What changes AI citation is not pretty formatting but structure: a +17.3% citation study

What decides citation in AI answer engines is content structure, not design. A March 2026 paper, "Structural Feature Engineering for GEO," found that structural optimization raised citation rate by 17.3% and quality ratings by 18.5% across six mainstream generative engines. It designs structure across three levels: macro, meso, and micro. ASAP summarizes how this meshes with our earlier analysis.

Start with the numbers: citation rises without touching content

The paper's central claim is that optimizing content structure raises citation without changing meaning. Across six mainstream generative engines, it showed a 17.3% lift in citation rate and an 18.5% lift in subjective quality. Worth flagging is that these are two separate measures. Citation rate is about whether a link appears in the answer at all; quality is about how usable that answer is. Both rising together suggests structural work moves visibility and trust at the same time.

Macro, meso, micro: taking the three levels apart

The paper's structure model is three levels: macro, meso, and micro. Macro is document architecture, meso is information chunking, and micro is visual emphasis. From a practitioner's view this breakdown is useful because it makes the edit points concrete. Macro is your table of contents and section order, meso is whether a paragraph holds a single answer, and micro is emphasis markup. The key is designing these three levels so answer engines can parse them.

"Formatting" and "structure" sit on different layers

The structure here is different from pretty formatting. ASAP's "What Gets Cited" (arXiv 2605.25517) found that surface formatting like bold or ratings has little effect. The structure this paper means is the semantic structure a machine chunks and parses, which is a different layer. Put simply, if the earlier study is about "applying paint," this one is about "splitting paragraphs and ordering them."

Reading the two studies together (ASAP's take)

The two studies do not conflict; they complement each other. Surface decoration like color and weight has small effect, but the information structure a machine parses, like chunking and architecture, changes citation by 17.3%. Translated to practice, priorities shift. The time once spent tuning fonts and colors is better moved to structural work: one answer per paragraph, subheadings that expose the question form. The outline meeting should come before the design meeting.

What it means for Korean content, and the open question

For Korean-market practitioners this result matters in particular. Domestic content competes hard on visual design while information-structure design is relatively empty, so getting chunking and architecture right leaves real room to lead. The limits are just as clear, though. The paper's experiment runs on six English-language engines, and whether the same margin holds under Korean tokenization or domestic answer engines is untested. Rather than copying the numbers verbatim, the safer move, in ASAP's view, is to adopt only the direction: structure comes before presentation.

Source: ASAP summary of "Structural Feature Engineering for Generative Engine Optimization" (arXiv 2603.29979, March 31, 2026; Junwei Yu et al., six generative engines, +17.3% citation and +18.5% quality, macro-meso-micro three levels).