What it is and what the spec says
AgentReady classifies it as AR-DISC-03 (SHOULD) and AR-DISC-04 for the extended version (MAY). Not MUST. The official spec is at llmstxt.org: a markdown index at the root summarizing the product and linking the documents agents should read first. llms-full.txt is the same but ingestible in one request within the context window.
isitagentready detects it but doesn't score it as critical; its real weight is in 'Content Accessibility' and 'Discoverability', not as a citation magic wand.
The numbers — with traceability
| Fact | Observed data | Honest implication |
|---|---|---|
| llms.txt adoption top 1k | 8,7 % (jun 2026, n≈1.000, fuente: llmstxt.org + plan interno — sin dataset público aún) | File alone doesn't differentiate — don't use as authority badge |
| Google sobre llms.txt | Not used for ranking (declaración pública Google, 2026) | Don't sell as SEO |
| Real fetch belmon.tech | ≈0,12 % (logs propios 01–15 ago 2026, n=1 dominio, 15 días — no generalizable) | Site is the base, not the file |
| llms.txt → citation correlation | r ≈ 0,03 (estudio externo 300k dominios citado en plan — dataset y metodología pendientes de publicar) | Having the file doesn't predict being cited — near-zero correlation |
| What does correlate (observed) | Answer-first content (glosario belmon.tech: 1 definición autocontenida por concepto) | Predictable structure + sources = cited by LLM (see glossary) |
Honest rule: Implement llms.txt/llms-full.txt IF your audience are developers consuming docs via LLM (Vercel, Supabase, GitHub do). Don't promise GEO ranking. Available evidence (fetch n=1 + r≈0.03 pending replication) doesn't support it. What does move citations in our measurements is the answer-first glossary and markdown-for-agents (see isitagentready validation).
Methodology & how to replicate
- Fetch 0.12%: Nginx/Cloudflare logs from belmon.tech, 01–15 Aug 2026, fetch count / total hits filtering verified bots (GPTBot, ClaudeBot, PerplexityBot). Replicable on your own logs.
- r≈0.03: external 300k-domain study cited in agent readiness.md:28. No public dataset yet — marked as pending verification, not fact.
- Primary sources: agentready.org/spec.json AR-DISC-03/04, llmstxt.org, isitagentready.com.
What to do today
- Generate /llms.txt from your real navigation (not by hand) and keep it in sync with /llms-full.txt. Validate both serve as text/plain.
- Advertise them via Link header rel=describedby, not just as loose files.
- Measure fetch in logs and citations in ChatGPT/Perplexity (weekly sample). If they don't rise, invest in glossary + JSON-LD, not more llms.txt.
Primary source: AgentReady v1.0.0 spec AR-DISC-03/04 + isitagentready validation + belmon.tech logs.