You cannot measure AI agent sales with one report, because agents reach your store in three different ways. Protocol orders arrive already tagged by the agent platform. Referral clicks from ChatGPT, Gemini or Copilot appear in web analytics when the referrer survives. Browser agents that click through your site themselves often appear only in server logs. This guide sets up all three, based on vendor documentation as of November 2026.

AI traffic growth to US retail, holiday 2025 (Adobe)
693.4%
Paths agent sales take
3
OpenAI bots with published IP lists
4
Adobe's description of the user base
Modest

Why is agent traffic hard to measure?

Web analytics was built around people running JavaScript in a browser. Agents break that in three ways:

  1. Protocol checkout skips your site. Under OpenAI's Agentic Commerce Protocol or Google's Universal Commerce Protocol, the agent calls your checkout API. No page view happens.
  2. Referrals get stripped. A shopper who reads a ChatGPT answer and then taps a link may arrive with a referrer, a UTM tag or nothing at all, depending on the app and browser.
  3. Browser agents look like bots. Claude Duet, Meta Muse and ChatGPT agent drive a real browser, but your analytics tool may filter them as bots or never load.

Put together, the usual "AI referrals" number in analytics is a floor, not a total.

Path 1: How do you track protocol orders?

These are the easiest because the platform tells you.

Route What identifies the order Where to look
ACP (ChatGPT merchant integrations) checkout_session_id on each session; order_created and order_updated webhooks you send Your checkout API logs and order records
Shopify Agentic Storefronts Shopify says orders show ChatGPT referral attribution Shopify Admin orders and reports
Google UCP checkout Orders created through your UCP integration Your UCP integration and Merchant Center
PayPal ACP server Orders placed via PayPal's integration PayPal and your order system

For ACP, OpenAI's checkout spec defines the session endpoints and the webhooks you send back, so tag every order created through /checkout_sessions/{id}/complete with a channel value such as agent_acp_chatgpt. Do the same for UCP orders. Google has not published a detailed reporting guide for UCP orders that we could find, so add your own tag at the integration layer.

Path 2: How do you track referral clicks?

When a shopper clicks from an AI answer to your site, analytics can catch it.

  1. List AI referrerschatgpt.com, gemini.google.com, copilot.microsoft.com, perplexity.ai and similar.
  2. Create a custom channelGroup those sources as AI assistants, separate from organic search and referral.
  3. Check UTM parametersSome AI apps add utm_source values to outbound links; capture whatever arrives.
  4. Add landing-page reportsSee which products AI answers send people to.
  5. Compare conversionAI referrals versus organic search, by product category.

We have not found an OpenAI document that specifies exactly how ChatGPT tags outbound links, so rely on what you see in your own data rather than a fixed rule. Shopify's statement that ChatGPT orders carry referral attribution suggests platforms are starting to pass this information more reliably.

Watch the gap between assisted and direct sales

Many shoppers research with an agent and then buy later by typing your URL. Those sales show up as direct traffic. A rough check: compare direct-traffic orders for products that are heavily recommended in AI answers against similar products that are not.

Path 3: How do you see browser agents in logs?

This is where most agent activity hides. Use your server, CDN or WAF logs, not JavaScript analytics.

Match User-Agent strings

Common shopping-relevant strings include ChatGPT-User, OAI-SearchBot, Claude-User, Perplexity-User and GoogleAgent-Shopping. OpenAI's bot documentation explains that ChatGPT-User visits pages when a user asks ChatGPT something and is not used for automatic crawling, which makes it a reasonable proxy for "a shopper asked about this page".

Confirm with published IP ranges

User-Agent strings can be faked. OpenAI publishes JSON IP lists for OAI-SearchBot, GPTBot, ChatGPT-User and OAI-AdsBot. Anthropic says Claude Duet's browser sends a distinct User-Agent header and that it publishes the IPs Duet browses from, in its Duet safety article. Count a hit only when string and IP agree.

Check signatures where available

Agents that support Web Bot Auth send Signature, Signature-Input and Signature-Agent headers. Log the Signature-Agent value and whether verification passed. Cloudflare's signed agents category, which included ChatGPT agent from August 2025, does this for you if you are a Cloudflare customer. See Web Bot Auth for stores.

Agents that do not identify themselves

Some agents send an ordinary browser User-Agent. Amazon cited exactly this when it blocked Meta Muse on 20 September 2026. You will not be able to count these reliably; note them as a known gap rather than guessing.

How do you tie agent visits to orders?

Once you can flag agent sessions in logs, join them to orders:

  • Tag the session or cart with an agent flag at the edge when a verified agent request arrives.
  • Carry that flag into the order record as a custom attribute or note.
  • Report orders by flag alongside protocol orders and AI referrals.

Pros

  • Protocol orders: precise and platform-tagged
  • Referrals: easy to set up in existing analytics
  • Log analysis: catches agents analytics misses

Cons

  • Protocol orders: only cover integrated channels
  • Referrals: stripped referrers make this a floor
  • Logs: unidentified agents remain invisible

What numbers should you expect?

Small ones, for now. Adobe's 2025 holiday report found AI traffic to US retail sites up 693.4% year on year, while noting "the base of users remains modest". Visa's October 2025 Trusted Agent Protocol release, citing Adobe, put the year-on-year rise at "over 4,700%". Growth rates that large from a small base mean your share can change within a season, so set up measurement before Black Friday on 27 November 2026 rather than after.

What should you do with the data?

  • If AI referrals convert well, invest in your product feed and checkout routes.
  • If agents visit but rarely buy, check whether your checkout is blocking them with CAPTCHAs or forced sign-ups.
  • If unwanted bots dominate, tighten rules using robots.txt and signature-based bot management.

Bottom line

Build one "AI agents" channel out of three feeds: platform-tagged protocol orders, AI referrals in analytics and verified agent visits from logs. Accept that some agent activity will stay invisible, and judge the channel on trends over a quarter. For the full readiness checklist, see making your store agent-ready, and if you are on Shopify, start with Shopify and agentic commerce.