Can AI Buy From You?

Can AI Buy From You?

The honest answer for products and for services, why it depends on where you are, and the audit that gets your store ready for ChatGPT, Gemini, and Copilot this week.

⏱ 14 min read ● Intermediate

Somewhere in the last year, a lot of business owners got told that AI assistants were about to become shopping malls, and that if they weren't listed, they'd be invisible. Some of that is true. Quite a lot of it isn't. And the parts that are true are not the parts most people are working on.

This article is the audit. By the end you'll know whether an AI assistant can sell your product at all, what's blocking you if it can't, and exactly what to fix. Most of the work takes a few hours. One part of it is probably costing you money right now without you knowing.

Section 01

What actually happened, and what most articles get wrong

In September 2025, OpenAI launched Instant Checkout in ChatGPT. You could see a product in a conversation, press a button, and buy it without leaving the chat. Etsy sellers went first, Shopify merchants followed, and the whole industry decided this was the future.

Then it didn't work. Conversion inside the chat was poor. Walmart, one of the biggest partners, pulled its ChatGPT checkout after about five months and said sales through it were disappointing, with buyers preferring to click through to Walmart.com. By March 2026, OpenAI had retreated.

Here's what OpenAI's own merchant page says today, and it's worth reading twice:

From chatgpt.com/merchants

"We're moving away from a standalone Instant Checkout experience in ChatGPT and prioritizing better shopping discovery and merchant-owned checkout experiences."

Translation: your products get recommended inside ChatGPT, and the purchase happens on your site.

So the headline everyone repeats, that ChatGPT is becoming a store, is out of date. What ChatGPT became is a very good product recommendation engine that sends people to your site. In-chat purchase still exists for approved partners and for big retailers who built their own apps inside ChatGPT, but it's no longer the main event.

Meanwhile Google went the other direction. Google's Universal Commerce Protocol launched in January 2026 and put a real buy button inside AI Mode and the Gemini app, where checkout happens on Google's surface and you stay the merchant of record. Microsoft did something similar with Copilot Checkout, and charges nothing for it. So the picture right now is: Google and Microsoft have the working buy button, and OpenAI has the traffic.

⚠️

A number to stop repeating: you'll see it claimed that Shopify charges 4% on ChatGPT checkout sales. Shopify's own pricing page shows its Agentic plan at $0 per month with the standard 2.9% plus 30 cents card rate, and no AI channel surcharge. OpenAI's merchant page says flatly that there are no fees on purchases that start in ChatGPT. The 4% figure traces back to a press report that appears to describe a different fee entirely. Don't budget around it.

The practical consequence for you is that the two halves of this have split apart. Discovery is global, free, and available to almost everybody. Checkout is narrow, regional, and closed to a lot of businesses. Nearly everyone gets this backwards and spends their effort on the half they can't have.

Most effort goes into the box on the right. Most of the available return sits in the box on the left.

Section 02

Three questions that decide everything

Before you touch a single setting, answer these. They take five minutes and they'll save you from building something you're not eligible for.

Question 1: What do you sell?

This is the one that stops most people, and almost nobody mentions it. Google's own documentation for merchants tells you to mark a product as not eligible for checkout if it falls into any of these categories:

CategoryNative checkout

Physical goods, in stock, fixed priceOpen. This is what the whole system was built for.

Services and consultingClosed. Explicitly excluded.

SubscriptionsClosed.

Courses and digital productsClosed on Google. Technically possible in OpenAI's spec, but no live example exists.

Personalised or made to orderClosed.

Pre-orders and rentalsClosed.

Alcohol, nicotine, firearms, gamblingProhibited outright by OpenAI's commerce policy.

Supplements with health claimsProhibited.

Anything age restrictedClosed.

If you sell services, courses, coaching, or software subscriptions, the buy button is not coming for you and there's no announced date when it will. Skip ahead to Section 06, which is written specifically for you. The rest of this article still applies, but your goal changes from being bought to being recommended.

Question 2: Where are you, and where are your buyers?

