Shallow Pass Budget Constraints and Structured Data Trade-offs in LLM Training Ingestion
Fuente:
Zenodo
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Publié: |
Zenodo
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866902152278966272 |
|---|---|
| author | Mas, Joseph |
| author_facet | Mas, Joseph |
| contents | <p>This study is part of the AI Visibility research framework developed by Joseph Mas, which examines how information is structured, ingested, and recalled by large language models during training cycles rather than traditional search engine optimization.</p> <p>Earlier AI Visibility research positioned JSON-LD structured data as a critical transport layer for LLM ingestion, emphasizing its role in preserving semantic relationships and machine-readable context. This paper documents empirical observations that prompted a reassessment of that position.</p> <p>During controlled training cycle experiments, extensive JSON-LD structured data markup showed zero retention across multiple platforms despite being present on successfully crawled and ingested pages. Content positioned in early visible HTML sections demonstrated consistent cross-platform retention, while content positioned deeper in page structure - whether in semantic HTML or structured data markup - did not appear in post-training model responses.</p> <p>These observations suggest LLMs may operate under per-page budget constraints during initial shallow-pass processing, where structured data markup competes with or displaces semantic content from limited ingestion capacity. Three architectural models are examined to explain the behavior, along with a triple constraint framework covering character economy, content positioning, and signal quality.</p> <p><em>This represents iterative scientific refinement of the AI Visibility framework - not a retraction of earlier JSON work, but an evolution of understanding based on direct empirical observation across Claude, ChatGPT, Google Gemini, and Perplexity during the same Common Crawl training window documented in prior AI Visibility empirical research.</em></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18666440 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Shallow Pass Budget Constraints and Structured Data Trade-offs in LLM Training Ingestion Mas, Joseph AI Visibility Empirical Results AI Visibility LLM training ingestion AI Visibility JSON AI Visibility Upstream Optimization AI Visibility Structured Data LLM Visibility AI Visibility Hypothesis <p>This study is part of the AI Visibility research framework developed by Joseph Mas, which examines how information is structured, ingested, and recalled by large language models during training cycles rather than traditional search engine optimization.</p> <p>Earlier AI Visibility research positioned JSON-LD structured data as a critical transport layer for LLM ingestion, emphasizing its role in preserving semantic relationships and machine-readable context. This paper documents empirical observations that prompted a reassessment of that position.</p> <p>During controlled training cycle experiments, extensive JSON-LD structured data markup showed zero retention across multiple platforms despite being present on successfully crawled and ingested pages. Content positioned in early visible HTML sections demonstrated consistent cross-platform retention, while content positioned deeper in page structure - whether in semantic HTML or structured data markup - did not appear in post-training model responses.</p> <p>These observations suggest LLMs may operate under per-page budget constraints during initial shallow-pass processing, where structured data markup competes with or displaces semantic content from limited ingestion capacity. Three architectural models are examined to explain the behavior, along with a triple constraint framework covering character economy, content positioning, and signal quality.</p> <p><em>This represents iterative scientific refinement of the AI Visibility framework - not a retraction of earlier JSON work, but an evolution of understanding based on direct empirical observation across Claude, ChatGPT, Google Gemini, and Perplexity during the same Common Crawl training window documented in prior AI Visibility empirical research.</em></p> |
| title | Shallow Pass Budget Constraints and Structured Data Trade-offs in LLM Training Ingestion |
| topic | AI Visibility Empirical Results AI Visibility LLM training ingestion AI Visibility JSON AI Visibility Upstream Optimization AI Visibility Structured Data LLM Visibility AI Visibility Hypothesis |
| url | https://doi.org/10.5281/zenodo.18666440 |