A new study submitted to arXiv by Nick Koudas challenges the prevailing wisdom of the generative engine optimization (GEO) industry. While publishers are scrambling to add FAQ blocks and structured data to win citations in ChatGPT, Gemini, and Perplexity, the data suggests these tactics are largely ineffective for internal ranking. The research, based on two 90-day collections totaling 158,099 answers and 468,037 citations, indicates that what drives citations is fundamentally different from traditional SEO metrics.

The Myth of Structured Data

The study compared pages within the same website to isolate the impact of content properties. It found that while pages with titles and paragraphs matching user questions were cited more frequently, technical features like FAQ blocks, summary boxes, structured data, and page speed had negligible impact on citation likelihood within a domain. These elements only showed effectiveness when comparing entirely different websites, suggesting they help with initial discovery but do not influence the model's decision to cite a specific page over its peers.

Sub-Queries and Semantic Matching

The primary driver of citations appears to be the engine's internal search process. The researchers replayed questions through the engines' APIs and discovered that every page cited by ChatGPT was retrieved by its own search mechanism. Crucially, the model rewrites queries into sub-queries that rarely coincide with Google’s or Bing’s results. These rewritten queries explain 32% of the variation in citation selection, whereas page properties explain only 2.8%. Content matters primarily because it matches the semantic intent of these rewritten sub-queries, not because it adheres to schema.org standards.

Brand Memory as a Ranking Factor

Beyond search retrieval, the model’s internal knowledge base plays a decisive role. The study found that ChatGPT tends to write brand names it already knows into its search queries and prefers websites it has encountered during training. Controlled experiments demonstrated that brand names can reorder rankings for products with identical factual content. In some cases, the model cites brands without even retrieving a supporting page, relying entirely on parametric memory. This behavior replicates in small B2B markets, indicating it is a systemic feature of current LLM architectures rather than an artifact of large consumer datasets.

Key Takeaways

  • FAQ blocks and structured data do not significantly increase citation rates for pages within the same website.
  • ChatGPT’s internal query rewriting explains 32% of citation variation, dwarfing the 2.8% explained by page properties.
  • Model memory and brand recognition directly influence search retrieval and citation decisions, sometimes bypassing retrieved pages entirely.
  • Traditional SEO metrics like page speed are irrelevant to the specific mechanism of AI citation selection.

The Bottom Line

The GEO industry is selling snake oil if they claim schema markup drives citations; the real game is semantic alignment with opaque model rewrites and establishing brand presence in the training data.