68.7% of ChatGPT-cited pages followed proper sequential heading structure (AirOps).
  • Fake lists and tables. A "table" laid out with divs and CSS grid reads as one undifferentiated blob. Real ul, ol, and table tags let a model extract each item as a discrete fact.

The fix: one H1 per page, H2s for each buyer question, H3s only to break down an H2. No skipping. To confirm headings are real, right-click and Inspect. If the tag says div or span, it's decoration.

Google has deprecated HowTo structured data, so step content must live in visible content under a clear heading rather than relying on that markup. The heading structure has to stand on its own.

Audit which passages get quoted and fix the ones that don't

Run this audit on any page you suspect should be getting cited but isn't.

  1. Prompt each engine with the real buyer question. Open ChatGPT, Gemini, Perplexity, Claude, and Grok. Type the exact question your buyer asks, not your target keyword.
  2. Record the quoted passage verbatim. Copy the sentence or block the engine used. Paste it into a spreadsheet next to the engine name.
  3. Map the quoted passage back to its H2. Find that passage on your page and note the heading directly above it. This tells you which H2 actually won the match.
  4. Flag every H2 that earned nothing. Mark the ones no engine quoted from. Usually H2s phrased as topics ("Our approach") instead of questions.
  5. Rewrite the losers with the answer-first pattern. Turn each dead H2 into the buyer's question, then put a 40-75 word direct answer right under it.
  6. Re-prompt the same five engines after the page is live and crawled. Compare to your first spreadsheet. Either your rewritten passage now shows up, or the H2 still needs work.

Manual auditing works for one page. It falls apart at scale. CiteTrack's citation tracking runs your prompts against ChatGPT, Perplexity, Gemini, Claude and Grok on a schedule, logs which passage each one quotes, ties the quote back to the page, and flags when a competitor takes a citation you used to hold. The Chrome extension scores heading structure and answer placement per page without needing a subscription.

When chasing quotable H2s backfires

Optimizing for citations goes wrong when you treat "quotable" as the only goal.

Question headings on every line reads like a robot. Question H2s correlate with a 2.8x lift (AirOps), but that doesn't mean every heading needs a question mark. When six headings in a row start with "how" or "what," the page loses its shape. Use questions where a buyer prompts that way, and plain declarative headings for context and transitions.

Forcing a 40-75 word answer on a topic that needs 200 words. The word band describes where citations land, not how much the reader needs. Put the tight answer first, then keep writing. A pricing comparison or migration guide deserves the depth. Chopping it to hit a word count leaves the reader with a slogan.

Definitive language on uncertain claims. Cited passages are nearly twice as likely to use definitive phrasing, which tempts writers to state uncertain claims as settled fact. That works until the claim is wrong. State what you know plainly. Flag what is still moving. Google has already deprecated HowTo schema and changed the model behind AI Mode, so anything written as permanent about those features ages fast.

Watch for these signals:

  • Every H2 is a question and you can't tell the page's spine at a glance.
  • An answer feels amputated because you cut real explanation to stay under 75 words.
  • You wrote "always," "never," or "the only way" about something you haven't verified.
  • A test reader asks a follow-up your page should have answered but skipped for brevity.

Structure serves the reader first. Citations follow answers that are genuinely useful.

Frequently asked questions

How long after I rewrite a page will an AI engine start quoting the new passage?

The engine has to re-crawl and re-index the page first, so timing depends on how often it revisits your site. Re-prompt only after the page is live and crawled. If a competitor still gets the citation weeks later, the H2 phrasing or answer structure still needs work, not the crawl timing. For crawl frequency details, see how often AI crawlers revisit a page.

Should I use FAQ schema on my question-based H2s?

FAQ schema reinforces what a question H2 already says, but it doesn't replace clean heading structure. Treat schema as a supporting signal. Validate with Schema.org's validator and CiteTrack's free Schema Markup Generator.

What do I do when the same passage answers two different buyer prompts?

Pick the single prompt with the most buyer intent for the H2. Let query fan-out capture the second question as its own H2 with its own answer-first passage. Trying to make one heading match two prompts dilutes the semantic match for both.

How many question-style H2s is too many on a single page?

No fixed number, but watch whether you can still spot the page's main point at a glance. When several headings in a row all open with "how" or "what," the page loses its shape. Use questions where a buyer prompts that way, and declarative headings for context.

Does passage-first structure hurt traditional Google rankings?

Citation-friendly structure isn't identical to classic ranking structure, and the overlap between cited URLs and Google's top results is small. But answer-first paragraphs, clean heading hierarchy, and question H2s are broadly compatible with good on-page SEO. Test both outcomes separately.

How do I track citations across engines without re-running every prompt by hand?

CiteTrack's citation tracking runs your prompts against ChatGPT, Perplexity, Gemini, Claude and Grok on a schedule, logs which passage each one quotes, ties it back to its source page, and flags when a competitor takes a citation you held. $59/yr.

Where to go from here

For the full editorial playbook on writing citable content beyond H2s, see how to write content ChatGPT will cite. For how chunk extraction works under the hood, see what a chunk actually is. For the practical overview of chunk size optimization, see chunk size content optimization. And the free Chrome extension and AI-Readiness Grader let you audit heading structure and answer placement on any page right now.

Guides · , ,

Want to see if AI cites pages like this?

CiteTrack AI tracks your AI crawlers, citations, and the revenue they drive, inside WordPress.

Get CiteTrack AI

Explore: Features · Crawl Academy · Glossary · Compare · Pricing