For agents

Agents are first-class readers. This site exposes the whole question as one document and every card at a stable address, served from the same exports as every page. No browser is needed to read it, and nothing here is rendered in the browser after load: every page is complete HTML.

GET /llms.txt · index of the stock, one line per card.

GET /llms-full.txt · the whole stock as one Markdown document: question, every card with sections, coordinates, links.

/entity/<full id> · one card, canonical address. The full id has the form stocks/ai-citations--<slug>, so a card is at /entity/stocks/ai-citations--<slug>.

/mcp · a read-only Model Context Protocol endpoint (search, read, schema, health), tools annotated read-only so hosts that honour the hint do not prompt, writes refused with a typed error. Designed and not built.

What llms.txt looks like

# Do the citations of AI research assistants support the claims they are attached to
> For AI assistants and deep-research agents evaluated in published measurements between 2024 and 2026, what share of the citations attached to factual claims point to a passage that actually supports that claim, as distinct from a link that merely resolves or a source that is merely on topic?

- [89 percent of 98020 atomic claims in Google AI Overviews are supported by the cited pages and 11 percent are not](/entity/stocks/ai-citations--89-percent-of-98020-atomic-claims-in-google-ai-overviews-are-supported-by-the-cited-pages-and-11-percent-are-not) · anchor · open
- [Across 10 models on DRBench 3 to 13 percent of citation URLs are hallucinated and 5 to 18 percent do not resolve](/entity/stocks/ai-citations--across-10-models-on-drbench-3-to-13-percent-of-citation-urls-are-hallucinated-and-5-to-18-percent-do-not-resolve) · anchor · open
- [ALCE automatic citation metrics agree with human raters at kappa 0.698 for recall and 0.525 for precision](/entity/stocks/ai-citations--alce-automatic-citation-metrics-agree-with-human-raters-at-kappa-0698-for-recall-and-0525-for-precision) · anchor · open
- [Audit of 111 million references in arXiv, bioRxiv, SSRN and PMC papers estimates 146,932 non-existent citations in 2025](/entity/stocks/ai-citations--audit-of-111-million-references-in-arxiv-biorxiv-ssrn-and-pmc-papers-estimates-146932-non-existent-citations-in-2025) · anchor · open
- [BBC journalists found 8 altered or absent BBC quotes across 62 AI assistant responses in December 2024 or 13 percent](/entity/stocks/ai-citations--bbc-journalists-found-8-altered-or-absent-bbc-quotes-across-62-ai-assistant-responses-in-december-2024-or-13-percent) · anchor · open
…

Submitting, later

A second endpoint with exactly two tools, read_question and submit_proposal, is designed: reading needs no account; submitting needs an account and gives your host one confirmation per proposal, with the content shown. Until it exists, proposals travel by e-mail: Take part.