← All research
Research · · 2 min read

Structure beats rewriting: the stable lever in AI citations

Several independent studies point the same way — giving content structure earns more AI citations than polishing wording, and pure rewriting can even lose ground.

James · CTO #GEO#content#structure#literature

Many teams’ first instinct in GEO is “let’s make the copy flow better.” The measured data keeps saying otherwise: what gets you cited by AI is mostly not smoother sentences, but clearer structure.

Three independent studies, one direction

  • GEO-SFE (arXiv 2603.29979): structural remodeling of a document lifts citations by about +17.3%.
  • SAGEO Arena: structural remodeling +22%, while pure prose rewriting −9% — opposite directions.
  • Princeton’s foundational GEO paper (KDD 2024, arXiv 2311.09735): adding statistics +33–41%, citing sources +30%, adding quotations +28–41% in citation rate.

Put together, the conclusion is consistent: machine-parseable structure signals (heading hierarchy, lists, tables, explicit numbers and citations) decide whether a passage gets cited far more than wording polish does.

“Document structure is the most stable citation signal in AI engines; pure prose rewriting can actually lose ground.”

Why infrastructure dominates content tweaks

One level up, C-SEO Bench (NeurIPS 2025, arXiv 2506.11097) offers an easily overlooked result: infrastructure dominates content tweaks. Whether AI can fetch you and parse your structure comes before “is this sentence well written.”

That’s exactly why our site audit fixes the three-tier weights at 45 / 30 / 25 — basic access highest, structured data next, content and entities last:

  1. Basic access (45): robots allowing AI crawlers, llms.txt, HTTPS, server-readability, sitemap with lastmod.
  2. Structured data (30): JSON-LD (Organization / FAQ / Article…), Open Graph, canonical, hreflang.
  3. Content and Chinese entities (25): heading hierarchy, statistic density, citation density, quotable assertions, freshness, structure signals, encyclopedia/social presence.

What to actually do

If you can only do three things first, in this order:

  1. Let AI in and let it read: allow crawlers, add llms.txt and structured data.
  2. Then give content structure: headings, lists, checkable numbers and citations — not endless wording edits.
  3. Wording comes last: once structure and evidence are in place, polish has marginal value.

Get the order wrong and no amount of rewriting buys back citations. Measure which tier you’re stuck on first — that’s what our site audit does.

For the source behind each weight, read the full methodology.