How website structure and formatting affect AI search citations

AI search systems like ChatGPT, Perplexity, and Claude don’t read your pages the way a human does. They retrieve individual chunks of roughly 100–300 tokens using a process called Retrieval-Augmented Generation (RAG), which means each section of your site either stands on its own as a citable unit or gets filtered out before the model decides what to quote. Structural changes alone, without rewriting your content, have been shown to improve AI citation rates by more than 17%.

Why structure matters: how AI search actually retrieves content

AI search systems don’t read your page from top to bottom and decide whether to cite it. They break your page into fragments (chunks of roughly 100–300 tokens, or about 75–225 words), convert each chunk into a mathematical representation called a vector embedding, and retrieve whichever chunks are most semantically relevant to the user’s query. Your page is never evaluated as a whole. Each chunk competes independently.

This process is called Retrieval-Augmented Generation (RAG), and it’s the underlying architecture behind how ChatGPT, Perplexity, Claude, and Gemini pull in web content. As growth marketer Liam Dunne puts it: “If your content is not structured for this retrieval process, it gets filtered out before the AI even considers citing it.”

The practical implication is that formatting decisions you might treat as aesthetic (where you put headings, how long your paragraphs run, whether your sections open with a direct answer) are actually retrieval decisions. A GEO structural engineering study (Yang et al., 2026) found that structural optimization alone produced a 17.3% improvement in AI citation rate across six generative engines, without changing the underlying content. Structure is the prerequisite. Everything else, including the trust and recency signals that decide citation among eligible pages, follows from it.

Which structural elements do AI retrievers actually respond to?

HTML heading tags (H1, H2, H3) are the primary chunking boundaries for most RAG pipelines. Tools like LangChain’s MarkdownHeaderTextSplitter and LlamaIndex use heading tags as natural breakpoints to segment pages into retrievable chunks. When those headings are clean and hierarchical, the chunker produces sections that each address a single coherent topic. When they’re not, the chunker groups unrelated ideas together into what researchers call a “mixed vector”: a chunk that matches no specific query well and gets retrieved for none of them.

The citation difference is measurable. Heading-based chunks show approximately 3× higher LLM citation rates and 85%+ retrieval accuracy compared to generic character-count chunks, which achieve only 45–60% retrieval accuracy. According to Profound (2025), pages with proper H1–H2–H3 hierarchy are 2.8 times more likely to be cited by AI systems overall.

A common heading mistake is jumping levels for visual styling: using an H4 where you want smaller text, or skipping H3 entirely. From a human perspective, the page looks fine. From a RAG pipeline’s perspective, the heading jump signals that the content under that tag belongs to a different logical level than it does, producing chunks that lose their context.

Each H2 section should function as a self-contained mini-article: a clear topic sentence, supporting evidence, and a conclusion. Ramanath, CTO and Co-Founder of Presenc AI, frames the requirement directly: “Each heading should clearly describe what the following section covers, ideally matching the language patterns users use when querying AI assistants.” If a chunk gets retrieved in isolation, which it will, it needs to make sense without the surrounding sections. Sections that open with “Additionally” or reference data from a previous section without repeating the key numbers fail this test.

Section length matters too. Sections shorter than 100 tokens often lack enough context to be useful as retrieval results. Sections longer than 300 tokens risk being split at non-semantic boundaries by the chunking algorithm, breaking coherence mid-argument.

One more structural factor worth auditing: the HTML underneath your headings. Sites built on React or Tailwind often produce “div soup”: nested layers of generic <div> tags with no semantic meaning. When a RAG chunker processes that markup, it sees a stream of text with no signal about which heading and answer belong together. HTML5 semantic tags like <article>, <section>, and <aside> signal logical boundaries that DOM-aware splitters use to keep related content together. If your site is built on a modern JavaScript framework, this is worth checking with your developer.

How does prose formatting change what an LLM can extract and cite?

44.2% of AI citations come from the first 30% of a document, according to analysis of AI citation patterns by JCT Growth (2024). This pattern, sometimes called the “ski ramp” attention effect, means the opening of your page carries disproportionate weight in whether any of it gets cited. An LLM retrieving chunks from your page will weight content that appears earlier more heavily, and the first 100–200 words of a page are the highest-value real estate you have.

The practical implication: answer the question your page is about in the opening paragraph, not after context-setting. The inverted-pyramid structure from journalism (lead with the conclusion, then support it) is the LLM-friendly format. A product page that buries the key answer under a hero video and three paragraphs of brand positioning is functionally invisible to AI retrieval, regardless of how good the content is further down.

LLMs also show what researchers call the “lost in the middle” effect: attention is strongest at the beginning and end of content, and weakest in the middle. A long page where the best answer is buried across sections 4 through 7 will be cited less often than a shorter page where that answer appears in the first two paragraphs. Research from Princeton and Georgia Tech (2024), as reported by Over The Top SEO, found that formatting optimizations, including answer-first structure and fluent, quotable language, increased AI citation rates by 15–40% depending on query type.

At the paragraph level, the principle is one point per paragraph. When an LLM retrieves a chunk, it pulls 200–500 tokens at a time. A dense paragraph making three related points produces a chunk where any single point is diluted by the others. A paragraph making one point produces a chunk that matches a specific query cleanly. Senior Software Engineer James Patterson puts it plainly: “AI citations flow toward content that is clearly structured, factually dense, and written in a form that survives extraction. Generic marketing prose does not.”

The practical technique is what some practitioners call the “atomic paragraph” approach: if a paragraph contains a transition word like “also,” “additionally,” or “furthermore” mid-paragraph, or joins two concepts with a semicolon, it’s probably carrying more than one point and should be split. Red flags include paragraphs longer than four or five sentences and multiple distinct concepts in a single block.

