Agentic Commerce Links: What AI Agents Get From Yours

AI agents are starting to shop and book for people, and a short link is one of the first things they meet. We read the Elido edge code to see what an agent actually receives. Known AI crawlers and assistants get a JSON description of the link instead of a redirect, and that response isn't counted as a click. Everything else follows the normal redirect, and scripted clients are filtered out of click analytics as bots. The payment protocols cover checkout and trust. They don't tell you who followed your links.

This sits next to answer engine optimization, which is about being cited by AI in the first place.

What agentic commerce is, from primary sources

Three open specs come up most. Each project's own page was read on 2026-10-25 and claim only what those pages say.

• Agentic Commerce Protocol (ACP). Its repository describes an open standard for connecting buyers, their AI agents and businesses. OpenAI and Stripe are listed as Founding Maintainers, the license is Apache 2.0, and the project is marked beta. The latest stable version it lists is 2026-04-17. OpenAI's commerce docs cover agentic checkout, delegated payment and product feeds for ChatGPT.
• Agent Payments Protocol (AP2). The AP2 site calls it an open protocol for the agent economy, built around signed "mandates" (checkout and payment) that form an audit trail. It says further standardisation continues in FIDO working groups.
• Universal Commerce Protocol (UCP). ucp.dev describes catalog search, cart, checkout, identity linking and order management, and says it supports AP2, A2A and MCP. Stripe's docs list it as the protocol they document under this heading.

The Visa and Mastercard agent programs weren't checked against their own pages, so this post says nothing about them. Partner counts and adoption figures weren't confirmed either, so you won't find any here.

All three cover discovery, cart, identity and payment. None says anything about the URL an agent was handed on the way in. Short, but it's the gap this post is about.

What an Elido link does when an agent follows it

This is how the edge redirect service behaves today, checked in services/edge-redirect:

1. Known AI agent User-Agent. The edge keeps a list of 15 signatures, including chatgpt-user, gptbot, claudebot, perplexitybot and amazonbot. A match returns JSON-LD with HTTP 200: the short URL, destination URL, title, link status and a safety score and abuse status. It's cacheable for five minutes and marked noindex.
2. Accept: application/json without text/html. Same JSON response, whatever the User-Agent. Appending + to a slug, or ?preview=1, gives a human-readable preview page instead.
3. Everything else. A normal redirect (302 by default, 301 or 307 per link). The click event is emitted unless the User-Agent matches the bot list. That list is a set of substrings: anything containing bot, crawler, spider, headless, curl/, python-requests, go-http-client, node-fetch, axios/ and more, plus an empty User-Agent.

Two consequences matter for anyone measuring agent traffic, and neither is obvious from the dashboard, which shows only counted clicks and says nothing about the visits it dropped.

• An agent that identifies itself as a known AI crawler never reaches your destination through the short link. It gets metadata. That's useful for an LLM that wants to know where a link leads before visiting, but the visit isn't a click in your analytics.
• An agent written as a script, or a headless browser that says so in its User-Agent, follows the redirect and then disappears from click analytics. An agent that drives a regular browser with an ordinary User-Agent looks like any other visitor.

The old version of this post claimed a short link records every machine visit. That was wrong. A short link does give you a stable address you can repoint without changing what the agent was given. It doesn't give you an agent-traffic report.

Check it yourself

I'd rather you test this than take our word for it. Request one of your own short links twice from a terminal:

curl -s -o /dev/null -w "%{http_code}\n" -A "Mozilla/5.0" https://go.yourbrand.com/slug
curl -s -H "Accept: application/json" https://go.yourbrand.com/slug

The first should print a redirect status (302 unless you changed it) and, because it looks like a browser, the visit is counted. The second returns the JSON description with a 200. Swap in -A "GPTBot" on the first command and you get the JSON too. This is the honest way to see what a given agent sees, because every agent stack sets its own headers.

What breaks for links that agents read

Cookies and referrers are browser features. An agent calling an API or fetching a page usually has neither, which is why page tags and referrer-based channel grouping miss it. The same gap shows up in dark social, and cookieless attribution covers the server-side alternative. GA4's AI assistant channel helps for named assistants that send a recognisable referrer.

For links you hand to agents, three habits hold up:

• Keep the destination stable. Repoint the short link when a page moves instead of publishing a new URL. A branded link on your own domain stays valid across redesigns.
• Make the destination answer a plain GET. Don't put a cookie wall, a JavaScript-only redirect or a pop-up consent step between the link and the content you want read.
• Keep your own logs. Server logs filtered by user-agent are the only complete record of what fetched a URL. Link analytics tell you about human-style clicks.

Closing the loop on revenue

If an agent does complete a purchase through a link you issued, Elido can credit it only when the order arrives through a Stripe or Shopify webhook that carries the elido_link_id. There is no public conversions API and no click ID passed to the destination, so an agent flow that pays elsewhere isn't attributed. Better to say so plainly than to imply a closed loop that doesn't exist. If agent-driven sales matter to you, route the payment through a Stripe or Shopify checkout that your own site controls, and tag the link there.

For agents that call tools directly, the repository includes an MCP server in the stdio form, covered in Elido MCP for Claude and Cursor. That post explains what it supports today.

What isn't built yet

Counting AI-agent visits as their own segment, and a way to return the redirect to a known agent when you want the click, are both on the product backlog. Until then, treat link analytics as a measure of human-style traffic, and read your server logs for the rest. If you're deciding how much to invest in agent-specific link handling today, start with the stable destination and the plain-GET check above.

Related on the Blog

• Answer engine optimization: get cited by AI
• GA4's AI Assistant channel: what it tracks and misses
• Elido MCP for Claude and Cursor
• Dark social attribution
• Cookieless attribution, explained