On November 10, 2026, Semafor published the kind of story that does not survive a pitch meeting at most AI conferences: actual order counts. Airbnb CEO Brian Chesky said the platform is seeing “very, very low numbers” of bookings made by Meta’s Muse and Instinct AI agents. DoorDash CEO Tony Xu, at the same event, put his company’s lifetime total of orders from personal AI assistants at “maybe several hundred orders in total.”

Several hundred. Not per day. In total. At a company that processes hundreds of millions of orders per quarter, that is not a channel. It is a rounding error with a press strategy.

The industry has spent 2026 broadcasting the opposite impression. McKinsey’s widely cited research finds that more than half of global consumers now use AI to shop, a 70 percent jump over last year. Juniper Research projects agentic commerce spend hitting $1.5 trillion by 2030. Adobe’s actuals, which we have covered in detail, show AI-referred retail traffic up 693 percent during the 2025 holidays, up 393 percent in Q1 2026, and converting 40 percent better than non-AI channels on Prime Day.

All of those numbers are real. So are Chesky’s and Xu’s. The reconciliation between them is the most important story in commerce right now, and almost nobody is telling it straight.

Discovery is booming. Checkout is a ghost town.

The Adobe numbers measure something specific: a human clicks a link an AI surface showed them, lands on a retailer’s site, and buys the way they always have. That is AI-assisted discovery, and it is scaling exactly as advertised.

Autonomous checkout is a different product with a different adoption curve. Walk the public record assembled by AI Agents Info’s rollout tracker and the state of play in mid-November 2026 is small and oddly shy:

  • Google: Universal Cart went live in the US on October 15 in Search and the Gemini app, on top of what Google calls 60+ billion continually refreshed listings, with Nike, Sephora, Target, Ulta, Walmart, Wayfair and many Shopify merchants named. On November 4, direct checkout powered by the Universal Commerce Protocol began rolling out in Canada with lululemon, Shoppers Drug Mart, Simons and select Shopify merchants.
  • Microsoft: Copilot Checkout is live on Copilot.com and the Copilot mobile app, restricted to English-language merchants selling to US buyers in US dollars. Microsoft’s own page says it cannot give a timeline for other countries.
  • OpenAI: ChatGPT shopping is live for US users, but purchases complete on the merchant’s own site. The company retired Instant Checkout on March 24, 2026, saying it “did not offer the level of flexibility that we aspire to provide.” That is a polite way of saying agents buying on their own did not work well enough to keep.
  • Amazon: Rufus, now surfaced in-store as Alexa for Shopping, added text-based shopping for some US customers on November 6. Meanwhile, as Marketplace reported on November 10, Amazon actively blocks outside AI shopping agents from its storefront.

Four platforms. One week in November. Every capability live somewhere, none live everywhere, and the two companies with the most transaction data, Amazon and OpenAI, are respectively blocking external agents or walking back autonomous checkout entirely. That is not what a technology-limited market looks like. The technology shipped. The usage did not follow.

The door problem: half open, half bolted

Retailers are split on whether to let agents in at all, and the split has a rational cause. The same tooling that lets a legitimate agent browse also lets a scraper wear an agent’s name. According to DataDome, a bot management firm, 16.4 million spoofed requests have posed as Meta’s external agent across the web in 2026 alone.

This puts merchants in an impossible position, and it is the real story behind the blocking decisions:

  • Block everything that claims to be an agent, and you turn away the AI-referred traffic that converts 40 percent better on Prime Day.
  • Let everything through, and you serve your pricing, catalog and inventory data to an unattributable swarm of impersonators.
  • Gate by assertion, checking what an agent says it is, and you have solved nothing, because a bot can claim any identity it likes.

The IBTimes piece on retailer readiness quotes Albert Castellana Lluís, CEO of GenLayer Labs, on the underlying principle: businesses need mechanisms to confirm who an agent is, who authorized it, and what it is allowed to do, “for every single agent visiting their website.” That is an identity and verification problem, not a commerce problem, and it sits one layer below every checkout protocol in production.

The trust gap runs in both directions

Most agentic commerce coverage treats trust as a one-way street: how does the retailer know the agent is legitimate? The deeper gap runs the other way. When an agent is finally let in and asked to buy, what does it actually know about what it is buying?

