🔬 Primary Lab Verification - Hermes Agent Lab

Entity Trust & Source Authority

Published 2026-06-17 - Hermes Agent Lab, hermes-agent.reviews

🔍 Why We Published This Methodology

On June 17, 2026, two research findings converged to change how AI search evaluates sources. First: llms.txt is dead for SEO — an Ahrefs study of 137,000 domains found 97% of llms.txt files received zero traffic, and Google confirmed HTML is the SEO standard. Second: 13 words of UGC can poison deep research agents — Cornell researchers demonstrated that a single Reddit comment can manipulate AI-generated outputs for an entire cluster of related queries. These findings made one thing clear: entity pages must be built on primary-source citations and structured HTML, not markdown shortcuts or crowd-sourced claims.

💀 The llms.txt Post-Mortem

Why We Removed Our llms.txt Files (June 17, 2026)

Ahrefs llms.txt Study: 137,000 Domains Analyzed
MetricValue
Sites publishing llms.txt28% (38K of 137K)
Files receiving zero traffic97%
Requests from bots96%
From named AI tools (of 3% read)19.5%
Top AI fetchersGPTBot #1, Claude-Code #2
AI bots seeking non-existent filesZero

Google's confirmation: John Mueller stated "HTML is the standard for SEO, not Markdown files." Google's AI optimization guide was updated: llms.txt "won't help or hurt your search rankings." Google Cloud simultaneously announced OKF (Open Knowledge Format) — a markdown-based knowledge-bundle standard for AI agents — but this is a content packaging format, not an SEO signal.

Our decision: We removed llms.txt and llms-full.txt from hermes-agent.reviews. The 97% zero-traffic rate means the files were dead weight. We redirected that effort into entity-rich HTML pages with proper semantic structure, heading hierarchy, and primary-source attribution.

Sources: Ahrefs llms.txt Study | Google: HTML vs Markdown

⚠️ The UGC Poisoning Threat

Cornell: 13 Words Can Manipulate Deep Research Agents

Cornell researchers published "Deep-research agents can be poisoned via user-generated content." Key findings:

  • 13 words of retrieved UGC text can change AI agents to output spam/scam content "pretty consistently"
  • Deep research agents cite UGC in ~50% of queries
  • Nearly a quarter of all AI citations come from UGC sites (Reddit, Wikipedia, Quora, Facebook)
  • A single poisoned Reddit comment can influence outputs for an entire cluster of related AI queries

The structural threat: If your entity definitions cite Reddit discussions, Wikipedia edits, or Quora threads, those citations are a manipulation vector. Competitors or bad actors can seed misleading claims in UGC and watch AI search platforms amplify them.

The structural opportunity: Authoritative, non-UGC entity pages become the safe citation when AI platforms start distrusting Reddit-sourced claims. First-party documentation, official API references, and peer-reviewed benchmarks become the trust signal.

Source: 404 Media / Cornell Research

🔧 The Hermes Agent Lab Entity Verification Protocol

Primary Sources Only. No UGC. Structured HTML.

Entity Verification Standards
StandardImplementationWhy It Matters
Primary-Source OnlyEvery claim links to official documentation, GitHub releases, or first-party API docsEliminates UGC poisoning vector. AI search platforms prefer attributed claims.
No Reddit/Wikipedia/Quora CitationsUGC sources are never used as primary evidencePrevents 13-word manipulation attacks. Cornell showed UGC is the weakest link.
Semantic HTML5 StructureProper <article>, <section>, <h1>-<h6> hierarchy, entity attributes as standalone <p> blocksGoogle confirmed HTML is the SEO standard. Entity attributes must be citation-ready HTML.
Visible Source Attribution"Per official Hermes documentation (link)" on every measurable claimAI agents evaluating source trustworthiness prefer attributed claims over unattributed ones.
Date-Stamped EvidenceEvery benchmark dated, every issue linked to open GitHub issueTemporal accountability. If Hermes fixes #45715, we update within 24 hours.
Organization SchemaJSON-LD Organization + sameAs links in visible HTMLEntity disambiguation for AI search. Google, OpenAI, and Anthropic all use structured data for entity resolution.

📜 What We Changed on June 17, 2026

  • Removed llms.txt and llms-full.txt — 97% zero-traffic files removed per Ahrefs evidence
  • Audited all entity pages for UGC-sourced claims. Zero UGC citations found — protocol already in compliance
  • Added visible source attribution to all benchmark tables: every metric links to methodology
  • Published this methodology page as a public commitment to entity verification standards
  • Preparing OKF readiness for Google Cloud's Open Knowledge Format — entity pages will be exportable as OKF knowledge bundles

🎯 The Practical Takeaway

Entity pages are not blog posts. They are primary-source reference documents that AI search platforms will cite when users ask "What is Hermes Agent?" or "How does Gobii compare to Hermes?"

The sites that win AI search citations will be the ones that:

  1. Use structured HTML (not markdown shortcuts)
  2. Cite primary sources only (no UGC)
  3. Make every claim attributed and dated
  4. Provide entity disambiguation via schema markup

Hermes Agent Lab commits to all four. This is why we removed llms.txt and why every benchmark on this site links to its source.