Two separate gates that get confused constantly. Your buyers' location decides whether a buy button appears. Your business's location decides which payment rails you're allowed to use. You can pass one and fail the other.

Question 3: What runs your checkout?

This decides your effort, not your eligibility. Shopify and Etsy sellers have almost nothing to do. WooCommerce, BigCommerce, Wix, and Squarespace sellers install something. Custom builds are a real engineering project.

🎯

Worth knowing before you read on: if you're on Shopify, you may already be selling in AI channels without having decided to. Shopify says its Agentic Storefronts feature is "enabled automatically for eligible stores." Go and look at your admin before you do anything else in this article.

Section 03

US, UK, and Europe are three different situations

Most coverage treats "the West" as one market. It isn't, and the differences are big enough to change what you should do this month. Here's how the three break down.

You're based inCan your local buyers buy in-chat?What you can do right now

United States Yes, across every assistant that offers it Everything in this article. You're the only market with no restrictions.

United Kingdom Not yet. Announced, no date. A surprising amount. The UK is one of only four countries where a merchant is eligible for Google's checkout system, and Stripe's agentic rails cover the UK. You can also sell to US buyers today.

European Union No, and not soon Your payment rails work. Stripe's agentic commerce suite covers 32 European markets. You can sell to US buyers now, and you can do the full discovery play at home.

Why the UK is in a better position than it looks

Google's documentation states that its checkout system is available to merchants who fulfil from within the United States, Canada, Australia, or the United Kingdom, and who hold a local bank account. That's the entire list. A UK business is in a club of four while most of the world is locked out entirely. UK buyers can't use the button yet, but UK merchants can be fully set up and ready for the day they can, and can sell to US buyers in the meantime.

Why Europe is stuck, and it isn't Google's fault

The blocker in the EU is payments law. Under PSD2, strong customer authentication is required per transaction, and none of the existing exemptions cleanly covers an AI agent buying from a merchant the customer has never dealt with before. That means an authentication step the agent can't complete. Until PSD3's delegation framework is finalised and actually implemented by payment providers, in-chat checkout for EU buyers stays blocked. My honest read is 2027 at the earliest, and it'll start with one or two countries rather than the whole bloc.

What European businesses do get, right now, is everything on the discovery side. ChatGPT's advertising product reached 31 European markets in August 2026. AI assistants recommend European products to European buyers every day. The traffic is real even where the button isn't.

💡

The workaround that works today, wherever you are: Shopify states that its Agentic Storefronts and free Agentic plan are "globally available to businesses selling to US buyers." For ChatGPT and Microsoft Copilot, any merchant in Shopify's Catalog qualifies regardless of store location. Google's AI Mode is the exception and needs a US-based shop. So a business in Manchester, Munich, or Manila can be bought inside ChatGPT today by pricing in USD and shipping to the US, with no US company and no US Stripe account.

Section 04

The audit: four checks to run this week

These apply to every business reading this. Physical products, services, US, EU, Shopify, custom build, all of it. They cost almost nothing and they're the entire game for anyone who failed Question 1 or Question 2. Do them in this order.

Check 1: Your robots.txt is probably blocking your own customers

This is the one costing people money right now. Back in 2024 and 2025, a lot of website owners blocked "AI bots" to keep their content out of training data. Reasonable instinct. But the AI companies run several different crawlers with different jobs, and blocking all of them removed those sites from AI shopping results too.

These are different decisions. Blocking a training crawler has no effect on whether you show up in ChatGPT's shopping answers. Blocking a search crawler removes you completely. Here's the split:

CrawlerJobVerdict

OAI-SearchBotPuts you in ChatGPT search and shoppingAllow

ChatGPT-UserFetches a page when a user asks about itAllow

GooglebotThe only crawler needed for Google's AI ModeAllow

Storebot-GoogleGoogle Shopping indexingAllow

PerplexityBotPuts you in Perplexity answersAllow

Claude-SearchBot, Claude-UserClaude search and live fetchesAllow

ApplebotSiri and Apple Intelligence answersAllow

Amzn-SearchBotAmazon and Alexa search resultsAllow

GPTBotOpenAI model training onlyYour call, no shopping impact

ClaudeBotAnthropic training onlyYour call

Google-ExtendedGemini training. Google states it does not affect Search ranking or inclusion.Your call

Applebot-ExtendedTraining permission flag, does not crawlYour call

Open yoursite.com/robots.txt in a browser right now and read it. If you see a blanket block on any of the crawlers in the top half of that table, that's your first fix.

