Agents that use your normal checkout now place 70% of AI-agent orders. Forter's data, charted.

For most of this year, the conversation about AI agents buying things has been a conversation about protocols. OpenAI has its checkout inside ChatGPT. Google has its Universal Commerce Protocol. Shopify, Walmart and BigCommerce have each announced ways for agents to connect. The assumption behind all of it was that agent shopping would arrive through these front doors, built and negotiated by platforms.

New order data from Forter suggests most of it arrived through the side door instead.

Forter, a fraud prevention company that sees checkout data from a large network of online merchants, published an analysis of agentic orders on October 23, covering July 23 to October 20, 2026. Forter splits [AI shopping agents](term:AI shopping agent) into two groups. Protocol agents (OpenAI's checkout in ChatGPT, Google's checkout in Gemini) buy through a commerce protocol. Screen-scraping agents (Meta's Muse, Instinct, Grok Bot and ChatGPT Agent) "use screen scraping, driving the merchant's own checkout pages." The second group is the story.

A note before the charts: this is one company's network, and Forter counts the dates it "started detecting agents reliably, which may be later than the dates they actually started placing orders." The data is about orders Forter saw, not the whole market. It is still the most detailed public look at agent checkout we've found so far.

The side door overtook the front door

Here is the headline number.

Show the data

Period	Share of agentic orders
Late July 2026	1%
Week of October 12, 2026	70%

Source: Forter, AI Agent Order Data: What's Actually Happening at Checkout, October 23, 2026. Screen-scraping agents in Forter's data: Muse, Instinct, Grok Bot and ChatGPT Agent.

In less than three months, agents that use a store's normal pages went from a rounding error to 70% of agentic orders. Forter's wording is that Muse "ushered in a shift in AI agent order data, with screen-scraping agents overtaking commerce protocols."

Even OpenAI, which runs both kinds, shows the same drift. ChatGPT Agent is a browsing agent that Forter reports separately from OpenAI's protocol checkout.

Show the data

Period	Share of OpenAI's agentic orders
Start of the period	1%
Three months later	15%

Source: Forter, AI Agent Order Data, October 23, 2026. Forter says ChatGPT Agent went from 1% to 15% of OpenAI's agentic orders within three months; exact start and end weeks are not given in the text.

That matters because protocol checkout is the part a brand can configure: you opt into a feed, set up a catalog, approve a channel. Screen scraping needs no permission. Forter's CEO, Michael Reitblat, put it plainly to eCommerceNews: "Merchants are effectively opted in by default, whether they've prepared for agentic traffic or not."

One launch changed the mix

Most of the shift traces to a single product. Forter first detected Muse orders on September 21, and the timeline in its post reads like a market being reset.

Period	What Forter reports
Throughout August	OpenAI placed 82% to 93% of agentic orders every week
August to September	Agentic orders grew 24% month over month, or 48% counting Instinct and Grok Bot, first detected September 1
Week of September 21 (Muse first detected)	Agentic orders up 229% month over month; Muse placed more than half
First eight days of Muse	Muse placed two-thirds as many agentic orders as Forter saw in the whole previous month
Week of October 12	Muse placed 43% of agentic orders; screen-scraping agents 70%
Whole period	All agentic orders roughly 5x their level in the first week of data

Source: Forter, October 23, 2026.

Two details stand out. First, Muse's growth was "additive, rather than coming at the expense of other agents." Since it arrived, OpenAI's daily orders are up 25% week over week. The new agent grew the pie rather than taking OpenAI's slice. Second, Muse's share is smaller by spend than by order count. Forter says the two are 16 percentage points apart on share of orders and 21 points apart on share of spend, "since Muse's orders skew smaller than OpenAI's."

The scale makes sense next to what Meta has said. As The Verge reported this week, Meta told investors on its Q3 earnings call that shopping was the most common use case for Muse so far, with purchases over $200 on average, and that 10 million people in the US downloaded the app in its first 40 days.

Agents reach more stores than the protocols do

The reach numbers are, to us, the most useful part for a brand that sells direct.

Show the data

Agent	Share of stores
Meta Muse	74%
ChatGPT Agent	45%

Source: Forter, AI Agent Order Data, October 23, 2026. Forter describes ChatGPT Agent's 45% as nearly double OpenAI's commerce protocol; it does not give the protocol's exact figure.

Muse bought at 74% of the stores that received any agentic order. ChatGPT Agent bought at 45%, which Forter calls "nearly double OpenAI's commerce protocol." That gap is the whole argument in one line. Protocol checkout only reaches stores that have joined it. A browsing agent reaches any store whose checkout it can work through.

