Can an AI agent buy from your online shop? Visa’s test of 5 November, what Clarity can see, and what to check in your store.
On 5 November Visa published the first test of AI agents doing the shopping: they fail 42.8% of runs, and none of the runs happens on a shop’s website. Here is the checklist for yours.
Updated 11 November 2026.
Yes, an AI agent can already try to buy from your website the way a person would; in Italy the direct link to the assistants’ checkouts is still missing. The OpenAI and Google protocols are for approved partners, in the United States or through a waiting list.
In Visa’s test of 5 November 2026, 22 AI models failed 42.8% of shopping runs, and in 54% of the failures they bought something anyway. The test measures product choice, not checkout.
Since 4 November Microsoft Clarity, which is free, shows which agents visit a website and where they get stuck. On 11 November the European Commission proposed the first EU rules on purchases made by agents.
The eight-line checklist, with the per-product sheet, can be downloaded below in Word and PDF.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The “can my shop handle an agent that buys?” checklist, to download
Eight checks for an online shop, from the simplest to the most technical, each with the source it comes from, plus a sheet with one row per product and the EU rules proposed on 11 November. The first two lines take one free tool; the rest is work on product pages and checkout, which already helps your flesh-and-blood customers.

| File | Contents | Link |
|---|---|---|
| Wordto fill in, in Italian | The eight checks with source and how to verify them, the per-product sheet and the proposed EU rules. | |
| PDFto print, in Italian | The same checklist, to tick off with a pen, product page by product page. |
| N. | What to check | Where it comes from |
|---|---|---|
| 1 | Clarity installed, with the Agent Activity dashboard: which agents arrive and where they stop. | Microsoft Clarity |
| 2 | A simulated purchase of your best seller, from product page to confirmation. | Microsoft Clarity |
| 3 | Every product page has a name, image, price with currency, availability, brand, and a GTIN or manufacturer part number. | Google, OpenAI |
| 4 | Price and availability match across the page, the structured data and the feed. | Google, OpenAI |
| 5 | Shipping and returns spelt out, with costs, lead times and days. | OpenAI, Google UCP |
| 6 | The product page answers every question a customer might ask: sizes, compatibility, variants. | Visa test |
| 7 | Checkout can be completed: promotions visible, guest checkout, a clear confirmation. | Microsoft Clarity |
| 8 | Bot protection recognises verified agents instead of blocking them all. | Visa, Cloudflare, Akamai |
On 7 September I published my analysis of the websites of 142 Italian companies that also sell abroad. In 61 of them a foreign buyer had no real way to get in touch, and in nineteen the online shop could not be used from abroad. A person facing those shops gives up. An agent sent by that person stops at the same point, and tells nobody.
I closed that post with a line that matters even more now: “Nobody will ever give a conference talk about a contact form that works.” The full reasoning is in the post on the 142 websites. This is why lines 3 to 7 of the checklist deal with product pages and checkout, and only the last one with technology.
Can an AI agent buy from my website today?
Yes, in the simplest way: an agent with a browser opens the product page, adds to basket and fills in the checkout as a person would. Clarity, Microsoft’s free website analytics tool, already counts these visits for agents run by Google, OpenAI and Anthropic. What an agent sees on a page is explained in the guide to agent-ready websites.
A direct connection is a different matter. With agentic commerce protocols the assistant talks to the shop through a feed and APIs, and the customer pays without leaving the chat.
For an Italian shop, as of 11 November 2026, this route is not documented. OpenAI only admits approved partners and its standard feed targets the United States; Google UCP has a waiting list and does not name the countries. The details are in the guide on how to sell on ChatGPT.
In Europe, though, real purchases have already happened. On 2 July 2026 Visa announced live transactions in which AI agents searched for products, chose them and started the purchase on the websites of independent merchants enrolled in its programme, including lastminute.com and Frasers. Nexi is among the banks that completed transactions. On 2 June Mastercard described controlled transactions on its own Priceless.com platform, with no Italian bank on the list.
Visa’s test of 5 November 2026: what it really measured
Visa put 22 AI models from 12 labs through 15 shopping scenarios, such as choosing a TV, a laptop or a small personal care appliance, with three attempts per scenario. The AI had to talk to the customer and uncover every requirement, including the unspoken ones, then decide whether to buy or not.
