When AI Agents Start Booking Things Without You

In September 2025, Etsy's stock jumped 16% in a single trading day — real money moving on a real exchange — because OpenAI announced that ChatGPT users would soon be able to buy things without ever leaving the chat window. Shopify popped more than 6% the same day. Six months later, according to Forrester research discussed on the Ecommerce Fastlane podcast, roughly 30 Shopify merchants had actually turned the feature on. Out of literal millions. By March 2026, OpenAI had quietly shelved the whole thing and gone back to the drawing board. That gap — between a 16% stock pop and 30 merchants bothering to flip the switch — is basically the entire current state of agentic commerce in one story: real, moving fast in some corners, and nowhere near as far along as the announcement headlines make it sound.

That doesn't mean the underlying shift isn't happening. It means it's happening unevenly, in fits and starts, with real infrastructure being built by real companies right alongside some very public stumbles. If you run a business with a website, the question worth asking isn't "is this real" — it clearly is, on some level — but "what does it actually require from my site today, versus what's still speculative." That's the honest version of this story, and it's worth walking through in order: what's shipped, what failed, what standards are actually competing for dominance, and what a business can reasonably do about any of it right now.

The Checkout That Wasn't

The mechanics of what OpenAI built were genuinely sound. Instant Checkout, launched September 29, 2025, let a user ask ChatGPT a shopping question, see a set of results the company says are "organic and unsponsored, ranked purely on relevance," and tap Buy without leaving the conversation. Behind it sat the Agentic Commerce Protocol (ACP), which OpenAI co-developed with Stripe and released as an open-source, Apache 2.0-licensed spec at agenticcommerce.dev — per Stripe's own announcement, merchants already on Stripe could turn it on with "as little as one line of code." Etsy sellers went first; Shopify's roughly one million merchants were coming soon. OpenAI disclosed, per SiliconANGLE's coverage of the launch, that over 700 million people were already using ChatGPT weekly — a genuinely enormous potential audience for a checkout button.

Then the actual usage numbers came in, and they told a much duller story. Modern Retail's reporting quotes Walmart EVP Daniel Danker describing conversion rates for products bought directly inside the chatbot as "three times lower" than for products where the customer clicked out to Walmart's own site. Etsy told the outlet it "did not end up seeing a large volume of sales" from the feature. OpenAI's own research, per reporting from The Information cited in the same piece, found that ChatGPT users were mostly researching purchases, not completing them inside the chat. And the infrastructure gaps were oddly basic — the Ecommerce Fastlane episode notes that as of February 2026, OpenAI still hadn't built the ability to collect US sales tax through the flow. By March 24, 2026, OpenAI pivoted, telling merchants plainly that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide," and moved toward letting retailers build their own in-chat apps instead of routing every purchase through one generic checkout box.

Anthropic entered the same space from a different angle almost exactly a year later. On September 2, 2026, per Digital Commerce 360's coverage, the company released a "blueprint" — pre-built patterns and guardrails, not a finished product — covering two roles: a Shopper Agent embedded in a retailer's own app to handle search, recommendations, and checkout, and a Merchant Agent for internal teams to watch inventory and flag pricing issues. Early partners include Accenture, Mastercard, Visa, and Shopify. Anthropic's own claim — that retailers running Claude-based shopping agents have seen carts "up to 35% larger" and shoppers "60% more likely to complete" a purchase — is worth reading with the same skepticism you'd apply to any vendor's own case study, since it isn't tied to a named retailer or an independent audit. The honest read on both companies' efforts so far: real infrastructure, genuinely early, and one very public example of the harder version of this problem — getting people to actually trust an agent with a real transaction — not yet solved.

What's Actually Shipping Right Now

Strip away the announcements and look at what people are actually running today, and the picture gets more specific. OpenAI's Operator, and its successor ChatGPT Agent mode, browses websites and clicks through them autonomously. Anthropic's Claude gained a comparable ability on October 22, 2024, called Computer Use — as Simon Willison documented in detail the same day, the model works from screenshots, calculates pixel coordinates, and issues mouse and keyboard commands in a loop, with Anthropic explicitly warning developers about prompt injection — malicious instructions hidden in a webpage's own content that can hijack an agent mid-task. That warning alone tells you a lot about how early this still is: the company shipping the capability is telling you, in the same announcement, not to point it at anything sensitive without a sandbox around it.

