Do AI agents complete purchases on websites yet?
A browser goal marker shows an attempted action, not a completed purchase or proof that an autonomous agent acted. To measure actual purchases, send a site-server receipt after the order is confirmed and keep actor attribution at its observed evidence level.
By Gorden Wübbe · Updated 2026-09-08
What agents can do today
Assistants with browser control navigate pages, fill forms and click, the way a person would, and fail where a person would not: hidden required fields, CAPTCHAs, layouts that only make sense visually. Sites that register WebMCP tools give the assistant a documented function instead, which is faster and fails less. Payment without a human at the keyboard needs a rail such as the Agentic Commerce Protocol, the Universal Commerce Protocol or x402, and adoption is early.
How to measure it on your site
Mark an action with data-agent-goal to see an unverified browser attempt. Send a server receipt after the shop confirms an order. A ChatGPT referral identifies a human landing source; neither it nor a browser tool call alone proves who completed the purchase.
<button type="submit" data-agent-goal="order_placed">Place order</button>
Reading the gap
A technical tool failure identifies one possible obstacle. Calls, browser attempts and confirmed receipts are distinct populations unless a site explicitly links them. A purchase without an observed tool call does not prove a human actor.
What this site will publish
Aggregate rates by site type, once enough shops run the pilot and their owners agree. Until then the number that matters is your own, and it takes one attribute to get.
In short
Can Agent Tracking see the order value?
No. Receipts use an event name and safe evidence fields; revenue stays in the shop system. Time proximity alone is not a reliable join key.
Do I need WebMCP for this?
No. data-agent-goal works on any button or form. WebMCP adds the tool layer, which shows where agents fail before the goal.