So if you sell DTC and have not signed up for a single agent program, you shouldn't assume agents aren't buying from you. On this data, the more likely case is that some already are, through your normal pages, and your analytics files them as ordinary sessions or as bot traffic.

What agents spend

Forter removed the effect of agents buying on price alone to keep the comparisons fair, then reported these order-value facts.

Measure	Forter's figure
Average agentic order	$226
Google protocol orders vs. average	55% below (single-item checkout only)
Orders of $50 or less	27% of agentic orders
Most common range, all agents	$50 to $200
Largest share of orders over $500	Grok, which also has the highest average order, just ahead of OpenAI

Source: Forter, October 23, 2026.

For CPG, the low end is the interesting part. Forter describes agents as tending to be used "for larger, more considered purchases" and calls Google's small single-item orders atypical, suggesting "single-item or repeat purchases." Only 27% of agentic orders were $50 or less. A single bag of coffee or a tube of sunscreen sits below that line. A subscription box, a multipack or a bundle sits above it.

Our reading, and it is a reading rather than a finding: agents look more likely to buy a CPG brand's bigger baskets first. That points toward making bundles, variety packs and subscription options easy for an agent to understand on the page, with the count, size and price written out.

There is a reporting wrinkle worth flagging. The eCommerceNews write-up of the same data says ChatGPT has the largest share of orders above $500 and that Muse held 46% of agentic orders in "the final week of the month." Forter's own post says Grok has the largest share above $500 and gives 43% for the week of October 12. We've used Forter's post throughout, since it's the primary source.

Where agents shop, and where they don't

Forter says agentic orders are "heavily concentrated in North America." Google's and OpenAI's protocol pilots are US-only, and Muse is limited to the US and Canada. ChatGPT Agent is the exception, placing nearly a quarter of its orders outside the US.

On categories, Forter sees "very little overlap in agents' specialties," with OpenAI dominating "shipped retail categories: apparel, beauty, sporting goods, and consumer electronics." Beauty is the one squarely CPG category named. Forter doesn't break out food, household or personal care in the post, so we can't tell you how agent volume looks in grocery-like categories from this data.

What we'd take from this as a CPG brand

The data supports a few concrete moves, most of which cost nothing.

1. Assume agents read your product pages literally. A screen-scraping agent completes checkout using what's on the page. Pack count, size, flavor, price per unit, stock status and shipping rules should be in plain text, not only in images or a variant picker that needs three clicks. Our tip on letting agents buy covers the checkout side.
2. Check your bot protection on purpose. The RealReal tightened its defenses after The Verge showed Muse could hold items in carts for hours, while still letting agents surface listings. That's a legitimate choice for a resale site with one-of-a-kind stock. For a CPG brand with thousands of identical units, blocking every agent at checkout may turn away real customers. Decide which you want, rather than letting a default rule decide.
3. Separate agents from bots in your reporting. If Forter is right that agents already reach most stores that see agentic orders, some of your "bot" traffic is shoppers. Our audit of agent checkout settings before Black Friday has a list of places to look.
4. Keep store facts consistent everywhere an agent might look. Walmart's October terms for agents, its Agent Gateway announcement, and Shopify's new knowledge base in the admin (see today's news and tip 37) all push toward one set of structured store facts. A browsing agent will compare it with your pages. They should say the same thing, which is the core of tip 5.
5. Watch the answer before the order. Before any agent buys, an assistant decided your product was the one. That recommendation step is where most brands have no visibility at all. You can check it by hand on a monthly list of shopper questions, or with a tool; XENA Intelligence scores products on how they show up when assistants such as ChatGPT, Perplexity and Gemini are asked shopper-style questions, and shows the answers in SearchPanel.

What the data doesn't tell you

It doesn't tell you how many of your own orders are agentic; that depends on your traffic and your fraud tools. It doesn't give weekly totals in absolute numbers, only shares and growth rates. It covers about three months, during which one big launch reshaped everything, so the shares could move just as fast in the next three. And it is a fraud company's view, which naturally focuses on whether agent orders can be trusted.

What it does show, clearly, is direction. In late July, agent shopping on Forter's network was mostly OpenAI's protocol checkout. By mid-October it was mostly agents using the same pages as everyone else. Reitblat's estimate is that merchants who might normally have 6 to 12 months to react to a change in shopping behavior have "closer to 6-8 weeks" this time. With Black Friday a month out, that sounds about right to us.