Writing for retrieval, not for ranking
A page that ranks is a page a person will click and scroll. A page that gets cited by an assistant is a page a retrieval system will slice, embed and hand to a model in fragments. Those are different demands, and the second one has a small but genuine research base behind it.
Position matters, inside the model’s context too
The canonical finding is that models do not read a long context evenly. Performance “is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts”. The authors call it a U-shaped curve, with primacy and recency bias at the ends (Liu et al., arXiv:2307.03172, TACL). At its worst, a model given the answer in the middle of its context did worse than the same model given no documents at all.
A SIGIR 2024 study of retrieval pipelines specifically found a related effect: “the position of relevant information should be placed near the query; otherwise, the model seriously struggles to attend to it”. It also found that plausible-but-irrelevant documents, the ones that score well and answer nothing, actively degrade the output (Cuconasu et al., arXiv:2401.14887).
Neither study is about your web page directly. Both are about the machinery your page passes through, and the read-across is the same: a claim buried in the middle of a long undifferentiated passage is the claim least likely to survive.
What cited domains look like
The largest citation study to date compared domains favoured by assistants with those favoured by search engines. The assistant-favoured set showed “more structured, hierarchical HTML, easier-to-read text, lower domain popularity, and more outlinks to reputable sources” (Zhang et al., arXiv:2512.09483).
Read that carefully: it is correlation across sampled URLs. It does not show that restructuring your markup causes citations. What it does do is make semantic HTML, plain language and honest outbound links the best-evidenced bet available, which is convenient, because they are also just good practice.
The honest gap
There is no credible study isolating the causal effect of on-page formatting, headings, lists or tables on citation by public assistants. Anyone who tells you that adding an FAQ block or a summary box lifts AI citations by a specific percentage is extrapolating from their own dataset at best.
Google is blunt about several popular tactics, telling site owners not to chunk content into tiny pieces and not to rewrite content for AI systems (Google). Fragmenting a good page into snippet-bait is not supported by the research and is explicitly discouraged by the one platform that has published guidance.
What this supports in practice
- Make each section self-contained. A passage that answers its own heading survives being extracted; one that depends on three paragraphs above it does not.
- Front-load the claim. Answer first, elaborate after. Position bias is real at every level, including within a chunk.
- Put the evidence in the text. Figures and quoted sources are the one content change with experimental support behind it.
- Use real structure. Headings that describe content, lists that are lists, tables that are tables, and one h1.
- Write for a reader anyway. The correlational evidence favours readable text, and readability is the only one of these you can verify yourself.
Then measure whether it changed anything, which needs more than a handful of prompts, and remember that structure only helps once the page can be reached and read at all.