Agentic commerce, also called agentic shopping, lets an AI agent buy for a person. The agent finds products, compares price and stock across stores, and pays on the shopper's behalf, sometimes without that person ever opening the store's website or app. The merchant still receives the order and payment through a protocol supported by both sides.
Rıza Doruk Ercan described this on episode 24 of our podcast: a shopper opens ChatGPT, buys from Walmart and leaves, without touching Walmart's site or app. Walmart then realized that flow was costing it first-party data.
Most analytics tools expect a browser between the ad and the order. When an agent shops, that browser may belong to the agent, or there may be no browser at all.
Where AI shopping agents check out today
This table reflects checkout options on 14 October 2026. OpenAI launched in-chat checkout in September 2025 and retired it six months later.
| Platform | Protocol | Where checkout happens | What the merchant receives |
|---|---|---|---|
| ChatGPT | Agentic Commerce Protocol (OpenAI and Stripe) | On the merchant's site, or in the merchant's own app inside ChatGPT. Instant Checkout was retired in March 2026. | A referred visit. ACP product feeds decide which products ChatGPT shows. |
| Google AI Mode and Gemini | Universal Commerce Protocol, with the Agent Payments Protocol for payment | Inside Google's Universal Cart, paid with Google Pay. US first, with Canada, Australia and the UK to follow. | A UCP order. You stay merchant of record. |
| Shopify Agentic Storefronts | Shopify Catalog and UCP | Depends on the surface: inside the chat on Google and Copilot, on your store through an in-app browser on ChatGPT. | A Shopify order in your admin. |
| Perplexity | Instant Buy, paid through PayPal | Inside Perplexity. | An order on your PayPal account. You stay merchant of record. |
| Visa and Mastercard | Visa Intelligent Commerce, Mastercard Agent Pay | Wherever the agent checks out. | A payment made with a token issued to the agent, not the shopper's card number. |
Three ways an agent touches your store
Hardal AI Visibility groups agent activity into these three categories.
It reads your pages. Fetchers such as ChatGPT-User and Perplexity-User request your HTML when someone asks a question. They don't run JavaScript, so no tag fires and only your server logs record the hit. See how to measure ChatGPT-User.
ChatGPT agent is a different case. It loads your page in a real browser with a normal Chrome user agent, so your tag can fire and count it as a visitor. It signs each request with an HTTP message signature and a Signature-Agent: "https://chatgpt.com" header, which is how you tell it apart from a person.
It sends a person to you. The visit arrives with utm_source=chatgpt.com, a chatgpt.com referrer, both, or neither. Since ChatGPT moved checkout back to merchants, this is how most ChatGPT orders reach you. ChatGPT referral tracking covers which markers survive and which don't.
It buys without a visit. Checkout runs inside the agent: Google's Universal Cart, Perplexity's Instant Buy, Shopify checkout inside Copilot. Your backend gets an order with no browser session behind it.
What agentic checkout breaks in your analytics
Checkout inside Google or Perplexity leaves three gaps in browser-based analytics.
- The sale has no browser trail. No page view, no pixel, no cookie, no click ID. GA4 records a purchase when its tag fires on the confirmation page. With no confirmation page, the order shows up as direct, if it shows up at all.
- Ad platforms never see the conversion. Meta, Google Ads and TikTok tie a purchase to a click through their pixel or click ID. An agent order carries neither, so the campaign that created the demand gets no credit. ROAS reads low, and automated bidding moves budget away from campaigns that are working.
- Browser agents distort your session data. An agent that opens 20 product pages in 40 seconds counts as 20 page views from one very fast visitor. Unless you separate agents from people, sessions, bounce rate and conversion rate drift with every agent that visits.
How to see agent orders in first-party data
Every path in the table still ends with an order in your system, so measure from the order.
Collect orders from the backend. Send each purchase to Hardal from your order system over the server-side REST endpoint when the order is created. An order placed in Universal Cart reaches your backend the same way a web order does, so it gets counted. A tag on the confirmation page only counts orders that had a confirmation page.
Keep the source fields the protocol sends. Both protocols leave room for attribution, and both make it optional.
- UCP identifies the calling platform in a
UCP-Agentheader that points to the platform's profile. The checkout can carry anattributionobject of URL-style parameters: campaign IDs, click IDs, source and medium. That object is copied onto the order as a read-only snapshot (spec version 2026-08-25). - ACP lets the agent attach
affiliate_attributionwhen it creates or completes a checkout: provider, publisher ID, campaign ID and a token. It's stored with the order and isn't returned in the response. It was built to credit affiliate publishers, so expect publisher and campaign IDs rather than UTM-style parameters.
Store whatever arrives as raw fields on the order event. Don't fold them into a channel when you collect them. Both specs are versioned every few months, and raw fields let you re-map history after the next change.
Separate agent visits from human sessions. Hardal AI Visibility reports requests from identified crawlers and agents, and the pages they hit, apart from your human traffic. It's in beta. Join the waitlist for early access.
Send the order to ad platforms server-side. Once the order is a server event, Hardal routes it to Meta, Google, TikTok and OpenAI Ads without a browser pixel, the same way the OpenAI Ads destination already delivers ChatGPT Ads conversions. The Measurement Pixel post compares OpenAI's pixel. Without a click ID, matching depends on the customer details on the order, such as a hashed email, and on the consent you hold for them.
Agent-ready store checklist
- Product and Offer structured data on every product page, with price, currency, availability and GTIN.
- A clean product feed. Google Merchant Center for Google surfaces, an ACP feed if you want ChatGPT to show your products, or Shopify Catalog if you're on Shopify.
- robots.txt that lets agents in. Allow OAI-SearchBot and don't block ChatGPT-User or Perplexity-User. Blocking GPTBot, the training crawler, is a separate decision.
- Server-rendered product pages. Fetchers don't run JavaScript. If your price renders client-side, the agent reads a page with no price.
- Price and stock that match the feed. Google disapproves offers when the feed and the landing page disagree, and an agent quoting a stale price loses the sale at checkout.
Run your domain through isvisible.ai to see which AI crawlers can reach it today. The AI visibility playbook covers the rest.
What no tool can attribute yet
- A source the platform never sends. Both protocols make attribution optional. When the platform leaves it empty, no tool can reconstruct it, and any tool that claims to is modeling.
- The conversation before the order. You see the order. You don't see the prompt, the products the agent compared, or the competitor it rejected.
- Unsigned browser agents. An agent that doesn't sign its requests and uses a normal browser looks like a person. Behavioral detection can only estimate.
- Research in one place, purchase in another. A shopper who compares products in ChatGPT and buys on their phone two days later arrives as direct. That's the same gap as dark social, and it's still open.
Get started
AI Visibility is in beta. Join the waitlist for early access. To start collecting orders from your backend and routing them to your ad platforms now, book a demo.