Content structure

Chunking exists because embedding and chat models cap input length, and because one vector describes a long mixed page badly^[1]. Microsoft's retrieval documentation recommends starting at 512 tokens per chunk with 25% overlap, and names HTML heading syntax as a way to cut a document into sections^[4]. A 2024 evaluation found that semantic chunking, which places boundaries by meaning rather than by length, does not reliably beat fixed-size splitting^[5]. Headings on a page therefore act less as a ranking signal than as a boundary: they decide where a retrievable unit starts and stops.

The paper that named generative engine optimization tested nine content edits against a benchmark of 10,000 queries^[6]. Adding quotations, statistics or source citations produced a 30 to 40% relative gain on its visibility metric^[7]. Keyword stuffing scored 17.8 against a 19.5 baseline, worse than leaving the text alone^[8]. The pattern favours text carrying checkable specifics over text shaped for keyword match, one of the practical splits between SEO and GEO.

Google runs a passage ranking system that identifies individual sections of a page to judge relevance^[9], so a buried answer can surface without the whole page being about it, the mechanism behind featured snippets. Inside the model, position still costs: answers are strongest when the relevant text sits at the start or the end of the context and weaken when it sits in the middle^[10]. An analysis of 1.4 million ChatGPT prompts found cited page titles closer to the prompt than uncited ones, 0.602 against 0.484 cosine similarity^[11].

No answer engine publishes its chunker. Google's guidance says a page needs only to be indexed and snippet-eligible^[12], which leaves formatting advice unverified against the ranking system it claims to address. The format studies are correlational, and the one controlled benchmark optimises a proxy score over a synthetic query set rather than live traffic^[6]. Engines also attribute imperfectly: a 2023 evaluation of four generative search engines found 51.5% of generated sentences fully supported by their citations and 74.5% of citations supporting the sentence they sat under^[13].