Agentic Commerce: what an AI Agent finds on your Ecommerce site

Who do you turn to for advice when shopping? Two out of three Italians have no doubts: more than their best friend, they would trust an AI agent, to use even during the purchase process.

For your ecommerce, that means working not just for people and for visibility on Google, but also for a system that shops on the customer’s behalf: it doesn’t read newsletters, doesn’t open abandoned-cart emails, and only knows about your store what it finds written, on your site and elsewhere.

What Agentic Commerce Is

Agentic commerce is a purchase where a person hands off to an AI agent part of the commercial work they’d normally do themselves: searching for the product, comparing offers, choosing the store, preparing the order and, with their authorization, completing it. An AI agent is a system, built on a language model, that receives a task, decides the steps to carry it out, and uses external tools to get it done: web search, product catalogs, the browser, the systems an ecommerce uses to manage carts and orders.

The task carries the person’s constraints with it: the product, the size, the budget, and the conditions that decide the purchase for them, like free returns or assembly. To carry it out, the agent works with the information it finds about products and stores and with the actions it can perform, on sites or through store systems. What comes back to the person is the result: a recommendation with its reasons, a ready cart, an order to confirm, or a question about whatever the agent couldn’t resolve.

The Step After the Money Prompt

Faced with a request like “what running shoes should I buy under €150,” ChatGPT, Gemini, and AI Mode often already answer with products, stores, and links. Requests like this are money prompts: brand-free, high commercial intent questions addressed to a generative engine, which SEOZoom uses to measure which products and stores the answers recommend. In the purchase recommendation, the engine chooses what to suggest, and the person does the rest on your product page, where they check the price for their size, the delivery, and the return policy, and decide.

Agentic commerce starts from the same request, written as a task: “buy them for me,” “find them and put them in the cart,” “order them for under €150.” The answers to money prompts already show which product an assistant chooses; the task adds what comes next, which until now happened in front of the person, on your pages, and now passes to the software: reading the conditions, comparing them against the constraints, choosing the variant, reaching the cart and, where the person allows it, the order.

Where the task comes from matters too. An ecommerce chatbot speaks on behalf of the store and proposes its own catalog, because the store gives it the task. An agent works for the buyer, even when it acts on a single site, and places your offer alongside the ones the buyer asks it to consider.

How Much the Agent Decides, How Much the Person Does

The constraints of the task and the limits of what the agent can do on its own make up the mandate, which sets how much of the purchase the person hands over to the software. A mandate that stops at the proposal stays close to a buying recommendation. One that extends to the cart leaves the agent to pick the variant, the address, and the shipping; one that extends to the order leaves the agent to finalize it too, either with the person approving the total or with price limits, timing, and conditions fixed in advance. These aren’t mandatory steps: the person may have already chosen the product and the store and hand the agent only the order, or delegate the search while keeping the choice of store.

Any decision the mandate doesn’t cover, or that the agent can’t make with the information it has, goes back to the person: a condition the agent can’t find written down, a price above budget, a variant that’s out of stock. That’s when your offer comes back in front of a person, and not always on your page: in a purchase made inside the platform, the person sees the summary the assistant presents, with the total to approve, and decides whether to go ahead based on that summary.

How an Agent Reaches Your Products

The model an agent runs on already knows the brands and products it encountered in its training data, and that’s the territory of GEO. The agent recognizes brands by name from that memory, but it can’t get today’s price from it, nor the availability of a requested size, nor delivery times: for that information it has to search, on store pages or in product catalogs, unless the person tells it where to go, naming the store or pasting the link to a listing.

Live search starts from the search engine indexes, Google first among them: the assistant runs its queries and works from the pages the engine returns, and to be picked up in that retrieval a listing needs to be indexed and show up among the results for those searches, meaning it needs to rank for the queries in your category. This is AEO territory. With agents, then, AEO doesn’t replace SEO, it depends on it: SEO remains the door through which search passes, and a listing left outside it won’t reach an agent searching live.

Inside Google the link is direct. AI Overviews and AI Mode are still SEO, because they run through the same systems as Search, and to show up as a source a page needs to be indexed and eligible to display a snippet, with no other technical requirements. The search engine also stays present along people’s paths: in the United States, in desktop purchases, 89% of AI-assisted shopping journeys also include a search engine.

