Short answer: on Forter’s network, agents that drive a merchant’s own checkout pages (screen-scraping) went from 1% of agentic orders in late July 2026 to 70% in the week of October 12. Agents that use a commerce protocol such as ACP or UCP can only reach stores that integrated it. Browsing agents can reach any store. If you have not built a protocol integration, agents may still be buying from you, through the page a human would use.

The figures below come from Forter’s own post, AI Agent Order Data: What’s Actually Happening at Checkout (published October 23, 2026). It covers Forter’s network from July 23 to October 20, 2026.

What Forter measured

  • Growth. August to September agentic orders grew 24% month over month, or 48% counting Instinct and Grok Bot, which Forter first detected on September 1.
  • Muse. Forter began detecting Meta’s Muse orders on September 21. That week agentic orders rose 229% month over month. Muse’s share has since waned, but it still placed 43% of agentic orders in the week of October 12.
  • Additive, not substitutive. Forter says Muse’s growth did not come at the expense of other agents: all agentic orders are roughly 5x their level in the first week of the data, and OpenAI’s daily orders are up 25% week over week since Muse arrived.
  • Reach. Muse bought at 74% of the stores that received any agentic order. ChatGPT Agent, a browsing agent Forter reports separately from OpenAI’s protocol checkout, bought at 45%: nearly double the reach of OpenAI’s commerce protocol.
  • Basket size. The average agentic order was $226. Only 27% were $50 or less, and $50 to $200 was the most common band.

How Forter classifies agents

Forter counts OpenAI’s ACP / ChatGPT Instant Checkout and Google’s UCP / Gemini as protocol checkouts, identified by where the shopper came from. Muse, Instinct, Grok Bot and ChatGPT Agent are screen-scraping agents, “driving the merchant’s own checkout pages,” identified by the agent platform. Within three months, ChatGPT Agent went from 1% to 15% of OpenAI’s agentic orders.

What this means for a merchant

Forter gives its own reason for the shift: browsing agents reach far more stores, because the protocols only reach stores that integrated them. For a merchant, the consequence is practical:

  • The checkout page is the interface. A browsing agent meets the same product page, cart and checkout a shopper does. We covered how Meta’s Muse reads a page: through an accessibility tree, with no JavaScript execution in the page context, per Meta’s own security write-up.
  • Protocol work and page work are not substitutes. Forter’s data says the protocol pilots are US-only and reach fewer stores; it does not say they matter less over time. Both paths can bring orders.
  • Your analytics may not label them. Forter says it identifies these agents by platform. A store without that visibility may see browsing-agent orders as ordinary traffic.

Limits of the data

This is one fraud-prevention vendor’s network, not all e-commerce. The post gives shares and growth rates, not the underlying order counts. Forter also notes that detection dates may be later than the dates agents actually started placing orders. Treat the percentages as a direction, not a market share.

FAQ

Are AI shopping agents buying through protocols or by browsing?

On Forter’s network, mostly by browsing. Screen-scraping agents placed 70% of agentic orders in the week of October 12, 2026, up from 1% in late July. Forter classes OpenAI’s ACP and Google’s UCP checkouts as the protocol side of that comparison.

Do I need ACP or UCP for AI agents to buy from my store?

Not for every agent. Forter reports that browsing agents such as Muse and ChatGPT Agent place orders on stores’ own checkout pages without a protocol integration. Protocols only reach stores that have integrated them, which is why Forter says browsing agents reach more stores.

Which AI agent placed the most orders in Forter’s data?

OpenAI placed 82% to 93% of agentic orders every week in August. After Meta’s Muse was first detected on September 21, it accounted for more than half of agentic orders in its first week, and still 43% in the week of October 12.

Sources

If agents can buy through the same page a human uses, the transaction has to work on that page: that is what AgentReady audits, and you can run it from the scanner.