Can Google Analytics or Plausible track AI agents?

What web analytics sees

A person arriving from chatgpt.com or perplexity.ai is a normal visit with a referrer. GA4 can analyze that source; Plausible now has an AI Assistants channel with individual sources, landing pages and conversions. A missing referrer still cannot be reliably reconstructed. Agent Tracking also attributes identifiable referrals but does not replace a full human analytics view.

What requires extra instrumentation

GPTBot and similar crawlers usually fetch HTML without executing a browser analytics script; their requests require server or edge logs. A WebMCP call is a browser event that needs explicit tracking. Both GA4 and Plausible can record custom events when configured, but neither automatically identifies every on-page model-context call. Attribution of a completed goal depends on the available session and event signals; an AI-referred human purchase should remain distinct from an autonomous agent action.

What Agent Tracking adds

Origin-log fetches with supported IP verification and burst grouping, instrumented tool calls with technical outcomes, browser goal attempts, separate server-confirmed outcomes, and identifiable assistant referrals. It does not record raw tool input values.

Use the right source for each event

Use web analytics for visitors and conversions, server logs for crawlers, and specific event instrumentation for browser tools. The tools can overlap on identifiable AI referrals; keep definitions consistent instead of adding their totals together. See the full AI traffic measurement guide for a repeatable setup.

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

Compare the questions each product answers

For human acquisition, GA4 and Plausible are mature choices. They can show where identifiable visitors came from, which pages they landed on and which configured goals they completed. Plausible explicitly groups known assistant sources under AI Assistants, and GA4 can be configured to report these sources through its traffic-source dimensions. Agent Tracking also shows assistant referrals, but its primary purpose is to put them beside requests from crawlers and actions by agents. It is not intended to replace the broad reporting of a general web analytics tool.

For crawler activity, a browser tag is the wrong data source. Google says GA4 automatically excludes known bots, and Plausible filters known bots too. That is sensible for human analytics; it means a crawler report needs server logs or an edge service. Agent Tracking imports origin logs. A CDN such as Cloudflare can see requests handled at the edge, including some that never reach the origin.

For WebMCP actions, ask whether the tool call and its outcome are recorded automatically. A custom event in GA4 or Plausible can count a call when the site sends one, but somebody must write and maintain that instrumentation. Agent Tracking's snippet observes supported browser tool APIs and can pair calls with goal markers. Remote MCP servers require their own server-side logs; an on-page snippet cannot inspect them.

A concrete example with three numbers

Suppose a product page receives eight identifiable ChatGPT referrals this week. A configured analytics report can show those visits and goals. The server log also contains 120 requests claiming to be GPTBot, of which 95 match the operator's current published ranges. An availability tool was called six times with one validation error; two separate site-server receipts confirm reservations.

The eight visits, 95 verified fetches and six tool calls are separate populations. Do not add them to report 109 AI visitors. The bot requests may never have led to a visible answer; the referrals may come from answers generated from other sources; and the tool calls may occur during visits classified in several ways. The useful conclusions are narrower: the page received identifiable assistant traffic, a verified bot fetched it, and one tool error needs attention.

A user who clicks through from ChatGPT is an AI-referred human. A browser tool call is an observed tool interaction, not proof of the caller. The six calls and two server receipts have different denominators unless individually linked. Instrument the final outcome before claiming a conversion rate.

Where GA4 and Plausible are stronger

GA4 and Plausible cover the rest of the human visitor journey: campaigns, landing pages, engagement and configured outcomes. Plausible is especially straightforward when the question is simply which assistants send people and whether those people convert; it already maintains the AI channel. GA4 can be a better fit if the organization relies on its broader reporting and advertising integrations. Prices, privacy settings and limits change, so compare their current official documentation and contract terms rather than a copied feature table.

If assistant referrals are the only question, start in the analytics product you already use. Introducing a second product for an eight-click sample may create more work than insight. Add log processing when crawler behavior matters, and browser tool instrumentation when agents can actually perform tasks on the site. This order makes the measurement proportional to the decisions you will make.

Where Agent Tracking adds a different view

Agent Tracking is designed for websites that need the three sources together: identifiable assistant referrals, verified crawler fetches from an uploaded log, and calls to on-page MCP or WebMCP tools. It displays verification status, fetch bursts, tool errors and goal events. The code is public under AGPL-3.0, can be self-hosted and has a hosted Free plan with limits. That makes its data flow inspectable, but self-hosting still requires a server and maintenance.

The practical choice is often coexistence. Leave GA4 or Plausible in place for human analytics. Add Agent Tracking when the team has a real question about crawler requests or agent actions that the existing dashboard cannot answer from its current data source. Use consistent dates and page definitions when comparing their referral counts; privacy settings, ad blockers and classification lists can cause differences.

In short