The Already-Structured Offer in Catalogs

Your offer can reach an agent even without going through a search, via product catalogs. The feed you upload to Merchant Center feeds the Shopping Graph, the archive of products for sale that Google’s shopping features draw on, combined with structured data and whatever Google reads on the pages. ChatGPT receives product data from direct and third-party providers, and OpenAI uses the Agentic Commerce Protocol, built for buying inside ChatGPT, also to collect merchants’ feeds and promotions. In a feed, price, availability, and shipping travel as fields, product by product, and reach shopping systems without anyone having to open the listing.

A catalog carries whatever is in the fields, and the fields need to say the same thing the page says. Google compares the feed price with the listing price and, if they don’t match, may not approve the product until a new crawl finds them aligned; OpenAI warns that in ChatGPT, price and shipping changes can arrive with a delay. So every time you change a price or delivery terms, you should update the listing, the markup, and the feed together, and state your shipping and return policies once for the whole store in structured data, as Google recommends, leaving exceptions to individual products only.

Blocking or Letting Agents In

In ChatGPT search, pages get in through OAI-SearchBot, and pages from a site that blocks it in robots.txt don’t show up in responses, except as plain navigation links. GPTBot is a different crawler, one that collects content for training the models: it has its own group in robots.txt, and you can block it without dropping out of search.

For agents that open pages at a person’s request, the rule changes. Google states that systems of this kind, like its own Google-Agent, generally ignore robots.txt, and OpenAI says that for ChatGPT-User those rules may not apply: whether it gets through depends on the firewall and bot-protection services. Cloudflare, for instance, distinguishes between search crawlers, training crawlers, and agents, meaning automated activity that acts in real time on a person’s behalf, and lets you handle them separately. Blocking agents closes off every path that depends on that access, including in-browser purchases, while separate categories let you allow them through while still keeping training crawlers out, provided you can identify them.

Some agents identify themselves with a proper name in the user agent and come from published IP addresses, but the user agent can be spoofed, and the stronger proof is request signing, which ChatGPT Work’s cloud browser applies following the RFC 9421 standard and which services like Cloudflare and Akamai verify automatically. Blocking, though, can move the agent without stopping it. ChatGPT’s dots run on a remote machine, and if a site rejects their browser, OpenAI suggests having them retry from the computer the person has connected.

The Reasons AI Gives for Recommending Your Offer

On Google, the person who opens your listing sees it the way you built it, with photos, description, reviews, and price. In an AI response, your offer appears alongside others, with the reasons the assistant chose it written for that person.

On October 15, 2026, using SEOZoom’s AI Prompt Tracker, we submitted twelve mattress purchase requests to ChatGPT, Gemini, and AI Mode, including two that explicitly asked to buy. These are responses, not orders: ChatGPT and AI Mode said they couldn’t complete the purchase, yet they still picked a product and a store. ChatGPT reported each store’s terms with actual figures and used them as reasons for the recommendation: “Free shipping on orders over 79 euros,” “Free returns within 30 days,” and, for one brand, “100-night trial, free delivery and returns, according to the official site.” In those responses, a stated term with its own figure reached the person as a reason to choose that store.

Reasons also come from sites that aren’t yours. Gemini built one of its recommendations on excerpts from a comparison site, and Altroconsumo, Reddit, and Trustpilot showed up in both Gemini’s and AI Mode’s responses: this is the territory of reviews and reputation, the same one as money prompts, where you can only act indirectly and over a long timeframe.

The Same Product, Sold by Multiple Stores

A product often has multiple offers: the brand, retailers, and marketplaces all sell the same pair of shoes at different prices, with different delivery times and services. To compare them, shopping systems need to recognize the same product across different sellers, and they do it through identifiers. In Merchant Center, the GTIN, the number printed under the barcode (the EAN in Europe), links your offer to the product. Without it, Google can’t guarantee the match, and the product’s visibility can suffer.

Among offers for the same item, ChatGPT says it ranks sellers based on factors like whether they’re the manufacturer or primary seller, availability, price, and quality. Being the manufacturer, then, doesn’t guarantee the top spot for the brand’s own site: your product might get recommended and still end up sold by a retailer with a better price or availability, and a retailer can land the order for a product whose reputation it never built. With a mandate that extends to the cart or the order, unless the person has already named the store, the agent is the one who decides where to buy.

The Conditions the Agent Can’t Find

In those same October 15 responses, a condition ChatGPT couldn’t confirm came through as a caveat. Asked for the most affordable pocket spring mattress with old-mattress pickup included in delivery, it recommended the cheapest offer while adding, “I couldn’t confirm, however, that pickup of the old mattress is included,” and listed separately the retailers that did state the service. The store might have offered pickup without ChatGPT finding it; either way, the decision went back to the person with a doubt attached, next to offers that carried no such doubt.