Perplexity's Comet browser is probably the most-tested consumer version of all this, and a review from Yahoo Tech is a genuinely useful data point precisely because of what the reviewer didn't do. Asked to book a London steakhouse reservation, Comet could clearly handle the task technically — but the reviewer stopped short of letting it actually complete the booking, writing plainly that they "didn't feel truly comfortable giving the AI my personal details" to finish it. That's not a bug report. That's a trust gap, from someone who was actively trying the product and still pulled back at the last step.

Real usage data backs up how early this all still is. HUMAN Security's State of Agentic Traffic report for June 2026 tracked which agents are actually out on the live web and what they're doing once they get there: Comet accounted for 47.6% of observed agentic traffic, Claude 20.8%, OpenAI's Atlas 16.5%, ChatGPT Agent 7.0%, and Genspark 2.6%. E-commerce made up the largest single category of that traffic at 43.8%, ahead of media at 41.3% and travel at 13.5%. But the number that matters most for anyone thinking about what to actually build sits in the page-type breakdown: 79% of agent activity on the web happened on product or search pages — browsing and comparing — while checkout and payment pages accounted for just 2.34%. Agents right now are overwhelmingly researching, not transacting. That's the single clearest, most current piece of evidence that "agentic browsing is real but not yet universal" isn't a hedge — it's what the traffic logs actually show.

The Standards Nobody's Fully Agreed On

Underneath all of this sits a genuinely unsettled technical layer, and it's worth naming the specific pieces rather than waving at "AI standards" in the abstract. The Agentic Commerce Protocol handles one narrow thing: the handshake between a buyer, an agent, a merchant, and a payment processor at the moment of checkout. Anthropic's Model Context Protocol, announced November 25, 2024, per Anthropic's own release, solves a different, broader problem — a single open way for an AI system to connect to "the systems where data lives," replacing what the company describes as a world where "every new data source requires its own custom implementation." Those are complementary, not competing, but they're also not the same thing, and an agency or developer who conflates them is going to build the wrong integration for the problem you actually have.

Then there's a genuine fight happening over how a webpage itself should talk to an agent that lands on it. Google's answer, WebMCP, is real and shipping in a limited form right now — published May 18, 2026 and available as a Chrome origin trial starting with Chrome 149, per Chrome's own developer documentation. Instead of an agent guessing what a page's buttons do by scraping the rendered DOM, a site can explicitly register a tool — "checkout," "filter_results" — with a JSON schema describing exactly what it expects and returns. That's a meaningfully different, more reliable mechanism than an agent trying to infer intent from HTML built for human eyes.

Compare that to llms.txt, a proposed convention that's been floating around for a couple of years asking sites to publish a plain-text summary for AI systems to read. Google's John Mueller addressed it directly in a Reddit thread in June 2026, and Search Engine Journal's report on the exchange quotes him calling it "purely speculative for now," adding pointedly that "the file has existed for years, yet none of the AI systems use it — what does it mean?" — before pivoting to say he preferred WebMCP instead. That's not a small disagreement. It's Google's own search team publicly dismissing one AI-agent convention while backing a competing one from the same company, in the same conversation. If the standards themselves are still this contested, "just add the file everyone's talking about" isn't a real strategy yet.

There's also a quieter, less discussed tension worth naming honestly: a meaningful share of the web is actively closing itself off to exactly the crawlers that agentic tools depend on. Cloudflare, which the company says handles traffic for roughly 20% of the web, announced on July 1, 2025 that it would start blocking AI crawlers by default for new domains unless a site owner opts in, alongside a "Pay Per Crawl" system letting publishers charge for access. More than a million Cloudflare customers had already chosen to block AI crawlers even before that default flipped, and more than 40 major publishers — Reddit, Pinterest, The Atlantic, and the USA TODAY Network among them — publicly backed the move. Being "agent-friendly" isn't a neutral default setting anymore. For a growing number of sites, it's a deliberate choice against a real, competing incentive to keep AI systems out.

What Machine-Readable Actually Requires

Put the standards fight aside for a second and look at what an agent concretely needs from a page to act on it reliably, because this is where most of the confusion for a business owner actually lives. There's a real difference between three separate things: content a human can read, content a chat assistant can summarize when someone asks it a question, and content an agent can act on directly — click through, fill in, and complete a transaction against. A page can be excellent at the first two and still be functionally useless for the third.

