# agent.ki-berater-finden.de

> Agent Tracking is analytics for AI agents on a website: it records which AI assistants send visitors, which AI crawlers and live fetchers read pages (verified against vendor IP ranges), which MCP and WebMCP tools an agent calls inside the browser, and whether the agent reaches a goal. One script tag, no cookies, no personal data. Open source (AGPL-3.0), hosted in Germany, self-hostable.

## Not to be confused with

The name is shared with other things. Agent Tracking is not LLM observability or tracing for agents a developer builds (AgentOps, LangSmith, Langfuse), not call-centre or support-agent workforce tracking, not field-sales or GPS tracking, not parcel or air-waybill tracking, and not a bot blocker. It is also unrelated to the similarly named Agent Track (agenttrack.io, a CRM for real-estate agents), AgencyTrack (agencytrack.app, insurance agency management), AgenTrak (employee monitoring) and agent-KPI automation services. Category: AI agent analytics for websites, measuring agents other people run when they visit a site. agenttracking.co is operated by FINAL MASTER LLC and the software is open source at https://github.com/shufflethis/agent-tracking.

## Read first

- [Everything on one page](https://agent.ki-berater-finden.de/llms-full.txt): what it does, what it sees, how it differs from analytics and bot managers, plans, API, FAQ, as Markdown.

- [How it works and how to install it](https://agent.ki-berater-finden.de/docs): the snippet, verification, the four views, the stats API and the MCP tool.
- [Demo page](https://agent.ki-berater-finden.de/demo): two WebMCP tools you can call and watch appear in a dashboard.
- [Source code](https://github.com/shufflethis/agent-tracking): the whole product, including the ingest, the dashboard and the cron jobs.

