Identify AI Agent Traffic in GA4 | UCP Analytics

The core problem: GA4 reports AI agent traffic as Direct, Bot, or Referral from chatgpt.com — with no way to tell which sessions are humans and which are agents crawling your content. This guide fixes that with four concrete implementations. See also: 2026 UCP Compliance and Risk Checklist for Merchants. For related reading, see WebMCP Is Here.

Why This Matters in 2026

If your site participates in agentic commerce via UCP, MCP, or any AI shopping integration, a meaningful percentage of your traffic is not human. AI agents from ChatGPT, Perplexity, Claude, Google’s Gemini, and custom enterprise agents are crawling your product data, checking inventory, and initiating checkout flows. GA4, as configured out of the box, cannot distinguish them from human sessions.

This creates two problems. First, your engagement metrics are diluted — a 2-second “session” from an agent crawling your /.well-known/ucp endpoint looks identical to a human who bounced. Second, you’re missing a critical growth signal: agent traffic volume is one of the earliest indicators of whether your UCP implementation is being discovered and used. You want to measure it, not lose it in the noise.

Method 1: Referral Source Segmentation (No Code, 5 Minutes)

The fastest win. AI platforms that send users to your site appear in GA4 as referral traffic from identifiable domains. In GA4, go to Reports → Acquisition → Traffic Acquisition and filter by Session source. Look for:

• chatgpt.com — ChatGPT users who clicked a link from a ChatGPT response
• perplexity.ai — Perplexity answer citations that drove a click-through
• claude.ai — Claude users referred to your content
• storage.googleapis.com — Often indicates automated agent workflows or Gemini staging environments pulling your content
• cn.bing.com — Copilot traffic (Microsoft’s AI) routed through Bing’s China CDN node

Create a saved comparison in GA4 (Explore → Free Form) with chatgpt.com vs google / organic as your segment split. Compare which pages each source visits. AI-referred users will show a completely different page distribution than organic search — they skip landing pages and go directly to deep technical content.

Limitation: This only captures AI traffic that clicked a link. Agents crawling your site directly (hitting your UCP endpoint, fetching product data) appear as Direct traffic or don’t appear at all if they don’t execute JavaScript.

Method 2: Custom Dimension for Agent-Origin Requests (Intermediate)

AI agents identify themselves in HTTP request headers. The User-Agent string for known agents follows identifiable patterns. Add a server-side check that tags requests from known agent user-agents and fires a custom GA4 event.

Known AI agent User-Agent patterns to detect (2026):

• GPTBot — OpenAI’s crawler
• ChatGPT-User — ChatGPT browsing agent
• Claude-Web — Anthropic’s web-access agent
• Googlebot + Google-Extended — Google’s AI training crawler (separate from standard Googlebot)
• PerplexityBot — Perplexity’s indexing crawler
• anthropic-ai — Anthropic API requests
• cohere-ai — Cohere’s crawler

Implementation via Google Tag Manager:

1. In GTM, create a new Variable → Custom JavaScript that reads navigator.userAgent and returns a boolean for known agent patterns
2. Create a Trigger that fires on All Pages when your agent-detection variable returns true
3. Create a GA4 Event Tag: event name ai_agent_session, parameter agent_type with value from your detection variable
4. In GA4, create a Custom Dimension: agent_type scoped to Session

Limitation: Client-side UA detection is unreliable for agents that don’t execute JavaScript (most crawlers). This catches browser-rendered agent sessions, not raw API calls to your UCP endpoint.

Method 3: Server-Side Logging for UCP Endpoint Requests (Recommended for Agentic Commerce)

Your /.well-known/ucp endpoint is the most reliable signal. Every agent that is evaluating your store for agentic commerce will hit this endpoint. Log every request to it at the server level, then pipe that data into GA4 via the Measurement Protocol.

What to log per request:

• Timestamp
• IP address (anonymized to /24 for GDPR)
• User-Agent string (full, for agent identification)
• Requested endpoint path
• Response code (200, 404, 429)
• Request frequency (rate — is this a one-time check or a polling agent?)

Piping to GA4 via Measurement Protocol:

POST https://www.google-analytics.com/mp/collect?measurement_id=G-XXXXXXXX&api_secret=YOUR_SECRET

{
  "client_id": "agent-{{hashed_ip}}",
  "events": [{
    "name": "ucp_endpoint_hit",
    "params": {
      "agent_ua": "GPTBot/1.0",
      "endpoint": "/.well-known/ucp",
      "response_code": 200,
      "session_id": "{{timestamp}}"
    }
  }]
}

This creates a dedicated event stream in GA4 for agent endpoint activity, completely separate from human session data. You can then build a GA4 Exploration report that shows agent request volume over time — this is your UCP adoption metric.

Method 4: Server-Side GTM for Privacy-Conscious Audiences (Advanced)

Your technical audience — Linux users on Chrome, Sunday-night developers, VPN users — is systematically stripping your client-side tracking. Server-Side Google Tag Manager (sGTM) solves this by moving event collection from the browser to your server, bypassing ad blockers and browser privacy restrictions.

Setup overview:

1. Deploy a sGTM container on Cloud Run (GCP) or any server you control — costs roughly $5–15/month at your traffic volume
2. Point your first-party measurement endpoint (e.g., analytics.yoursite.com) to the sGTM container
3. In your GTM web container, change the GA4 configuration tag to send to your first-party endpoint instead of google-analytics.com directly
4. The sGTM container proxies the request to GA4, but the browser only sees a first-party domain request — which privacy tools do not block

Why this matters for your site specifically: The GA4 data showed a high proportion of Linux/Chrome users and evidence of privacy tool usage (referral strings being stripped, form events not firing). Server-side GTM is the most reliable fix for this specific audience profile. It typically recovers 15–30% of sessions that were previously invisible.

Building Your AI Agent Traffic Dashboard

Once your tracking is in place, build a GA4 Exploration with these dimensions and metrics:

What Normal Agent Traffic Patterns Look Like

Understanding what you’re looking at once tracking is in place:

• UCP endpoint polling: Regular hits every 30–60 minutes from the same IP range. This is a configured agent keeping your product data fresh. It’s a good sign — it means an agent has your store in its catalog.
• Content crawler burst: 5–20 page requests in rapid succession, no JavaScript execution, sequential URL patterns. This is a training or indexing crawler. Log it but it doesn’t convert to purchases.
• Referral from chatgpt.com with long session: A human user who asked ChatGPT a question, got your page cited, and clicked through. This is your best lead — they’ve already been pre-qualified by an AI that selected your content as authoritative.
• storage.googleapis.com referral: Usually Gemini’s automated research pipeline or a developer testing a Vertex AI workflow against your site. Engagement will be short — this is a machine, not a human.

Related Resources

• UCP Technical Specification — /.well-known/ucp Endpoint
• Configure Your /.well-known/ucp Discovery Endpoint
• UCP vs ACP vs MCP: Protocol Comparison
• Optimize Your Google Merchant Center Feed for AI Mode
• Where Can You Use Agentic Commerce Today?