A run passed only if the AI uncovered every requirement and made the right call, which sometimes meant buying nothing. The press release puts it plainly: “models failed 42.8% of test runs”. The study comes from Visa, a company that sells infrastructure for agentic commerce, and the release does not say which models did better or worse.
| What Visa measured | Detail | Value |
|---|---|---|
| Runs passed | Average across all models and all scenarios. | 57.2% |
| Runs failed | A requirement missed or a wrong decision. | 42.8% |
| Best model | Share of attempts passed. | 82.2% |
| Worst model | Share of attempts passed. | 28.9% |
| Three attempts out of three | The same scenario passed every time. | 34.2% |
| Failed runs with a purchase | Failed runs in which the AI bought anyway. | 54% |
| Models that uncover more requirements | The five models best at drawing out what the customer needs, against the five worst. | 70.2% against 44.4% |
| Most expensive models | The five with the highest average cost per run, against the five cheapest. | 67.1% against 50.2% |
What it leaves out: checkout, payment and order placement on a real website. Visa writes that it “will extend testing to checkout, payment, and order placement”. Reporting the 42.8% as “agents that cannot buy on websites” claims something the test has not tried. The benchmark can be explored publicly on Visa Payments Frontier, with replays of the attempts.
Why AI shopping agents fail to buy online
In Visa’s test they fail mostly when they do not find out what the customer really wants. The five models best at drawing out hidden requirements pass 70.2% of runs; the five weakest, fewer than half. In the release’s example the requirements are budget, format compatibility and HDR10+ support for family videos.
The test says nothing about shops, but one practical point follows from it. An agent cannot check a requirement that appears nowhere on the product page, even when the customer has told it. So line 6 of the checklist asks for product pages that answer every question: sizes, compatibility, variants, what comes in the box.
The second reason sits on the website. Clarity’s blog says: “An agent might find the right product but miss a promotion, or reach checkout and get stuck.” A promotion that only appears after a click, a mandatory account, an extra step at checkout: for a person these are irritations, for an agent they can be where the sale stops.
How to see whether AI agents visit your website: Clarity Agent Activity
Since 4 November 2026 Microsoft Clarity has a dashboard, Agent Activity, showing how much agent traffic arrives over time, which agents they are, what they do and where they get stuck. The actions counted are searches, forms, logins, downloads, bookings and checkouts. Like the rest of Clarity it is free, and Microsoft says it is available worldwide.
Next to the dashboard come simulations: Clarity tests whether an agent can complete a key task on the website, and reports whether it did, how long it took and how many interactions it needed. It starts with commerce and travel scenarios, so an online shop is among the first cases covered. This is line 2 of the checklist.
The other two launches of 4 November come with limits. Brand Agents, the shopping assistant you add to your website, brings products from eligible WooCommerce and Shopify stores into Copilot, Bing and MSN, but the blog does not say in which countries. Ask Clarity, the plain-language questions about your data, is in closed beta with a waiting list for retailers.
Citations in AI answers, which Clarity measures in a separate panel, are covered in the guide on how to measure GEO.
What agents do on websites: Clarity’s study of 5.2 million sessions
On 29 October 2026 Clarity published an analysis of more than 5.2 million agent sessions across 230,511 websites, from 13 August to 9 September 2026. It is Microsoft analysing its own data, not an independent study, and it covers agents from the OpenAI, Google and Anthropic ecosystems.
| Action recorded | What it means | Sessions with agents |
|---|---|---|
| Checkout | The session reached the till. In human-only sessions the figure is 14.7%. | 25.2% |
| Form submissions | Contact, quote and sign-up forms. | 19.9% |
| Outbound clicks | Links to other websites. | 14.3% |
| Bookings | Appointments, tables, rooms. | 8.4% |
| Searches | The website’s own search. | 7.2% |
| Login | Signing in to an account. | 6.8% |
| Add to basket | The product placed in the basket. | 2.3% |
The study states its own limit: the actions are events recorded in sessions, “not necessarily actions completed autonomously by an agent”. A session can pass from an agent to a person, and Clarity does not separate the two. The 25.2% says agents often reach the till. How many purchases they completed on their own, nobody has measured.
One piece of context, from another analytics vendor: according to Adobe, in September 2026 visits to US retail websites arriving from a link in an AI assistant brought 83% more revenue per visit than other visits. It measures people clicking, not agents buying.
Agentic commerce for an Italian shop: protocols and payments
For an Italian shop today, the OpenAI and Google protocols are something to prepare for, not something to switch on. In both, the shop remains the seller: OpenAI writes “OpenAI is not the merchant of record”, Google “You remain the Merchant of Record”. The shop takes the money, works out tax and shipping, and accepts or declines the order.