Check 2: Your security settings may be blocking the agent trying to buy

Shopping agents look exactly like bad bots to older security tools. They run without a visible browser, from data centre IP addresses, with no mouse movement and no typing rhythm. Every one of those is a signal that traditional bot detection uses to block traffic.

DataDome, one of the major bot management vendors, now says openly in its own guidance that identity-based bot detection breaks down once real customers start delegating purchases to agents. Cloudflare rebuilt its classification in 2026 to separate bots into search, agent, and training categories, and added a behaviour tag specifically for agents doing checkout on a user's behalf.

If you're on Cloudflare, go into your bot settings and look for that three-way classification rather than trusting whatever you set two years ago. This failure is invisible: a blocked agent doesn't generate an error you'd ever see. You just quietly don't get the sale.

A blocked agent returns no error you would ever notice. The traffic simply never arrives.

Check 3: Your product pages need to be readable by a machine

Adobe published analytics in 2026 on how machine-readable US retail sites actually are. Homepages averaged 75%. Support pages scored 79 to 82%. Product pages came last at 66%, which is the worst possible result, because the product page is the only one that matters for selling.

The usual culprit is JavaScript. If your product's price, availability, and description only appear after a script runs, agents often can't see them. Google's own documentation warns that JavaScript-generated markup reduces crawl frequency and reliability. Server-render your product data.

Then add structured data. Google accepts JSON-LD, RDFa, or Microdata, so you're not locked into one format, though JSON-LD is the sensible default. Required fields are the product name, an image, and an offer containing a price and a currency code. Strongly recommended on top of that: brand, GTIN or MPN, availability, item condition, shipping details, returns policy, and an aggregate rating.

A few things quietly disqualify you, so check these:

One currency per URL

If the same page serves three currencies, it fails. Each currency needs its own URL.

No zero prices

A price of exactly $0 disqualifies the listing. Common on "contact us for pricing" pages.

One product per page

Category pages and multi-product pages aren't eligible. Single product focus only.

You must be the seller

Pages that send buyers to a different retailer to complete the purchase don't qualify.

Use Google's Rich Results Test on three of your product pages. It's free, it takes a minute per page, and it tells you exactly which fields you're missing.

Test yours free in Google's Rich Results Test. Paste a product URL and read the merchant listing result.

Check 4: Get your product feed into both merchant centres

Google Merchant Center and Microsoft Merchant Center both accept product feeds from over a hundred countries, and here's the part that matters: there is no country gate on discovery. Only the buy button is restricted. A business anywhere in the world can submit a feed and start appearing in AI-generated shopping answers.

For most readers outside the US, this is the single highest-return item in this article. It's free, it works today, and it's the reason the discovery half is worth more of your attention than the checkout half.

Get the core fields right: id, title, description, link, image link, price, availability, brand, and GTIN or MPN. Then keep them accurate, which brings us to the thing that breaks most agentic sales.

⚡

Stale feeds fail sales silently. An agent that once saw "out of stock" keeps telling customers you're out of stock until it receives an explicit update. Stripe's documentation asks merchants to refresh inventory and pricing every 15 minutes, and to keep the rate of checkouts that fail on unavailable stock below 5%. A daily feed refresh isn't enough once real agents are buying.

Section 05

Your platform, your path

Once the audit is done, what you do next depends entirely on what runs your store. Effort figures below are my estimates based on the documented requirements, not vendor promises.

Four of these five routes need no protocol engineering at all. Start at the top and only move down if you have to.