In another response, Gemini stated a delivery time without reading it on the store’s site: for the mattress it recommended, it gave 2 to 5 business days, drawing that figure from the general category of mattresses shipped rolled in a box. A condition the system doesn’t read directly from the store, then, can reach the person either as a doubt or as a generalization about the category.

In a delegated purchase, the options for handling missing information expand. An agent can keep searching, on other pages of the store or on other sites, and it can stop and ask the person before moving forward; it can also present the offer along with the doubt, as happened in the October 15 responses. Through the store’s own systems, it can also get an answer straight from the store, confirming stock, costs, and total, or returning an error. Which path it takes depends on the tools it has, its mandate, and how it’s built: agentic commerce doesn’t guarantee that an agent will only move forward after verifying every condition.

Where Conditions Appear on Italian Stores

What does an agent find when it looks for those conditions on an Italian ecommerce site? On October 19, 2026, we used Claude, through the integrated browser in its desktop app, as an agent on six online stores selling mattresses in Italy, which we selected. The request was the same everywhere: a 160×190 cm queen-size memory foam mattress under 600 euros, with delivery to Milan and pickup of the old mattress; Claude chose the product at each store, and the process stopped at the cart, with no account or payment involved. One visit per store is enough to see where these conditions sit, not to measure how often this happens elsewhere.

Along the way, the conditions of the request surfaced one step at a time. At two stores, the listing opened on the price of the single-size option, and the queen-size price only appeared once it was selected. At IKEA, delivery cost and date appeared only in the cart, after entering the zip code, with the cost shown as a minimum figure; at Amazon, the delivery date changed depending on the zip code. Reading the listing alone stops short of these conditions, and the agent in the browser reaches them by doing exactly what a person would do.

Pickup of the old mattress, on the other hand, wasn’t an option to select at any of the six stores before checkout. The closest thing we found was Mondo Convenienza’s furniture pickup and disposal service: the customer books it after placing the order, by phone or WhatsApp, only at certain store locations, at a price the site doesn’t publish, and the page doesn’t say whether it covers a mattress. An agent finds nothing to select on the site. With pickup as a condition of the mandate, the consistent paths are either a different store or a question to the person; buying without the service requires the mandate to allow it.

A browser-based agent reaches conditions that appear after a choice or after you enter a ZIP code, but not ones your site reveals only before checkout, like in-store pickup for try-ons. Write the conditions your customers ask about where a system reads them before the order: in the listing text and, wherever a field exists to describe them, in structured data and in your feed.

Up to the Order

A mandate that goes all the way to the order sends the agent to your store. It can use the site the way a person would: opening the listing, choosing the variant, entering the ZIP code, adding to cart. With an integration, though, it handles those steps by talking directly to the store’s systems. In protocols like the Agentic Commerce Protocol, built by OpenAI and Stripe, and the Universal Commerce Protocol, the open standard Google built with Shopify and other commerce companies, the agent asks and the store checks stock, validates the address, updates shipping and total, returns errors, and accepts or rejects the order. Conditions that on the site show up in the cart arrive as responses from the store.

An order that reaches your store is still yours. The merchant of record, the actual seller, is the store, in both OpenAI’s and Google’s protocols as well as Shopify’s AI channels, and you keep the payment, through your payment provider, the returns, the refunds, and the customer relationship. An integration, on top of that, doesn’t buy visibility: on Google, choosing native checkout integration has no effect on how Google ranks offers.

Purchases completed inside platforms are, for now, aimed mainly at buyers in the United States, like Google checkout in AI Mode and Gemini or purchases made through Shopify stores’ AI channels, which only serves US buyers. And even there, the store has taken back ground. Walmart saw purchases made inside ChatGPT convert at a third of the rate of customers sent to its own site, and switched to its own app inside ChatGPT, which links the account, loyalty program, and Walmart payments. OpenAI then let merchants use their own checkout, focusing instead on product discovery. The choice can happen inside the assistant, while the account, services, and customer relationship stay with the store’s own systems.

Software on Your Site

To recognize the commands on a page, a browser-based agent can combine a screenshot, the page structure after scripts have run, and the accessibility tree, the description browsers prepare for assistive technologies like screen readers, where every element has a role, a name, and a state. Meta says Muse, its personal agent, navigates with a subagent that sees a snapshot of the accessibility tree, not the page structure, and can’t run JavaScript. An “Add to cart” button a person spots instantly, if it’s built from a generic element and made clickable by a script, doesn’t register as a button in the tree; a field with no label doesn’t say what it’s asking for; a transparent layer over the controls hides them. Google’s guidelines for sites that agents need to use call for native elements, labels tied to fields, and no layer over the controls. It’s the same work that accessibility compliance requires, which since June 28, 2025 the European Accessibility Act has required of online stores, microenterprises excepted.

