The Checkout Moved Into the Chat Window
AI shopping agents are already comparing prices, reading reviews, and in some cases buying, without a human ever landing on your homepage. What that costs, what it takes to be findable, and what changes next.
Somebody needs a new espresso machine. Two years ago that meant a Google search, six tabs open, and twenty minutes of comparing specs. Now it means one sentence typed into ChatGPT or Perplexity: "find me a good one under $300 with a milk frother, order it." The agent compares products across sites, reads the reviews, checks the price, and in a growing number of cases, buys it. Nobody visits your homepage. Nobody sees your banner ad. The first time you know a sale happened is when the order lands.
That is not a hypothetical anymore. It is a live, if still messy, part of how people shop, and it changes what "being found online" means for a business that sells things.
How much of this is actually real right now
Worth being honest about scale before anything else, because the hype and the reality are running at different speeds.
eMarketer projects AI platforms will account for roughly $20.6 billion in U.S. retail ecommerce sales in 2026, about 1.5 percent of the total, but nearly four times what it was in 2025. That is a real number and a small one at the same time. It only counts purchases completed inside the AI platform itself, not the much larger wave of people using AI to research and then buying the normal way.
That larger wave is where the traffic numbers get startling. Adobe Analytics measured AI-referred traffic to U.S. retail sites up 805 percent year over year on Black Friday 2025, and up roughly 693 percent across the full holiday season. Morgan Stanley's survey found 53 percent of ChatGPT users and 46 percent of Gemini users had used the platform to research or compare prices on a product in the past month. Most people are not letting an agent buy for them yet. A lot of people are letting one do the shopping-around part.
And here's the detail that should actually change how you think about it: when AI-referred visitors do land on a retail site, Adobe found they convert at rates roughly 31 percent higher than visitors from other sources. Someone arriving because an AI agent already vetted the options is closer to a decision than someone who just clicked an ad.
What an agent is actually grading when it looks at your store
An AI agent does not browse your site the way a person does. It does not scroll, admire your homepage banner, or forgive a confusing menu because your logo looks nice. It reads data. Specifically:
- Structured product data. Price, availability, condition, GTIN or MPN, and review ratings, marked up in a machine-readable format (schema.org Product markup, or a feed like Google Merchant Center). Most ecommerce sites have partial schema at best, missing exactly the fields agents rely on: real-time inventory and aggregate review data are the two most commonly skipped.
- Accuracy that matches everywhere. If your website says $89 and your feed says $94, an agent either skips you or, worse, sends a customer into a checkout that doesn't match what they were quoted. That mismatch is a fast way to lose a sale that was already halfway closed.
- Crawler access. GPTBot, PerplexityBot, and similar agents have to be able to reach your pages at all. A security plugin or an overly aggressive robots.txt file that blocks "bots" indiscriminately blocks these too, and there's no error message telling you it happened. You just quietly stop showing up.
- A checkout protocol, or a path around needing one. The two big ones right now are OpenAI and Stripe's Agentic Commerce Protocol (ACP) and Google's Universal Commerce Protocol (UCP), launched in January 2026 and backed by Amazon, Shopify, Target, Etsy, Walmart, and Wayfair among others. If you sell on Shopify, this may already apply to you without you doing anything: Shopify turned on "Agentic Storefronts" by default for eligible merchants in March 2026, syndicating store catalogs to ChatGPT, Microsoft Copilot, Google AI Mode, and Gemini automatically. Worth checking your own settings. You may already be listed somewhere you've never looked.
None of this is exotic. It's the same instinct as good SEO always was: give the machine clean, honest, structured information instead of making it guess. The difference is the machine is now the one placing the order, not just ranking the page.
What it's actually costing business owners
This is the part that gets skipped in most of the coverage, because "AI agents are shopping now" is a more exciting headline than "here's what it costs to participate."
Transaction fees on top of your normal ones. OpenAI charges merchants a 4 percent fee on completed Instant Checkout purchases, stacked on top of standard payment processing, which typically runs about 2.9 percent plus 30 cents. On a $100 order that's roughly $7.20 gone before you've shipped anything. That's a real cost of doing business in the channel, not a rounding error.