If you're on Shopify

You have the easiest route by a distance, and possibly nothing to do at all. Shopify is a co-author of Google's checkout protocol and has native support for OpenAI's. Its Agentic Storefronts feature is enabled automatically for eligible stores, which means your first job is to go and look at what's already switched on.

Shopify also runs a free plan, priced at $0 per month, built specifically for getting products into AI channels. It puts your catalogue into Shopify's Catalog, lets you choose which AI channels to sell in, and charges the standard 2.9% plus 30 cents online card rate when you make a sale. If you sell on a marketplace and have resisted running a direct store, this is a genuinely new reason to reconsider, because it's the only route into ChatGPT and Copilot that doesn't require building a storefront.

Effort: one to two days, almost all of it product data cleanup. Live in: days.

If you're on WooCommerce, BigCommerce, Wix, or Squarespace

All four route through Stripe's Agentic Commerce Suite, which means the first thing to check is whether your business is based in one of Stripe's 34 supported countries. That list is the US, Canada, and 32 European markets including the UK, Ireland, Germany, France, Spain, Italy, the Netherlands, and the Nordics. Good news for this audience specifically. Bad news for anyone outside those regions, because there's no APAC, Latin America, Middle East, Africa, India, Japan, or Australia on it.

WooCommerce publishes two extensions for this, one from Stripe and one from Google, and shipped native support for agent connections in WordPress 6.9 and WooCommerce 10.3. For the other three, you enable it from the Stripe dashboard: go to the agentic commerce page, onboard as a seller, then request a connection to the specific agent you want to sell through and wait for them to accept.

Two things people skip and then wonder why checkouts fail. First, configure a tax code for every taxable product, because Stripe warns explicitly that incomplete tax configuration causes checkout failures. Second, set up the feed refresh schedule properly: product data daily, inventory and pricing every 15 minutes.

Effort: three to ten days, mostly building the feed pipeline. Live in: two to four weeks including approval.

A daily feed refresh is enough to get listed. It is not enough once real agents start buying.

If you sell on Etsy, Amazon, or another marketplace

You don't control checkout, the marketplace does. That's mostly a limitation, but it also means the technical work is somebody else's problem. Your lever is listing quality plus one strategic decision.

Etsy syndicates its sellers centrally into ChatGPT, Gemini, and Copilot, so you're already there. Your entire input to that system is your listing data: title, description, attributes, and images. Etsy also gives sellers the option to opt out.

Amazon is different and worth understanding clearly. You're in Amazon's own shopping assistant automatically, as a normal search result. But being on Amazon does not put you in ChatGPT. Amazon blocks OpenAI's crawlers, sued Perplexity over its shopping agent, and publishes nothing that would let outside assistants sell Amazon inventory. Amazon is building its own agent and defending its marketplace, which is a coherent strategy but leaves its sellers invisible everywhere else.

If you have a custom or headless store

This is a real engineering project. You'd host five checkout endpoints, verify a signature header on every request, handle mandatory idempotency keys, emit order webhooks, allowlist the assistant's IP ranges, run TLS 1.2 or higher, possibly obtain a PCI attestation, and build a feed pipeline that hits the 15-minute refresh target.

My honest estimate is four to twelve developer-weeks for a first version, plus ongoing feed operations. Only worth it once you have measurable AI-referred traffic that converts. For almost everyone else, the right answer is to let Shopify, Stripe, Salesforce, or Adobe carry it.

Section 06

If you sell services, courses, or anything digital

Here's the straight answer, and it's not the one you were hoping for. Native in-assistant checkout is closed to you, and nobody has announced a date when it opens.

Google's merchant documentation requires that subscriptions, personalised goods, pre-orders, age-restricted items, services, rentals, and virtual or digital items all be flagged as ineligible for checkout. OpenAI's feed specification does contain a field for digital goods, and its 2026 spec update added consent fields that would support subscriptions, but I could not find a single named merchant actually selling a service or digital product through it.

That covers consultants, agencies, coaches, course creators, software businesses, and anyone selling made-to-order work. It's a big share of the people reading this, and it deserves a real answer rather than a vague "watch this space."

