What Three H2 Heading Types Were Identified?
Manual classification of all 232 H2 headings across 21 sites revealed three distinct taxonomic categories. Question-format headings are phrased as interrogatives and typically begin with "What," "How," "Why," or "Who." Declarative/benefit-oriented headings make statements about capabilities, services, or outcomes. Numeric/stat-callout headings use numbers as the primary heading content, functioning as visual anchors for supporting data.
The distribution of these types varies dramatically between cited and uncited groups. Uncited sites deploy 71 to 75% of their H2s in question format, reflecting a content strategy explicitly designed to match user queries. Cited sites show much lower question percentages, with most headings in the declarative/benefit category.
| Anonymised Site | Vertical | Score | H2 Count | Question % | Dominant Type |
|---|---|---|---|---|---|
| AU Design Alpha | au-design | 10 | 6 | 0% | Declarative / Action |
| DaaS Alpha | daas | 9 | 2 | 0% | Declarative / Benefit |
| Game Outsource Alpha | game-outsource | 8 | 10 | 20% | Declarative / Mixed |
| DaaS Beta | daas | 8 | 11 | 0% | Declarative / Benefit |
| DaaS Gamma | daas | 6 | 117 | 0% | Numeric / Stat-Callout |
| Game Outsource Delta | game-outsource | 0 | 7 | 71% | Question-Format |
| DaaS Delta | daas | 0 | 8 | 75% | Question-Format |
| GEO Agency Delta | geo-agency | 0 | 7 | 71% | Question-Format |
Table 1. Selected sites showing heading taxonomy distribution by citation score. All names anonymised. Full dataset contains 21 sites.
What Patterns Emerge from the Numeric/Stat-Callout Category?
One cited site in the DaaS vertical uses H2 headings exclusively as numeric stat callouts — headings like "150,000+," "3,000+," "$5,995," and "97+." This creates a distinct third category where the heading itself is a data point rather than a topic label. The site achieves citation score 6 despite using 117 H2s — the highest count in the dataset — suggesting that the heading format (numeric) may be more important than the heading count.
This pattern is unique in the dataset and warrants further investigation. The numeric heading approach converts each section into a self-contained data claim, which may facilitate extraction by RAG systems that are designed to identify and cite specific statistics.
Why Is the Question-Heading Correlation Confounded?
The three uncited sites with the highest question-heading percentages (71-75%) are also the three newest sites in the dataset, all using .one TLD domains with minimal training data presence. The question-format heading strategy may have been adopted specifically because these are new sites following current GEO best-practice recommendations to match user query formats. The heading format and the site newness are perfectly confounded — it is impossible to determine from observational data alone whether the headings, the site age, or an entirely different factor drives the citation outcome.
This finding creates an important tension with the introspective trust signal framework, where question-format H2 headings scored 8.0 (positive direction). The observational data shows the opposite: question H2s are associated with non-citation. Resolving this conflict requires the controlled experiment proposed in SIGI-2026-035.
Suggested Citation
Tavitian, V. & Tavitian, J. (2026). H2 Heading Taxonomy and AI Citation: A Classification Study of 232 Headings Across 21 Websites. The Scientific Institute for Generative Intelligence, SIGI-2026-039. https://generativeintelligence.institute/publications/SIGI-2026-039/