How do I measure what AI agents do on my website?

What counts as agent behaviour

A crawler reading 400 pages overnight, a live assistant fetching three pages in four seconds to answer a question, a person arriving from a chat, and an assistant calling a tool inside the browser are four different behaviours with four different meanings. Ordinary analytics folds the first three into one line labelled Direct or Other and never sees the fourth.

To measure behaviour you need each event with three facts: which agent, which page or tool, and what happened next.

Step 1: one line of script

Add your site in the dashboard and put the snippet on every page. It records the page path (no query string), whether the visit came from or was made by an AI assistant, and every MCP or WebMCP tool registration and call. No cookies, no network address stored.

<script defer data-domain="example.com" src="https://agenttracking.co/agent.js"></script>

Step 2: your server log for crawlers

Most crawlers never run JavaScript, so the snippet cannot see them. Upload your access log on the settings page or send it daily from a cron with the API token. Each line is matched against the agent list and checked against the vendor's published IP ranges; an impostor claiming to be GPTBot from a random address is listed as unverified, not counted.

curl -sS -X POST https://agenttracking.co/api/logs/example.com \
  -H "Authorization: Bearer wmt_your_token" \
  -H "Content-Type: text/plain" --data-binary @/var/log/nginx/access.log

Step 3: read the four views

Overview shows referrals, fetches, tool calls and conversions per day with the trend against the previous period. Agents lists who sent visitors and who fetched pages, with share, trend, bursts and verification. Tools shows calls, success rate, average duration and top errors per tool. Pages shows what agents fetch and where tools are called.

A burst is one agent fetching several pages within seconds: the footprint of an assistant answering one question about you. Watching which pages a burst takes tells you what the assistant considered relevant.

Step 4: ask instead of browsing

Every number is available as JSON and as an MCP tool. Point Claude, ChatGPT or Cursor at it and ask which agents read your site this week and which tool failed most.

{ "mcpServers": { "agent-tracking": { "url": "https://agenttracking.co/api/mcp",
  "headers": { "Authorization": "Bearer wmt_your_token" } } } }

In short