Most marketing teams treat headings as visual furniture \u2014 bigger text to break up the page. That habit is now actively costing them AI citations. When a retrieval system decides whether a page is worth pulling into a ChatGPT, Perplexity, or AI Overview answer, the first thing it scans is the document outline. If the outline is sloppy, the page is filtered out before a single sentence is read.<\/p>\n
A 2025 cross-engine analysis pinned the number down: 68.7% of pages cited inside AI answers follow a strict H1 \u2192 H2 \u2192 H3 hierarchy<\/strong>. The non-cited control group looked nothing like that. It was a mess of skipped levels, double H1s, decorative H4s, and headings nested for visual weight rather than meaning. Same words. Same content. Different fate inside the LLM.<\/p>\n If you\u2019ve been doing SEO long enough to remember when H1s were \u201cthe rule,\u201d this is going to feel like 2010 came back wearing a new badge. It kind of did.<\/p>\n Retrieval pipelines for ChatGPT, Perplexity, Gemini, and Google\u2019s AI Overviews don\u2019t read your page top-to-bottom like a person. They chunk it. The chunker uses headings as the segment boundaries. Your H2 isn\u2019t a design choice \u2014 it\u2019s a function call that says, \u201cthe next ~200 words are an answerable unit about X.\u201d<\/p>\n When the hierarchy is clean, the model gets a clear map: H1 names the entity, each H2 names a subtopic, each H3 narrows the subtopic into a concrete claim. The retriever scores those chunks against the user\u2019s query and picks the tightest match.<\/p>\n When the hierarchy is broken \u2014 H1, H3, H2, H4, H2 \u2014 the chunker still segments, but the segments don\u2019t align with concepts. The model can\u2019t tell whether your H4 is a sibling of the previous H3 or a stray label. Confidence drops. Citation probability drops with it. This is the same failure mode you used to see in featured-snippet eligibility, just with bigger consequences and no second chance.<\/p>\n Pair this with the front-load rule (44.2% of LLM citations come from the first 30% of the page) and the picture sharpens fast: the model is trying to find a confident, top-of-page, well-bounded answer block. Headings are the bounding boxes.<\/p>\n I audit a lot of agency client sites, and the same four patterns kill heading hierarchy in production:<\/p>\n 1.<\/strong> Double H1s.<\/strong> A logo or site name is wrapped in an H1, and the actual article title also gets an H1. Two H1s, no clear root. The retriever now has to guess which one names the entity.<\/p>\n 2.<\/strong> Visual-weight nesting.<\/strong> Designers reach for H3 because it \u201clooks right\u201d between paragraphs, even when the previous heading was an H1. The outline jumps from level 1 to level 3 with nothing in between. The chunker sees a missing parent.<\/p>\n 3.<\/strong> Decorative H4s.<\/strong> Sidebars, callout boxes, and \u201crelated reading\u201d widgets ship with H4s that have nothing to do with the article\u2019s structure. They contaminate the outline.<\/p>\n 4.<\/strong> Page-builder defaults.<\/strong> Elementor, Divi, Gutenberg blocks, half the React component libraries \u2014 they all have at least one preset that emits a wrong-level heading. Nobody catches it because the page looks<\/em> fine.<\/p>\n None of this shows up in a Lighthouse score. It shows up in your AI citation rate, which most teams aren\u2019t tracking yet, so the bleeding is invisible.<\/p>\n ` or `<\/p>\nThe mechanic: LLMs parse hierarchy the same way they parse a knowledge graph<\/h2>\n
Why most pages quietly fail the test<\/h2>\n
What to do this week<\/h2>\n
\n