What you can do right now is have the data both protocols ask for. OpenAI’s feed wants title, description, brand, image, availability and price, and treats an empty shipping cost as unknown rather than free. Google UCP relies on Merchant Center and asks for complete return policies and contact details. These are the same fields as the Product and Offer structured data Google has read for years.
On payment, Europe has one extra rule, the strong customer authentication of PSD2: Visa and Mastercard use passkeys for precisely this reason.
Then there is bot protection. Visa Trusted Agent Protocol lets a verified agent be told apart from other traffic, and Cloudflare and Akamai support it: a firewall that blocks every bot also blocks the customer who sent an agent. The crawlers and agents worth recognising are listed in the glossary of AI crawlers.
| Where | Agents that buy, as of 11 November 2026 | Source |
|---|---|---|
| Italy | No documented route to connect to the ChatGPT or AI Mode checkouts. Nexi among the banks in Visa’s transactions. | OpenAI, Google, Visa |
| EU | Live agent purchases on the websites of merchants enrolled in Visa Agentic Ready, with passkeys for PSD2. Rules proposed in Article 26 of the Digital Fairness Act. | Visa, Commission |
| United Kingdom | Frasers is among the merchants in Visa’s transactions. The Digital Fairness Act does not apply; no rules on agents read. | Visa |
| United States | OpenAI’s standard feed for the United States (the format also covers Canada and Mexico), checkout for approved partners. No federal rules on agents read. | OpenAI |
On 10 November OpenAI also opened Business Hub, where a company claims its own listing in ChatGPT: how it works and what applies outside the United States is in the guide on how to claim your business in ChatGPT.
The EU rules on the way: Article 26 of the Digital Fairness Act
On 11 November 2026 the European Commission proposed the Digital Fairness Act, and its Article 26 regulates for the first time automated systems that conclude contracts on a consumer’s behalf, AI agents included. As a proposed regulation, it has to go through the European Parliament and the Council, and would apply 12 months after entry into force, which has not happened yet.
Almost all the obligations would fall on the provider of the agent: letting the consumer choose which contracts to confirm, showing the pre-contractual information, disclosing sponsored suggestions. For the shop there would be three lines. The agent would pass on the customer’s contact details (Art. 26(7)); the shop would send the contract confirmation on a durable medium (Art. 26(8)); and it could not require customers to use an agent to access the shop or buy (Art. 26(9)).
Recital 78 adds that using an agent would not cost consumers any of their rights, the right of withdrawal included. The full proposal, with dark patterns, pricing and subscriptions, is in the guide to the Digital Fairness Act.
What is confirmed and what is not
News about agents that buy mixes lab tests, pilot programmes and draft laws. This table keeps the three apart, with the date of each fact and its source.
| Fact | Date and source | Status |
|---|---|---|
| AI agents fail 42.8% of shopping runs | 5 November 2026, Visa | Confirmed, in a lab, on product choice |
| Visa’s test covers checkout on a website | 5 November 2026, Visa | Announced, not live yet |
| Clarity shows the agents visiting a website | 4 November 2026, Microsoft Clarity | Confirmed, for Google, OpenAI and Anthropic |
| Brand Agents for Italian shops | 4 November 2026, Microsoft Clarity | Not stated: countries not listed |
| Checkout in 25.2% of agent sessions | 29 October 2026, Microsoft Clarity | Confirmed, does not prove purchases completed by the agent |
| AI agents have bought on European shops’ websites | 2 July 2026, Visa | Confirmed, for merchants in the programme |
| An Italian shop connects to the ChatGPT checkout | OpenAI, documentation read on 11 November 2026 | No: standard feed for the United States |
| An Italian shop sells in AI Mode through UCP | Google, FAQ updated on 3 November 2026 | Not stated: waiting list |
| EU rules on purchases made by agents | 11 November 2026, European Commission | Proposed, not yet law |
| OpenAI has scaled back Instant Checkout | Reported by the press | Unconfirmed report, no primary source |
Where a shop starts
With lines 1 and 2: install Clarity and watch for a few weeks which agents arrive and where they stop. Then the product pages, lines 3 to 6, starting with the best sellers, and the checkout, line 7. Prices and availability written by a script escape anything that reads the HTML: the subject of the guide on JavaScript and AI crawlers.
I learnt this on my own website before writing it here. In August ChatGPT could not find the founder of itria.io, and I had already done the things people usually recommend. In the post of 21 September I put it like this: “AI cites the rule-maker or the firm doing the work. Never the explainer.” With agents that buy it works the same way: what counts is the product page that answers, not the page that tells a story.
