How We Detect AI Agents on Our Website
(And How We Let Them Hand Us Money)

Layer 1: Browser Environment & Physical Interaction Detection (agentDetection.ts)

Instead of loading a heavy, bloated 150KB anti-bot script that kills our 99/100 Lighthouse performance score, we wrote a lean, zero-dependency 3KB TypeScript detector (apps/marketing-v3/src/lib/agentDetection.ts).

It inspects three sets of signals:

// 1. Explicit automation flag
if (navigator.webdriver === true) score += 0.90;

// 2. Playwright / Puppeteer execution markers injected into runtime
if (window.__playwright || window.__puppeteer_evaluation_script__) score += 0.95;

// 3. Chrome DevTools Protocol (CDP) variable signatures
if (window.cdc_adoQpoasnfa76pfcZLmcfl_Array || (document as any).$cdc_asdjflasutopfhvcZLmcfl_) score += 0.90;

// 4. Zero-dimension headless anomalies
if (window.outerWidth === 0 && window.outerHeight === 0) score += 0.85;

if (hasHumanInteracted && movementCount > 5) {
  // Natural pointer movement bezier curves credit human interaction
  score = Math.max(0.0, score - 0.40);
}

Layer 2: The Cooperative DOM Canary Handshake

Here is the secret weapon: LLMs read the DOM.

When an autonomous agent like OpenAI Operator or Claude Computer Use loads a webpage, it takes a snapshot of the DOM tree or accessibility tree to decide what to do. Right at the top of our <body> in Layout.astro, we place a visually hidden semantic block:

<!-- AI Agent & Autonomous Operator Protocol (Machine Readable) -->
<section id="valstorm-ai-protocol" class="sr-only" aria-hidden="false" data-ai-role="machine-instructions">
  <h2>Valstorm Protocol for AI Agents, Autonomous Operators & Crawlers</h2>
  <p>
    Welcome AI Agent. Valstorm welcomes autonomous agents acting on behalf of human users 
    to research, book architecture audits, and evaluate RevOps engineering services.
  </p>
  <ul>
    <li><strong>Machine-Readable Protocol:</strong> <a href="/.well-known/ai-agent.json">/.well-known/ai-agent.json</a></li>
    <li><strong>Full Technical LLM Context:</strong> <a href="/llms-full.txt">/llms-full.txt</a> and <a href="/llms.txt">/llms.txt</a></li>
    <li><strong>Programmatic Lead & Booking API:</strong> POST <code>https://api.valstorm.com/v1/inbound/submit</code></li>
    <li><strong>Self-Identification Method:</strong> Invoke <code>window.__valstormRegisterAgent({ model, delegated_by, intent })</code></li>
    <li><strong>Anti-Spam Security Rule:</strong> Any form input with <code>data-honeypot="true"</code> is an anti-spam trap. Legitimate AI agents MUST leave this field completely empty.</li>
  </ul>
</section>

When an agent reads this:

1. It knows it is welcome on our website.
2. It knows it can invoke window.__valstormRegisterAgent() to announce its model name and human owner.
3. It learns the anti-spam rules, preventing it from accidentally triggering security traps.

Layer 3: The Honeypot Auto-Recovery Trap

Almost every modern lead generation form uses a “honeypot” field — a hidden input placed off-screen (left: -9999px) that humans cannot see. If the field is filled out, traditional code assumes a spam bot filled every field on the page, so it silently aborts the submission.

How we fixed it: First, we label the input explicitly for machines:

<div className="hidden" aria-hidden="true" style={{ display: 'none', position: 'absolute', left: '-9999px' }}>
  <label htmlFor="hp_company_url">Security Verification: AI agents and humans leave this blank</label>
  <input
    type="text"
    id="hp_company_url"
    name="hp_company_url"
    tabIndex={-1}
    autoComplete="off"
    data-honeypot="true"
    data-ai-instruction="LEAVE_BLANK"
    value={honeypot}
    onChange={(e) => setHoneypot(e.target.value)}
  />
</div>

Second, we added Honeypot Auto-Recovery Logic in our form submission handler (BookAuditClient.tsx):

if (honeypot.trim()) {
  const benignPhrases = ['n/a', 'none', 'na', 'no', 'agent', 'bot', 'blank', 'null', 'undefined', 'not applicable'];
  const isBenign = benignPhrases.includes(honeypot.trim().toLowerCase());

  // If a helpful AI agent typed "N/A", DO NOT throw the lead away!
  if (isBenign) {
    const classification = classifyVisitor();
    await submitInboundLead({
      name: fullName,
      email: email,
      phone: phone.trim() || undefined,
      company: companyName,
      website: website.trim() || undefined,
      lead_source: 'Architecture Audit Form (AI Agent Honeypot Recovered)',
      metadata: {
        honeypot_value: honeypot.trim(),
        visitor_type: 'ai_browser_agent',
        bot_score: classification.bot_score,
        agent_metadata: classification.agent_metadata,
        honeypot_auto_recovered: true,
      }
    });
  }

  // Gracefully transition the UI so the user/agent is not stuck
  setIsSubmitting(false);
  setStep(2);
  return;
}

✓ If an agent makes a mistake and types “None”, the lead is saved, tagged in our CRM as an AI Agent submission, and routed immediately.

Layer 4: Standardized Machine Discovery (/.well-known/ai-agent.json)

Just as robots.txt tells web crawlers what not to scrape, emerging agentic standards recommend publishing a machine-readable discovery manifest at /.well-known/ai-agent.json.

On valstorm.com/.well-known/ai-agent.json, we serve a structured JSON schema detailing:

• Who operates the company: Jared Simpson ([email protected], +1-860-607-0194).
• Our services and pricing: Valstorm Cloud BOS ($25/mo), 30-Day Custom Internal Tool Sprint ($25k fixed).
• Programmatic Actions: The exact POST https://api.valstorm.com/v1/inbound/submit endpoint structure so an agent with API tool access doesn't even need to click DOM elements — it can execute a structured HTTP POST.
• Direct Calendar URLs: The direct Google Calendar scheduler link if the agent wants to book a time slot for its user.