What the third tier actually needs is fairly specific. Schema.org markup on a Product or Service — with an explicit Offer carrying a real price, a currency, and a current availability status — gives a model something unambiguous instead of a paragraph it has to interpret. A stable, predictable page structure matters more than it would for a human visitor, because tools like Claude's Computer Use are working from screenshots and coordinates, not a semantic understanding of your layout; a page that reorganizes itself dynamically between one screenshot and the next is a page an agent can lose track of mid-task. And where a business can offer it, an actual API or a WebMCP-registered tool beats prose every time, because it removes the guessing entirely — the agent isn't inferring what the "Add to Cart" button does, it's calling a function that tells it directly.

The HUMAN Security data adds one more useful detail here: agents aren't only touching product pages. Authentication routes made up 5.38% of observed agentic activity and account pages 5.7% — meaning a fair number of these tools are already logging in and managing sessions on a user's behalf, not just window-shopping. That's a genuinely different technical surface than a chatbot summarizing your FAQ page, and it's worth a business at least being aware of before assuming "AEO" and "agent-readiness" are the same checklist.

What People Are Actually Arguing About

The clearest, most honest debate about all of this isn't in a vendor's blog post — it's in the comments underneath one. A Hacker News thread discussing the Instant Checkout and Agentic Commerce Protocol launch split about as sharply as you'd expect, and it's worth reading the actual disagreement rather than picking a side in advance.

On the optimistic end, a commenter posting as podnami pushed back on skeptics with a blunt question: "would they rather click, scroll, swipe...or just ask their digital assistant to do it?" Another, ahmedhawas123, called the whole thing "a pretty obvious monetization path" and floated the idea that ChatGPT becomes a genuine "entry point to the internet," competing directly with Amazon and Google Search rather than sitting alongside them. theptip made a narrower, more defensible version of the same case — that "deep research really shines" for comparing products, and that frictionless purchasing would be well received as long as accidental purchases get prevented.

The skeptics landed some sharper hits. matsemann worried openly about "enshittification" — the mechanism by which "sponsored products or spam bots" quietly poison the underlying recommendations until users can't trust them anymore. Analemma_ drew a direct comparison to Amazon's Alexa voice-shopping feature, noting flatly that "normal people...never, ever use them to buy things" — a real historical precedent for exactly the kind of feature OpenAI just tried and rolled back. tedd4u and alach11 both zeroed in on the incentive problem underneath the technology: "I don't want GPT shilling its wares," and, more pointedly, "will the model steer me towards ACP-supported retailers at a higher rate?" — a fair question for any system where the company running the agent also collects a fee from the merchants it recommends. siva7 put the whole critique in five words: "the day OpenAI turned into an ad company."

"will the model steer me towards ACP-supported retailers at a higher rate?"

— alach11, on Hacker News

The most grounded comment in the whole thread came from throw-qqqqq, who pointed out that OpenAI has reportedly been losing something like $8 billion a year despite its revenue — meaning a transaction fee on every agent-completed purchase isn't a nice-to-have feature, it's a monetization strategy the company genuinely needs to work. dbbk added the obvious follow-on: merchants absorbing that fee tend to pass it straight through to the customer, quietly raising prices on exactly the purchases that were supposed to be more convenient. None of that makes agentic commerce fake. It does mean the incentives are real, contested, and not fully settled — which is a very different claim than "this is inevitable" or "this is nothing."

Where This Actually Leaves a Business Today

Put all of it together and the honest picture is this: autonomous agents that browse, compare, and occasionally transact are a real and growing category, built by companies with serious engineering behind them, and simultaneously a category where the highest-profile checkout product to date got quietly wound down after six months because almost nobody used it. Both things are true at once, and a business trying to plan around this needs to hold both.

What that means practically is less dramatic than "make your site agent-ready" implies. Most agent traffic right now, per the HUMAN Security numbers, is doing exactly what a careful human researcher would do — comparing products, checking availability, reading details — just faster and at scale. The businesses positioned well for that aren't the ones that bolted on a checkout API nobody's using yet. They're the ones whose pricing, availability, and service details are already correct, current, and structured cleanly enough that a research-mode agent parses them without guessing — the same foundation that makes a business more discoverable to a chat assistant answering a question, extended one step further to a tool that can act on what it finds. We built our own Agentic SEO work around that specific, current reality rather than a speculative future one, and if you want a sense of how your own site's structured data and entity signals currently look, the AEO Score tool gives you a real, evidenced starting point in a few minutes.

None of that requires betting the business on autonomous checkout becoming the default way people buy things — the evidence so far genuinely doesn't support that yet. It requires treating "can an automated process reliably read what I actually offer" as one more real, measurable thing about your website, alongside how it ranks and how often it gets cited — not because every customer is an agent today, but because a meaningful and growing share of the ones doing the early research already are.