The shop closed to foreign buyers. Of 142 websites of Italian companies that also sell abroad, 19 with an online shop that could not be used from outside Italy and 61 with no real way to get in touch, from the post of 7 September 2026.
The email address that was not there. An address removed from the HTML and put back by a script: for anything reading the raw HTML, the contact page was empty. The same flaw on 14 of 121 food producers’ websites, in the piece on the email address AI assistants cannot see and in the study of producers’ websites.
The result. Between 11 October and 7 November 2026 Bing’s AI Performance report counted 268 citations of itria.io pages in Copilot, and Search Console 3,550 impressions in AI Overviews and AI Mode.
The index of all the guides is in marketing for SMEs. How the work is done on a real company, so that assistants find it and choose it, is described on the page about presence on search engines and AI assistants.
Questions and answers
Can an AI agent buy from my website?
It can already try, the way a person would: it opens the product page, adds to basket and fills in the checkout. Agents from Google, OpenAI and Anthropic visit websites, and Microsoft Clarity has been counting them since 4 November 2026.
What is missing in Italy is a direct connection to the assistants’ own checkouts: the OpenAI and Google protocols are open to approved partners, in the United States or through a waiting list.
What did Visa's test of AI shopping agents measure?
Whether 22 AI models from 12 labs, across 15 shopping scenarios, uncover every requirement the customer has and decide correctly whether to buy. They failed 42.8% of the runs.
Checkout, payment and order placement on a real website are not in the test yet: Visa says it will add them.
How can I see whether AI agents visit my website?
With the Agent Activity dashboard in Microsoft Clarity, free, announced on 4 November 2026: it shows which agents arrive, how many sessions they run, what they do and where they get stuck.
For now it recognises agents run by Google, OpenAI and Anthropic.
Can an Italian shop sell inside ChatGPT or Gemini?
As of 11 November 2026 there is no documented route. OpenAI’s standard feed targets the United States and checkout is for approved partners; Google’s UCP protocol has a waiting list and does not name the countries.
What both ask for can be prepared today: prices, availability, shipping and returns.
Does a customer who buys through an AI agent keep the right to withdraw?
Today the consumer law in force applies, and it does not mention agents; no official source says the right of withdrawal changes. The Digital Fairness Act proposed on 11 November 2026 states that using an agent does not take away any consumer rights.
The text is still a proposal, not yet law. This page is for information only and is not legal advice.
Notes on sources
- Visa, Visa Releases First-of-Its-Kind Industry Benchmark for AI Shopping, 5 November 2026: sample, method and results of the Visa Agentic Commerce Benchmark, and the source of the two quoted lines. A study by a payments company. Re-read on 11 November 2026.
- Microsoft Clarity, Agents act, LLMs recommend, and people browse, 4 November 2026: Agent Activity, simulations, Brand Agents, Ask Clarity, supported agents. Re-read on 11 November 2026. Also summarised on the Microsoft Advertising blog.
- Microsoft Clarity, What AI Agents Do on Websites: Insights From 5.2 Million Sessions, 29 October 2026: the vendor’s analysis of its own data, 230,511 projects, 13 August to 9 September 2026.
- OpenAI, Agentic Commerce Protocol, key concepts and the product feed specification, read on 11 November 2026.
- Google for Developers, UCP guide and FAQ updated on 3 November 2026; Google Cloud, Agent Payments Protocol (AP2), 16 September 2025.
- Visa, Visa and Banks Across Europe Reach the Next Phase of Agentic Commerce, 2 July 2026; Mastercard, Europe is Building the Foundations for Trusted Agentic Commerce, 2 June 2026. Company sources on their own programmes.
- European Commission, proposal for a Digital Fairness Act, COM(2026) 1100 final, 11 November 2026, Articles 26 and 27 and Recital 78; press release IP/26/2407. A proposal, not law in force.
- Google Search Central, Product and Offer structured data for merchant listings, updated on 8 September 2026, CC BY 4.0 licence.
- Adobe, AI Traffic Drives 83% Higher Revenue Per Visit: data from an analytics vendor, US retail websites, September 2026.
- The cases come from the post of 7 September 2026, the post of 21 September 2026 and the itria.io labs; the itria.io figures from Search Console (generative AI report) and Bing Webmaster Tools, window 11 October to 7 November 2026.
The checklist tells you whether an agent can buy from you. Why it picks you over a competitor is the next piece of work. Positioning is built starting from the market.
Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.