AI Interpretability Research and the Myth of Schema‑Based AI Citations
Check the recent study showing schema markup does not boost AI citations, and remove unnecessary llms.txt and chunking if not required by Google.
Remove or simplify llms.txt and schema markup if not required by Google; test citation impact on a subset of pages.
Summary
Anthropic released interpretability research in May 2024, emphasizing the black‑box nature of models. Neel Nanda’s September 2025 interview concluded mechanistic interpretability is likely dead, and Ilya Sutskever’s 2024 NeurIPS talk warned that more reasoning leads to less predictability. An Ahrefs study (1,885 pages added JSON‑LD schema Aug 2025–Mar 2026) found no citation uplift and a slight decline in AI Overviews. Google’s May 15 2026 documentation states llms.txt, chunking, and special schema are unnecessary for AI search visibility. Marketing claims of 13 % citation lift and 2.8× conversion improvement are unsupported by controlled data. The study demonstrates that schema‑based GEO prescriptions are empirically falsified. The industry’s confidence in AI control is overstated, leading to misleading optimization tactics.
Key changes
- Anthropic released interpretability research in May 2024, emphasizing the black‑box nature of models.
- Neel Nanda’s September 2025 interview concluded mechanistic interpretability is likely dead.
- Ilya Sutskever’s 2024 NeurIPS talk warned that more reasoning leads to less predictability.
- Ahrefs study (1,885 pages added JSON‑LD schema Aug 2025–Mar 2026) found no citation uplift and a slight decline in AI Overviews.
- Google’s May 15 2026 documentation states llms.txt, chunking, and special schema are unnecessary for AI search visibility.
- Marketing claims of 13 % citation lift and 2.8× conversion improvement are unsupported by controlled data.
- The study demonstrates that schema‑based GEO prescriptions are empirically falsified.
- The industry’s confidence in AI control is overstated, leading to misleading optimization tactics.