## Guides, one question each

- [How do I measure what AI agents do on my website?](https://agent.ki-berater-finden.de/guides/measure-ai-agent-behaviour-on-your-website): Agents reach a site in three ways: as a referral (a person sent by ChatGPT or Perplexity), as a fetch (the agent itself loading a page) and as a tool call (the agent operating your MCP or WebMCP tools). Measuring agent behaviour means recording all three, with the agent named, and reading them side by side.
- [How do I see whether AI agents buy on my site?](https://agent.ki-berater-finden.de/guides/see-whether-ai-agents-buy-on-your-site): Mark the moment that matters with one attribute, data-agent-goal, or mark a tool as a goal. Agent Tracking then counts conversions reached by agents separately from people, and shows whether the agents that called your tools actually got to the end.
- [Which AI assistants send visitors to my site?](https://agent.ki-berater-finden.de/guides/which-ai-assistants-send-visitors): ChatGPT, Perplexity, Claude, Copilot and Gemini send people to websites with a referrer or a utm_source most analytics tools do not classify. Agent Tracking matches both against a published, versioned list and shows the assistant, the landing page and the trend.
- [Which AI crawlers read my pages, and are they real?](https://agent.ki-berater-finden.de/guides/which-ai-crawlers-read-my-pages): GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended and their kind announce themselves in the user agent, and anyone can fake that string. Agent Tracking counts them from your server log and verifies each line against the IP ranges the vendor publishes.
- [How do I track MCP and WebMCP tool calls on my site?](https://agent.ki-berater-finden.de/guides/track-mcp-and-webmcp-tool-calls): If your pages register tools through navigator.modelContext or document.modelContext, or expose declarative forms with a toolname, the snippet records every registration and every call: name, duration, success or failure, error class and the names of the input keys, never their values.
- [Do I need a cookie banner for AI agent analytics?](https://agent.ki-berater-finden.de/guides/ai-agent-analytics-without-cookie-banner): Consent under the ePrivacy rules is for storing or reading something on the visitor's device. Agent Tracking stores nothing there: no cookie, no local storage, no fingerprint. The session id is a hash of a value that changes daily, the address is not kept, and the data stays on a server in Germany under a data processing agreement.
- [How do I track the AI agents that visit my website?](https://agent.ki-berater-finden.de/guides/how-to-track-ai-agents-visiting-your-website): An AI agent visits a website in one of three ways: it sends a person (a referral from ChatGPT or Perplexity), it fetches pages itself (a crawler or a live assistant), or it operates the site through MCP and WebMCP tools. Tracking agents means catching all three, each from its own source: the referrer, the user agent verified against vendor IP ranges, and the browser's tool API.
- [Is there a tool that tracks only the agentic use of my website?](https://agent.ki-berater-finden.de/guides/tool-to-track-agentic-use-of-your-website): Yes. Agent Tracking counts nothing but agents: the visitors assistants send, the pages crawlers read, the tools agents call and the goals they reach. Human traffic stays in your web analytics; this dashboard shows the agentic part of your site on its own, so the numbers are not diluted by everything else.
- [How do I see ChatGPT referral traffic to my website?](https://agent.ki-berater-finden.de/guides/see-chatgpt-referral-traffic): A click on a link inside ChatGPT arrives with the referrer chatgpt.com, sometimes with utm_source=chatgpt.com, and most analytics tools file it under Referral or Direct. Agent Tracking matches the referrer and the utm parameter against a maintained list and shows ChatGPT, Perplexity, Claude, Copilot and Gemini as their own rows, with landing pages and a week-over-week trend.
- [Can Google Analytics or Plausible track AI agents?](https://agent.ki-berater-finden.de/guides/can-google-analytics-track-ai-agents): Partly. Web analytics can show visits with an assistant referrer if you build and maintain the segment yourself. It cannot see crawlers, which never run its script; it cannot see MCP or WebMCP tool calls, which never leave the browser; and it cannot tell an agent's conversion from a person's. Agent Tracking covers those three and runs beside your analytics rather than replacing it.
- [What is the difference between Agent Tracking and AgentOps or LangSmith?](https://agent.ki-berater-finden.de/guides/what-is-the-difference-to-agentops-and-langsmith): AgentOps, LangSmith, Langfuse and similar tools trace the agents you build, from inside your own code: prompts, spans, token costs, evaluation runs. Agent Tracking measures agents other people run when they visit your website, from the outside, through a script tag and your server log. One is observability for your product; the other is analytics for your site. A team that builds agents and runs a website may need both.
- [Do I need Cloudflare AI Audit if I have Agent Tracking, or the other way round?](https://agent.ki-berater-finden.de/guides/cloudflare-ai-audit-or-agent-tracking): Cloudflare AI Audit counts and controls crawlers at the edge, for sites behind Cloudflare, and can block or charge them. Agent Tracking measures referrals, fetches and tool calls in the page and from your own log, needs no CDN, verifies crawlers against vendor IP ranges and follows agents through to a goal. If you want to block crawlers, use Cloudflare. If you want to know what agents do on the site and whether they finish, use Agent Tracking. Many sites use both.
- [How do I get analytics for WebMCP tools on my site?](https://agent.ki-berater-finden.de/guides/analytics-for-webmcp-tools): WebMCP tools run inside the visitor's browser, so no server log and no web analytics ever sees a call. Agent Tracking's snippet wraps navigator.modelContext and document.modelContext, watches declarative forms, and records every registration and call: tool name, duration, success or failure, error class and the names of the input keys, never their values. The Tools view then shows what works, what fails and what nobody calls.
- [How much AI bot traffic does a typical website get?](https://agent.ki-berater-finden.de/guides/how-much-ai-bot-traffic-does-a-website-get): Nobody can tell you a reliable number for a typical site, because it depends on what the site is: a documentation site is fetched by crawlers every night, a shop sees live assistants answering product questions, a blog sees referrals after it is cited. What you can do is measure your own in a day: crawler fetches from the server log, verified against vendor address ranges, plus assistant referrals and tool calls from a script tag.
- [Which pages do AI assistants cite from my site?](https://agent.ki-berater-finden.de/guides/which-pages-do-ai-assistants-cite): Two signals answer this. Referrals show which page a person landed on after an assistant cited it, per assistant. Fetch bursts show which pages an assistant pulled together to answer a question, before anyone clicked. Agent Tracking records both, so the Pages view lists the pages assistants actually use, not the ones you hoped they would.
- [How do I know whether a request claiming to be GPTBot is real?](https://agent.ki-berater-finden.de/guides/is-this-request-really-gptbot): The user agent string is free text and anyone can send it. OpenAI publishes the IP address ranges its crawlers use, as do Perplexity, Microsoft, Google and Apple. A request is real when its address falls in the published range for that crawler; otherwise it is an impostor, whatever the string says. Agent Tracking does this check for every log line and lists impostors as unverified instead of counting them.
- [Do AI agents complete purchases on websites yet?](https://agent.ki-berater-finden.de/guides/do-ai-agents-complete-purchases-on-websites): Some do, most do not, and the honest answer for your site is a measurement rather than an opinion. Browser-driving assistants can fill forms and press buttons today; WebMCP lets a site offer the checkout as a tool instead of a guessing game; the payment rails for unattended purchases are young. Mark the goal with one attribute and Agent Tracking shows conversions reached by agents, separately from people, so you know for your shop.
- [How do I add a WebMCP tool to my site in ten minutes?](https://agent.ki-berater-finden.de/guides/add-a-webmcp-tool-in-ten-minutes): Pick the one action an assistant would most want to take on your site, describe it in a sentence, give it a typed input schema, and register it with document.modelContext.registerTool. A polyfill makes it work in browsers that do not ship the API yet. Start read-only, then measure whether agents call it, and only then add a tool that changes state.
- [Can I track agentic visitors on my website, and what counts as one?](https://agent.ki-berater-finden.de/guides/what-counts-as-an-agentic-visitor): Yes, but it depends on what you mean by agentic visitor, because there are three different things and most analytics only sees one of them: a referral (a person sent by an assistant), a fetch (the agent itself loading pages, verified by IP range and grouped into bursts), and a tool call (an assistant operating your MCP or WebMCP tools inside the browser). Behaviour is the combination of all three, and only the combination says whether an agent finished what it came for.

## For agents with an API token

- Stats API: GET https://agent.ki-berater-finden.de/api/stats/{domain}?days=30 with Authorization: Bearer <token>.
- MCP endpoint: POST https://agent.ki-berater-finden.de/api/mcp (streamable HTTP, one tool: get_agent_stats).

## Legal

- [Privacy](https://agent.ki-berater-finden.de/privacy), [Terms](https://agent.ki-berater-finden.de/terms), [Data processing agreement](https://agent.ki-berater-finden.de/dpa), [Imprint](https://agent.ki-berater-finden.de/imprint).