Visa announced card payments inside ChatGPT on June 10, 2026. How an AI agent purchase works end-to-end and the three gaps that keep most sites out of agentic commerce.
Yesterday, June 10, 2026, Visa stood on stage at its Payments Forum in San Francisco and announced that its payment network now plugs into ChatGPT. Users link a Visa card, set limits, and the agent can buy on their behalf at any merchant that accepts Visa. Payment was the last missing piece of AI agent commerce. The browsing, comparing, and selecting were already happening. Now there is a mainstream way for the agent to finish the job, and the question moves to your side of the counter: when an agent arrives with a credential and intent to buy, can your site actually sell to it?
The verified facts first, because announcements in this space have a habit of growing in the retelling. Per the June 10 announcement: users link their Visa cards to ChatGPT, and AI agents can then shop and complete transactions across merchants that accept Visa, not at a single partnered retailer. The network handles payment authorization and fraud monitoring. The guardrails named are spending limits, approval steps, and approved merchant lists. The integration builds on Visa Intelligent Commerce, the token framework the company has been developing for agent-initiated payments.
What was not announced matters just as much. Neither company disclosed financial terms, the fees merchants or customers will pay, a regional rollout map, or whether this is a limited pilot or general availability. Treat the timeline as open. The capability is announced; the footprint is not yet public.
Even with those gaps, the structure of the deal is the story. Compare it to OpenAI's Instant Checkout, which was retired in March 2026 after charging merchants 4% per transaction and seeing limited adoption. Instant Checkout asked merchants to opt in and pay a premium. The Visa approach inverts that: it rides the card rails you already use. If you accept Visa today, you are inside the addressable market for ChatGPT purchases without signing anything. There is no merchant gate to walk through.
Which means the competition just changed shape. The gate is no longer "did you join the program." It is "can an agent find your product, understand it, and complete your checkout." Plenty of sites will fail that test while their competitors pass it, and the agent will not file a complaint. It will simply buy elsewhere.
An agent purchase has four stages, and the payment announcement only solved the last one.
1. Browse. The agent fetches your pages. Some agents use a live crawler with a distinct user agent (ChatGPT-User, OAI-SearchBot, PerplexityBot); some drive a real browser. Either way, if your product pages render their content client-side only, the crawler-style agents see a blank shell. Server-rendered product data is table stakes here.
2. Extract. The agent has to turn your page into structured facts: product name, price, variants, availability, shipping. Pages with schema.org/Product markup and clean semantic HTML get parsed reliably. Pages where the price lives in a JavaScript-rendered widget, or where the spec table is an image, get parsed wrong or not at all. An agent that cannot confirm your price will not guess. It moves on.
3. Select. The agent compares candidates across merchants and picks one. This is where extraction quality becomes revenue. When three stores carry the same item, the one whose data parsed cleanly, with confirmed stock and a visible total, wins the slot. The other two were never really in the running, and nothing in their analytics will say why.
4. Pay. With the Visa integration, the agent presents a tokenized credential inside the user's limits, possibly pausing for an approval step. Some agents complete checkout by driving your DOM like a human would, filling fields and clicking buttons. Worth being precise here: Chrome's auto-browse, the highest-distribution agent of the moment, operates pages through the DOM and does not invoke WebMCP. The DOM-level readiness problem is its own topic, and we audited it in depth in our agent transactability guide. This post is about the upstream stages, because a flawless checkout never gets reached if the agent could not browse, extract, and select you first.
The Visa news is agents spending human money on human-sized purchases. There is a second, stranger thread worth watching: agents spending machine money on machine-sized purchases.
The x402 protocol, an open spec Coinbase introduced in 2025, revives HTTP status code 402 ("Payment Required") as a working mechanism. A server replies to a request with a 402 and a price; the client attaches a stablecoin payment, typically USDC, and retries. No account creation, no API key, no invoice. Payment becomes a property of the request itself.
It is already running in the wild, at small scale. Superhighway, a web-search API that surfaced as a Show HN demo, sells search, scrape, and research calls to agents and advertises that "agents can even pay for each call on their own — no API key required." One indie demo does not make a market, and to be clear, this is proof-of-concept territory, not something your finance team needs a policy for. But it demonstrates the end state: an agent that discovers a paid capability mid-task and buys access in milliseconds, with no human wiring anything. If your business sells data, APIs, or digital goods, per-call agent pricing is a distribution channel that did not exist two years ago.
The payment rails are one front. The advertising surfaces are another, and they are moving in the same direction. Google has begun bringing agent-mediated buying into its ad products, which means an AI agent can reach your offer two ways now: by crawling your site, or by acting on a paid placement it was served. Google's ads and commerce updates point at a near future where the entity clicking your ad and the entity completing the purchase can both be software.