What it doesn't mean is that AI assistants are irrelevant to you. It means your lever is a different one, and honestly a better one, because the discovery half of this market is global, free, and far less contested than paid search. Four routes work.

Route 1: Be recommended, and convert on your own site

Assistants recommend services constantly. Agencies, consultants, tools, courses, all of it. Getting into those answers is a content and structured data problem rather than a commerce protocol problem, and it has no country restriction whatsoever.

Run the full audit in Section 04. The crawler check matters most, because you cannot be recommended by an assistant whose crawler you blocked. Then publish the things an assistant needs in order to say something specific about you: your pricing, what's included, who it's for, who it isn't for, honest comparisons, and real outcomes. Vague positioning that survives a human skim will fail an agent that has to state a fact.

Expect the traffic to land deep rather than on your homepage. Shopify reports that roughly half of AI-referred sessions land directly on a product page, about 2.5 times more often than traditional search traffic does. Your inner pages are your front door now.

Route 2: Build an app inside ChatGPT that runs your own checkout

This is the real answer to "can services be sold inside an assistant," and it's the route OpenAI pushed everyone toward after retreating from Instant Checkout. A merchant-built app runs your flow inside ChatGPT: your booking system, your account linking, your payment. Walmart, Target, Instacart, and Klarna all took this path.

Because you control the whole experience, it sidesteps the product category restrictions entirely. You're not asking Google or OpenAI to process a subscription. You're running your own. The catch is that it's built for businesses with volume, and there's no self-serve version for a two-person consultancy yet.

Route 3: Stand up an MCP server

MCP is the open standard that lets an assistant query your systems live rather than guessing from a cached page. It can ask what you have available, what a package includes, or what your next opening is. Shopify, Adobe, and commercetools all run one. WordPress and WooCommerce shipped support for it natively.

For a service business this is the closest thing to being bookable by an agent that exists without building a full app. It won't take payment, but it can carry the conversation right up to the point of one. Reckon on one to three weeks of development.

Route 4: If you're in travel, hospitality, or food, use the aggregator

Google announced extensions of its checkout system for lodging and for food in May 2026, both still on a waiting list. The named participants are aggregators, not individual businesses: Booking.com, Expedia, Hilton, Marriott, and Trip.com for hotels, and DoorDash, Uber Eats, Square, and Toast for food.

If you run a restaurant or a hotel, your route into agentic checkout runs through those platforms. Your job is to be listed completely and accurately on them, not to implement a protocol yourself.

🔑

One more restriction to check: OpenAI's commerce policy names several service categories as prohibited outright, including certain legal and travel services, unregulated financial services, identity theft monitoring, account modification services, and social media engagement services. If you're in one of those, even the app route is likely closed.

Section 07

How to tell whether any of this worked

Almost nobody sets up measurement, then six months later wonders whether the effort did anything. Four places to look, ordered by how much they actually tell you.

Stripe Dashboard

The strongest signal available. Transactions are tagged with the agent that originated them and can be filtered by agent name. The feed history page shows how many products are actually syndicated to each connected agent, which is how you catch a feed failing quietly.

Shopify admin

Agentic Storefronts has its own reporting: sales by AI channel, plus prompts telling you which product listings are too thin to be surfaced at all.

Your analytics

Watch referral traffic from chatgpt.com, perplexity.ai, gemini.google.com, and copilot.microsoft.com. Add UTM tags to the URLs in your product feed at setup time so you can separate feed-driven visits from organic mentions.

Google Merchant Center

AI performance insights, available in the US, UK-adjacent markets, Canada, Australia, India, and New Zealand, plus a diagnostic assistant currently in US beta.

Set the UTM tagging up before you launch, not after. It's the one measurement decision that's painful to retrofit.

Numbers worth knowing, with the caveat attached

Shopify reported that in the first quarter of 2026, AI-referred orders grew 13 times year on year and traffic 8 times. New-buyer orders arriving from AI search converted at roughly twice the rate of organic search. By the second quarter, AI-driven traffic and orders had tripled year on year, and 75% of AI-attributed purchases fell outside the top hundred product categories, meaning the long tail does unusually well here.