The conversion isn't automatically better. Walmart, at real scale, measured checkout completed inside ChatGPT converting roughly three times worse than sending that same customer to walmart.com to finish the purchase there. By early 2026, OpenAI had scaled back in-chat Instant Checkout after it launched with a lot of fanfare in September 2025. Only around 30 Shopify merchants were actually live on it by February. The technology moved faster than the buying behavior did. That's a useful thing to know before you spend real time chasing it.
Margin pressure from being reduced to a spec sheet. When an agent is comparing you to four competitors on price, availability, and rating, the things that used to differentiate you, brand story, in-store relationship, the way your site feels, stop mattering to the decision. Industry analysts tracking this shift describe it plainly: AI shifts loyalty from brands to outcomes. That is a real cost even when no invoice shows up for it. It shows up later, as thinner margins and a harder time justifying your price against a competitor an algorithm decided was "close enough."
The cost of doing nothing. Harder to put a number on, but real. If your competitor has clean product data and a working feed and you don't, the agent isn't choosing between you on merit. It's choosing the one it can actually read.
Why this isn't a five-alarm fire, either
Worth saying plainly, because the opposite mistake, panic-adopting every protocol that launches, is its own waste of money. Consumer trust has not caught up to the infrastructure. Even with 60 percent of shoppers expecting to use an AI agent within the next year, only somewhere between 23 and 30 percent currently trust an agent to actually spend money on their behalf without a human double-checking first. Most people still want to see the checkout page with their own eyes before they hit buy, even after letting the agent do the legwork.
That gap is exactly why the traffic-and-research numbers matter more right now than the checkout-completed numbers. The agents are already in the room comparing your prices. Most of them just haven't been handed the wallet yet.
What the next few years look like
The forecasts vary in size but agree on direction. McKinsey estimates agentic commerce could orchestrate up to $1 trillion in U.S. retail revenue by 2030, and $3 to $5 trillion globally. J.P. Morgan puts it at up to 25 percent of U.S. online sales by 2030. Gartner projects 20 percent of digital commerce transactions running through AI platforms by the same year. Nobody serious is arguing this shrinks back down.
What's shifting underneath those numbers is the plumbing. Visa and Mastercard are both building agent-specific payment rails, Visa Intelligent Commerce and Mastercard Agent Pay, that bind a payment credential to a specific agent, merchant, and spending limit, so a business owner or a card issuer can actually control what an agent is allowed to spend without a person clicking approve on every transaction. Stripe, Google, and a growing list of retailers are pushing toward more than one open protocol rather than a single winner. Merchants who support more than one protocol are already seeing meaningfully more agentic traffic than those betting on just one. The practical takeaway: this is not a single standard to wait out. It's turning into infrastructure you plug into, the same way you plug into Google and Meta today, whether or not any one brand name wins the headline.
Where to actually put your attention
Not everything above needs action this month. A short, honest list of what does:
- Check whether your product pages have real, complete schema markup, price, availability, condition, GTIN, reviews, not partial markup that looked fine in 2023.
- Make sure your price and inventory are accurate and consistent everywhere they're listed. Agents penalize mismatches harder than people do.
- If you're on Shopify or a platform with a similar feature, check whether you've already been opted into agentic storefronts or similar syndication, and decide on purpose whether you want to be.
- Don't block AI crawlers by accident. Check what your security tools and robots.txt are actually doing.
- Watch conversion, not just traffic, before you invest real time chasing any one checkout protocol. The Walmart example above is worth remembering.
This is the same discipline that's always separated businesses that get found online from the ones that don't: clean, accurate, honest information, structured so the thing reading it, human or otherwise, doesn't have to guess. The reader changed. The job didn't.
If you want a second set of eyes on where your store actually stands in all this, reach out. No pitch, just a straight look at what an AI agent sees when it looks at your business today.
Found this useful?
Most of what I write about comes from real conversations with small business owners. If something here connects to what you are working through, let's talk.