In the October 19 test, size selection and the add-to-cart button were reachable on all six stores, and the mattress made it into the cart on five of them; on Amazon the add-to-cart failed both through clicks and through a command run on the page, with no error message and for a cause we couldn’t pin down. The accessibility tree also showed up some weak points that didn’t block the add-to-cart: a cookie banner that didn’t appear there, another with unnamed toggles already switched on, sizes offered as generic elements instead of as options, and a paid delivery option whose name was just “Price.”

Paying Within the Task’s Limits

In the Agentic Commerce Protocol, the agent pays without ever receiving the card: the data goes to the payment provider, which issues a single-use token valid up to a maximum amount, within a deadline, and for a single merchant, and the store’s provider charges it. In Google’s Agent Payments Protocol, the person signs a mandate: with the person present, the mandate approves the cart; in their absence, it sets price limits, timing, and the other conditions the agent can buy within beforehand.

Google Pay’s agentic checkout completes the purchase only if the final price is equal to or lower than the one the person confirmed, and a price that rises between confirmation and order cancels the sale. In Europe, strong customer authentication matters too: in the AI-agent transactions on European stores that Visa announced in July 2026, every purchase was tied to a person verified with their passkeys. For your store, payment still runs through the provider you already use.

After the Order

After the order, the store responds to the customer. In the Agentic Commerce Protocol, refunds and disputes stay between you and your provider, and Google, for its agentic checkout, sends people back to the merchant for cancellations, returns, wrong items, and late or missing deliveries. What remains unresolved is liability for an agent’s error: European consumer law assumes a person makes the purchase decision and, per the CERRE research center’s analysis, says little about how to split the risk between consumer, agent provider, and store when the agent buys the wrong product.

This issue weighs on trust: in an Adyen survey of 2,000 Italian consumers, among the things that matter most for trusting AI to complete a purchase, an easy return or refund ranks first (38%), with clarity on who’s responsible for errors close behind (32%). Of the two, returns are up to you, and an agent will pass that along to the person if it finds it written in the product page and the data.

What Shows Up in Your Data

In the tools you use every day, little of agentic commerce shows up. Search Console has a dedicated report for Search’s generative AI features, opened to all sites on August 31, 2026, that counts impressions of your pages in AI Overviews and AI Mode by page, country, and device, with no clicks and no queries; that data is already included in the performance report’s totals. In Analytics, visits from links in recognized AI assistants, like ChatGPT, Gemini, and Claude, land in the AI Assistant channel, which Google added to its default channels in May 2026. Server logs show agents that identify themselves, along with the pages they open, and your store’s systems log orders that come in through an integration, the way Shopify’s admin does with its AI source channel.

Direct referrals from AI platforms account for 0.4% of ecommerce visits, according to Similarweb, and an impression inside an answer isn’t the same as a click. The choice between your store and a competitor’s happens inside the answer or inside the agent’s own process, and it only shows up in your data when it ends with a visit or an order on your site.

Seeing that choice means querying the engines yourself. In SEOZoom, you do this with AI Prompt Tracker, which repeatedly queries AI engines on the prompts you choose and shows, engine by engine, which answers mention your brand and which cite it as a source. Among the metrics is Recommendation Rate, the share of classified mentions where the AI recommends your brand; each answer’s detail view gives you the full text and, when the answer specifies it, the store the AI sends the person to and the terms it cites. Recommendation Rate measures the recommendations you see in the answers, not what agents actually buy. It’s worth reading results engine by engine and over time: in the mattress answers from October 15, four out of twelve queries had no domain in common across any pair of engines, and a single round only captures one moment.

Reading a Response From the Store’s Side

An answer shows the outcome of a choice, not the alternatives it ruled out or the checks it ran. You can see whether your product shows up or is missing, with or without caveats, and, when the answer says so, which store the AI sends the person to and why; you don’t see why another product beat yours, or what the system actually read on your pages.

That outcome tells you where to start checking, not what’s wrong, because an absence or a caveat can also come down to how the engine phrased that particular answer.

In AI Prompt Tracker, then, write prompts the way your customers would, conditions and all: size, budget, returns, whatever service detail actually decides the purchase for the customer. These are the queries where a condition missing from your pages can turn into a caveat, and where you see which store, and for what reasons, the AI sends the person to.