1. Introduction
The H2 heading is the primary structural subdivision of web content. Search engine optimisation practitioners have long recommended question-formatted H2 headings to match user search queries. However, the relationship between heading format and AI citation has not been systematically examined. This study presents a taxonomy of H2 patterns observed across 21 service websites, classifying each of 232 headings by format type and examining how these patterns distribute across cited and uncited sites.
2. Taxonomy Development
Each H2 heading was classified into one of three categories based on its syntactic structure:
Type Q (Question-based): Headings phrased as interrogative sentences, typically beginning with "What," "How," "Why," or ending with a question mark. Example pattern: "What Makes [Entity] Different?"
Type D (Declarative): Headings phrased as statements, typically describing a section's content, a benefit, or a capability. Example pattern: "Full-Cycle Development Services."
Type N (Numeric): Headings consisting primarily or exclusively of numbers, statistics, or metric callouts. Example pattern: "600+" or "$185k."
3. Results
| Heading Type | Cited Sites (Score 5+) | Uncited Sites (Score 0) | Distribution Ratio |
|---|---|---|---|
| Type Q (Question) | 6% of H2s | 54% of H2s | 0.1x (uncited-dominant) |
| Type D (Declarative) | 52% of H2s | 46% of H2s | 1.1x (balanced) |
| Type N (Numeric) | 42% of H2s | 0% of H2s | ∞ (cited-only) |
3.1 Question-Based Headings and Citation
The most striking correlation in the dataset is the inverse relationship between question-formatted H2 prevalence and citation score. Uncited properties in this sample used question-formatted H2s at rates between 71% and 75%. Among cited sites, question-formatted headings were rare: the highest-scoring site in the sample (score 10) used zero question-formatted H2s. Cited game development outsourcing providers averaged 2 question-formatted H2s out of 8-10 total, while uncited providers in the same vertical used 5-6 question-formatted H2s out of 7 total.
3.2 Declarative Headings Across Verticals
Declarative headings were used by both cited and uncited sites, making them the least discriminating heading type in isolation. However, declarative heading content differed by citation status. Cited sites used benefit-oriented and outcome-oriented declaratives, while uncited sites used process-descriptive and self-identifying declaratives. The taxonomy alone does not capture this semantic distinction.
3.3 The Numeric Heading Pattern
One design-as-a-service provider employed 117 H2 headings, the vast majority consisting of single numeric values used as statistical callouts. This extreme pattern — using headings as display metrics rather than section labels — was unique to the cited corpus. While this site's citation score (6) was moderate, the pattern represents a distinct heading strategy that converts the page structure into a metrics dashboard rather than a narrative document.
3.4 Heading Count Distribution
Cited sites averaged 14 H2 headings versus 6 for uncited sites (2.3x ratio). However, variance was extreme: one cited site used 117 H2s while the highest-scoring site used only 6. H2 count alone is not predictive; the combination of count, format, and content appears more relevant.
4. Discussion
The question-heading penalty observed in this sample likely reflects a confound rather than a direct causal mechanism. Question-formatted headings are a hallmark of SEO-optimised content designed to match voice-search queries. Sites that rely heavily on this technique may be newer, less established, or more focused on search optimisation than on demonstrating industry expertise. AI systems may not be penalising the question format itself but rather selecting against the overall content profile that accompanies heavy question-heading usage.
The taxonomy presented here — Question, Declarative, Numeric — provides a classification system that can be applied to larger datasets for cross-industry comparison. Future research should test whether heading format has independent explanatory power after controlling for external validation signals, domain authority, and content age.
5. Conclusions
The 232 H2 headings in this sample cluster into three types with distinct distribution patterns across cited and uncited service websites. Question-formatted headings correlate inversely with citation score, declarative headings are format-neutral, and numeric headings appear exclusively in the cited corpus. These patterns are observational and confounded, but the taxonomy provides a replicable framework for future studies examining content structure and AI visibility.