Question-framed headings and dedicated FAQ sections also improve extractability, because they structurally match the question-and-answer format of AI conversations. A heading phrased as a question (“What does a GEO audit cover?”) signals to the retrieval system that the section below is designed to answer that question directly. Worth saying plainly: the independent impact of question headings and FAQ schema beyond strong fundamentals is contested in practitioner testing; the structural logic is sound, but don’t expect these tactics alone to rescue a page that fails on the basics.

What technical site-level factors determine whether AI crawlers can reach your content at all?

Prose formatting and heading structure only matter if AI crawlers can read your page in the first place. Three technical factors determine whether they can.

The JavaScript rendering gap is the most consequential. GPTBot, ClaudeBot, and PerplexityBot fetch raw HTML and parse it. They do not execute JavaScript, wait for API calls, or use a rendering engine. Vercel’s analysis of 569 million GPTBot requests found zero evidence of JavaScript execution. A 2025 SearchViu study found 69% of AI crawlers cannot execute JavaScript at all. (There is some disagreement about PerplexityBot specifically. One crawler reference table lists it as having full rendering capability, but the Vercel primary-source data, drawn from actual request logs, found no JavaScript execution for any major AI crawler. Treat PerplexityBot’s JavaScript capabilities as uncertain and optimize for the conservative case.)

AI crawlers also abandon requests after 1–5 seconds, compared to Googlebot’s 20-second timeout. A site that ranks well on Google via JavaScript rendering can be entirely blank to every AI crawler. As James Calloway of Geology writes: “A page that scores 95 on PageSpeed Insights can still be invisible to GPTBot if the answer lives inside a JavaScript-rendered component that loads after first paint.” The fix is ensuring your core content appears in the initial HTML response before any JavaScript runs. That means headings, body text, and structured data all present before any scripts execute.

robots.txt misconfiguration is blocking more sites than most owners realize. Approximately 40% of sites audited in 2025 were inadvertently blocking at least one major AI crawler through misconfigured robots.txt rules. AI crawlers are distinct from traditional search crawlers: allowlisting GPTBot has no effect on ClaudeBot or PerplexityBot. Each must be explicitly permitted. Only 35% of the top 1,000 websites explicitly address AI crawlers in their robots.txt at all. Checking your robots.txt takes five minutes and is the highest-leverage single action on this list.

Schema markup is the strongest measurable predictor of AI citation. A pre-registered OSF study analyzing 100,411 AI citation events from ChatGPT, Perplexity, Claude, and Google AI Mode (Lee, 2026) found schema markup had an odds ratio of 1.31, the highest of any content feature measured, above author bylines, answer-first coverage, and list structure. FAQPage schema has the strongest effect among individual schema types, with an odds ratio of approximately 1.69. A Semrush study of 5 million AI-cited URLs found Organization schema on 25–34% of cited pages and Article schema on 20–26%, significantly above the web average. Schema markup is how you make your page’s structure machine-readable to AI systems, not just human-readable.

Page speed thresholds also affect crawl frequency directly. Pages with Core Web Vitals scores above 80 receive 6× faster crawl frequency from AI crawlers compared to pages scoring below 50, per research from ryanshojae.com’s 2025 crawlability audit. The specific technical targets, per getgeology.com’s site speed analysis: Time to First Byte below 400ms (below 200ms is the safe zone), core content present in the first 100KB of HTML, and no broken internal links or redirect chains that waste crawl budget.

What does a same-day formatting audit look like for a site you already have?

You don’t need to rewrite your content to close most of these gaps. A practical GEO audit for an existing site runs across four areas, roughly in order of how quickly you can act on them.

AI crawler access. Open your robots.txt file (yourdomain.com/robots.txt) and check whether GPTBot, ClaudeBot, and PerplexityBot are explicitly allowed. If they’re not mentioned, or if a broad Disallow: / rule is catching them, add explicit allow rules for each. This is a five-minute fix with no developer required. Also check whether a llms.txt file exists at yourdomain.com/llms.txt. This file tells AI systems which pages are most important to crawl; one analysis found sites that have it receive 4.2× more AI crawl requests than sites without it, though evidence that the major AI systems actually read llms.txt is thin. Treat it as a low-cost, unproven bet rather than a guaranteed lever.

Schema markup. Use Google’s Rich Results Test or Schema.org’s validator to check whether your homepage has Organization schema, your blog posts have Article schema, and any FAQ content has FAQPage schema with JSON-LD. Missing schema is the single highest-leverage technical gap to close, and adding JSON-LD doesn’t require changing your visible content at all.

Content structure. Pick your five most important pages and check three things: Does each page have exactly one H1? Does the heading hierarchy run H1 → H2 → H3 without skipping levels? Does the page open with a direct answer to the question it’s trying to answer, in the first paragraph, not after context-setting? As GEO consultant Danielle Birriel notes, most businesses that run this audit discover their primary AI visibility constraint isn’t content volume. It’s that the content they have isn’t structured to be retrieved.

Extractability. Scan your key pages for multi-point paragraphs (longer than four or five sentences, or containing “additionally” / “furthermore” mid-paragraph) and split them. Add a FAQ section to any page that answers a question your customers commonly ask, and mark it up with FAQPage schema. 87% of AI-cited pages have a unique H1, per practitioner audit data, and pages with proper heading hierarchy are 2.8 times more likely to be cited. Both are things you can verify and fix today without touching your content strategy.

The audit doesn’t require a developer for most items. The JavaScript rendering check (whether your core content appears in the initial HTML response) is the one exception, and it’s worth asking your developer or hosting provider about if your site is built on a JavaScript-heavy framework. This is the same ground our structure lever covers from the content side; here it’s the machine’s side of the same story. See how we sequence the work →