This sounds like a reason to rethink your ad strategy. It is mostly a reason to fix the page the ad points at. An agent that taps a sponsored result still has to do the same three things any agent does before it buys: browse the landing page, extract the product facts, and decide your offer matches the task. If your landing page renders price client-side, hides variants behind JavaScript, or buries availability, the agent bounces, and you paid for the click anyway. Ad spend does not skip the readiness stack. It sits on top of it.
The practical takeaway: agent-readiness is upstream of agent advertising. Before you budget for agent-targeted campaigns, confirm that an agent landing on your page can actually parse and act on what it finds. A clean, server-rendered, schema-marked page is what converts an agent click into an agent purchase. Spending on placements that route agents to a page they cannot read is lighting money on fire one tokenized transaction at a time.
Run the four-stage purchase model against a typical e-commerce site and the same three gaps appear almost every time.
Gap 1: no bot visibility. Most operators cannot answer the basic question: are AI agents hitting my product pages right now? Google Analytics filters most bot traffic by design, so the answer is invisible in the tool everyone checks. Agent visits live in your server logs under user agents like GPTBot, ChatGPT-User, ClaudeBot, and PerplexityBot. Without that visibility you are making agentic-commerce decisions blind, including the decision that it is safe to ignore.
Gap 2: no llms.txt guidance. Agents work under context limits. A site with thousands of URLs and no map forces the agent to guess which pages matter, and it guesses wrong often. An llms.txt file hands the agent a curated index: here is the catalog, here are the bestsellers, here is shipping and returns. It is the cheapest fix on this list, a single markdown file at your root.
Gap 3: no machine-readable commerce layer. Prices rendered client-side, missing Product schema, variant pickers that only work via JavaScript events, totals that appear only after a multi-step form. Each of these is a place where an agent loses confidence in your data. The fixes range from adding structured data (hours of work) to exposing products as callable tools via WebMCP (a script tag, for the agents that support it).
Three jobs, in priority order.
Detect. Get agent traffic visible before you change anything. Crawlytics reads AI bot and agent visits from your traffic and shows which crawlers touch which pages, so you know whether ChatGPT-User is already fetching your product pages or whether you are early. The free Agent-Ready Grader gives you a baseline scan in about a minute.
Serve. Ship llms.txt, add Product schema, and make sure name, price, and availability exist in the server-rendered HTML. This is the layer that determines whether you survive the extract and select stages, where most agent purchases are actually decided.
Sell. Add the WebMCP commerce snippet so agents that support tool calls can query products and start checkout directly instead of scraping. On Shopify this is a 15-minute theme edit. Then audit the human checkout path itself for DOM-level transactability, using the readiness audit, because the Visa-credentialed agent still has to get through your actual checkout flow.
The payment rail arrived this week. The sites that win the first wave of agent purchases will be the ones that were findable, parseable, and buyable before their category filled in.
Written by Crawlytics Team. Crawlytics tracks AI bots, generates llms.txt, and powers WebMCP commerce, all from one snippet on any stack. See how it works →
Yes, in specific configurations, and the set is growing. As of the June 10, 2026 announcement, users can link Visa cards to ChatGPT and authorize agents to complete purchases at Visa-accepting merchants, within spending limits and approval steps. Separately, browser-driving agents like Chrome's auto-browse complete checkouts by operating the page directly, which works only when the flow is accessible to a non-human operator. What is not yet public for the Visa integration: regions, fees, and whether it is a pilot or generally available. Capability is real; ubiquity is not here yet.
You don't need it to be reachable, but it is the cheapest upgrade to being reliably buyable. Shopify stores already produce decent structured data, so agents can usually extract your products. WebMCP goes further by letting agents call your store as tools (search products, check a variant, begin checkout) instead of scraping rendered pages, which is faster and doesn't break when your theme changes. The snippet is a script tag that no-ops in browsers without agent support, so there is no downside risk. The Shopify install guide covers it in about 15 minutes.
Check for AI user agents in your server logs, because Google Analytics filters most of them out. Look for GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Claude-User, PerplexityBot, and Bytespider, and note which paths they request. ChatGPT-User hits in particular mean a real person asked ChatGPT something that led it to fetch your page live. Crawlytics automates this: it identifies AI crawlers and agents, shows which pages they touch and how often, and trends it over time, so you can see whether agent interest in your product pages is growing before the purchase data shows up.
Per visit, often yes, because an agent visit is usually further down the funnel. A human browsing might be window shopping; an agent fetching your product page is typically executing a specific task someone delegated, like "find me a standing desk under $400 and order it." Early evidence on AI-referred human traffic points the same direction, with several published analyses showing higher conversion rates than search referrals. The honest caveat: agent-completed purchase volume is still small, and nobody has clean public numbers yet on Visa-in-ChatGPT conversion. The value today is mostly option value, and it is cheap to position for.
This page is part of Crawlytics.app. View all pages: llms.txt · llms-full.txt