The honest answer today: almost nothing that is not self-reported. Merchant feeds, product catalogs and shopping graphs are all attested by the seller. As we documented last week in our analysis of agentic commerce optimization, not one major feed specification carries an independently verified signal for review authenticity or product quality. The agent inherits the retailer’s marketing department as its sole source of truth.

The practitioners interviewed by IBTimes describe the same wall from the merchant side. “For twenty years, ecommerce has been designed around one person scrolling one page,” said Shashank Agrawal, founder of StarApps. “Every decision assumed a human was looking: the hero image, the color swatches, the reviews placed just above the fold. That assumption is breaking.” Ruslan Okhrimovych, CEO of Effie.ai, notes that agents “are built to analyze structured information and metadata: listed attributes, prices, availability, shipping times, and return policies.” They do not fall for a hero image. They also cannot detect that a product’s five-star history was manufactured by purchased reviews last quarter, because that signal is not in the feed.

Cristian Orasanu, co-founder of Naratix, gets closest to the point: “Agents will only recommend what they can understand and trust, and that’s the same data that already powers search, conversion and recommendations today.” Exactly. And that data is unverified. An agent that cannot distinguish authentic reviews from fabricated ones is not a smarter shopper. It is a faster victim.

The low-volume era is the build era

If the numbers are this small, why should anyone care this quarter? Because Deloitte’s survey of 200 retail and consumer goods executives shows the market has not decided what it is building: 40 percent of consumer goods companies and 29 percent of retailers have no defined approach to agentic commerce, even while roughly half are already piloting capabilities. The window between “piloting” and “volume” is when infrastructure gets chosen, and it is open right now.

Look at how much plumbing has already been poured. By late 2026 there are at least six protocols in active use: ACP and ACS for checkout, AP2 for payment authorization, UCP for commerce capabilities, x402 for machine payments, MCP for agent-tool connectivity, and A2A for agent-to-agent tasks. The IETF holds the first session of its agentic commerce protocol work on November 16 in San Francisco, five days from publication. When standards bodies convene, the deployment phase has begun, whatever DoorDash’s order counter says.

Here is the structural read. Every protocol in that stack moves money, identity or capability between merchant and agent. None of them moves evidence about products. The six-protocol stack assumes the agent already knows what is worth buying. It does not. It is reading the same polluted review corpus humans do: Amazon alone blocked 275 million suspected fake reviews in 2024, roughly 750,000 a day, and the fakes that get through are precisely the ones an agent will weight most confidently, because they look like consensus.

Forecast math makes this concrete. Juniper’s $1.5 trillion projection for 2030 requires agents to transact at volumes six orders of magnitude above DoorDash’s “several hundred.” That jump is gated on two verifications happening in milliseconds at decision time: is this agent real, and is this product actually good. The first is being built by identity protocols and courts of the kind GenLayer is drafting. The second is what we build.

The evidence layer GoBuy is building

GoBuy exists because the question “should I buy this?” has an answer that is computable: filter the fake reviews out, weight the authentic ones, score the product 0 to 100 on review quality rather than review quantity, and surface only the top 7 products in a category instead of five thousand listings optimized against your attention span.

That machinery was built for humans staring at Amazon. It is now agent-native. The GoBuy MCP server at gobuy.ai/api/mcp lets any AI shopping agent consult verified product trust before it recommends or buys: Smart Score, review authenticity signals, and the GoBuy Verified badge, earned only by products holding a score of 80 or above across 90 days. For shoppers who stay in the browser, the GoBuy Chrome extension injects the trust panel directly onto Amazon product pages.

The volume counters from Airbnb and DoorDash are not a rebuttal of agentic commerce. They are a schedule. The low-volume era is when the trust layer gets decided, and every agent that ships without an independent evidence source is piloting on marketing data with a payment rail attached.

If you are building a shopping agent, do not ship it blind. Point it at gobuy.ai/api/mcp, read the agent docs at gobuy.ai/agent-docs, and give it the one input the six-protocol stack forgot: proof the product deserves the order.

Smart Score. Real reviews. Seven products. Before the cart.