The most useful stat for our purposes: Shopify found that catalogue-powered AI search converts about twice as well as AI search without structured product data. That's the clearest argument available for doing the boring work in Section 04. Adobe separately measured AI-sourced traffic converting 42% better than non-AI traffic.

⚠️

Read those numbers carefully. Every published growth figure blends ChatGPT, Copilot, Gemini, and Perplexity together, and blends referral traffic with actual in-assistant purchases. Nobody has published conversion or revenue isolated to a single assistant's checkout. Fast growth from a small base is the honest reading. If a vendor quotes one of these numbers as proof that agentic checkout works, they're quoting you a discovery statistic.

Section 08

What to ignore, at least for now

A fair amount of noise surrounds this topic. Here's what you can safely put down.

Crypto payment standards. You'll see x402 mentioned, an open standard for machine-to-machine payments that moved to the Linux Foundation in 2026 with backing from Visa, Mastercard, Google, Stripe, and Shopify. The membership list is impressive. Real usage is API billing and agent-to-agent payments, not shopping, and CoinDesk reported in March 2026 that on-chain analysts estimated roughly half of its transaction volume was artificial. No mainstream retail checkout runs on it. Ignore it.

The payment authorisation protocols. AP2 is a real standard, donated to the FIDO Alliance in April 2026, with 60-plus partner companies. But you don't implement it. Your payment provider does, if and when it becomes necessary. The same goes for Visa's agent identity protocol.

llms.txt. This is a proposed file that tells AI systems how to read your site, and plenty of tools now generate one automatically. Here's the honest position: OpenAI, Anthropic, and Google all publish one for their own documentation, but none of their published crawler documentation says they read a third party's llms.txt, and no vendor claims it affects shopping visibility. It's cheap and harmless. It is not a lever, and anyone selling it to you as one is guessing.

Waiting for the protocol war to settle. Two standards emerged, OpenAI's and Google's, and both have been frozen at their April 2026 versions ever since. That's not a race, it's a stalemate, and the platforms are absorbing the complexity for you anyway. Don't hold off on the audit while you wait for a winner.

Section 09

Your week

Here's the whole article compressed into things you can actually tick off. The first four apply to everyone reading this, whatever you sell and wherever you are.

Everyone, this week

Open your robots.txt and confirm the search crawlers in Section 04 are allowed

Check your bot protection settings for the search, agent, and training split

Run three product or service pages through Google's Rich Results Test

Submit or refresh your feed in Google Merchant Center and Microsoft Merchant Center

If you sell physical products

Shopify users: open Agentic Storefronts and see what's already switched on

Everyone else: check whether your country is in Stripe's 34, then enable from the Stripe dashboard

Set inventory and pricing feeds to refresh every 15 minutes, not daily

Configure a tax code for every taxable product

Add UTM tags to your feed URLs before you go live

If you sell services, courses, or digital products

Accept that the buy button isn't coming, and stop waiting for it

Publish the specifics an assistant needs to recommend you: pricing, inclusions, fit, comparisons

Treat your inner pages as your front door, because that's where AI traffic lands

If you have volume, scope a ChatGPT app that runs your own checkout

If you're in travel, hospitality, or food, get your aggregator listings complete

One last thought, and it's the thing I'd want you to take away if you forget everything else.

The instinct most business owners have about AI and their website is defensive. Block the crawlers, protect the content, wait and see. That instinct made sense in 2024 when the only thing AI companies wanted from your site was training data. It doesn't make sense now, because the same companies are also running the systems that recommend products to buyers, and those systems only know about businesses they're allowed to read.

The checkout half of this story is genuinely narrow, genuinely US-first, and genuinely closed to a lot of businesses. But the recommendation half is open to everyone, everywhere, for free, and it's growing at rates that would have been the whole story in any other year. The businesses that win the next two years won't be the ones who built the fanciest checkout integration. They'll be the ones whose product pages an AI could actually read.

Go and check your robots.txt.