AI SEO · Agentic Commerce · Product Data · Schema.org · 2026

Agentic commerce: how AI agents choose products and what online shops should prepare now

Agentic commerce describes retail in which AI agents search for, compare and increasingly also order products on people's behalf. For online shops this shifts the key question: it is no longer enough for a person to find the product page. An agent has to be able to read the product data, match it correctly and compare it with other offers. This guide explains how AI agents choose products, which data matters and how to prepare your product pages.

by Markus Röhler · November 2026 · 10 min read
In short

Agentic commerce is retail in which AI agents take over buying tasks: understanding the need, researching offers, comparing them and, once the person approves, placing the order. An agent can only choose what it can read in machine-readable form. Complete, structured and up-to-date product data is therefore the prerequisite for a product to make the shortlist at all.

What does agentic commerce mean?

An AI agent is an AI system that pursues a goal over several steps on its own and uses tools to do so: web search, product feeds, shop interfaces or payment services. In agentic commerce that goal might be: “Find lightweight trail shoes for wide feet under 150 euros that will be delivered by Friday.” The agent translates the request into criteria, checks offers and presents the result, or orders within the limits the person has set.

How far an agent goes varies. In practice, four levels can be distinguished:

Level What the agent does Role of the person
1. Research Answers product questions and summarises information Decides and buys in the shop themselves
2. Comparison Builds a shortlist based on criteria such as price, delivery time and attributes Picks from the shortlist
3. Purchase with approval Fills the basket and prepares checkout Confirms the purchase
4. Delegated purchase Buys independently as soon as predefined conditions are met Sets rules, such as a price limit

The technical foundations are being laid right now. Major providers have introduced open protocols through which agents, merchants and payment services exchange data:

As of November 2026, these protocols are evolving quickly, and which purchase features are available in which country keeps changing. For online shops, the individual protocol therefore matters less than whether their product data is readable for any kind of agent.

Context

In February 2024, Gartner predicted that traditional search engine volume would drop by 25% by 2026 as AI chatbots and other virtual agents take over search queries. This is a forecast, not a measurement. It does, however, show the direction in which product research is moving.

Source: Gartner press release, 19 February 2024

Providers do not disclose exactly how their systems weight offers. The following steps therefore describe what an agent technically has to do to complete a buying task, not a ranking formula. That is precisely what is useful for shops: a product can drop out of the selection at every step.

1

Translate the need into criteria

“Lightweight trail shoes for wide feet under 150 euros, delivered by Friday” becomes a category, an attribute (width), a price limit and a delivery deadline. The more precise the request, the more hard filters the agent applies.

2

Retrieve sources

The agent uses web search, the product data of the respective platform, merchant interfaces via commerce protocols, and reviews, forums and user reports. Which sources come into play depends on the system.

3

Extract and normalise the data

Price, currency, availability, variants, dimensions, shipping and returns are pulled from the sources and made comparable. If an item is missing or two sources contradict each other, the product cannot be compared reliably.

4

Filter and weight

Hard criteria such as price, size or delivery time rule offers out. Soft criteria such as user reports, brand awareness or return terms feed into the weighting of the remaining offers.

5

Justify the selection or prepare the purchase

The agent presents a few options with reasons. With a purchasing agent, this is followed by the basket and the person's approval or, for a delegated purchase, the order within the agreed rules.

The difference from traditional search

A search engine shows results and the person filters. An agent filters itself. Anything it cannot match unambiguously does not make the short list the person sees at the end. You can read more about the mechanisms behind this in the guide How AI systems recommend products.

Which product data matters for AI purchases

For an agent, a product page is a collection of facts. The more complete and unambiguous those facts are, the more reliably it can match the product to a request. The following table shows the information that is particularly relevant for comparison and purchase, along with the matching Schema.org property:

Information Why agents need it Schema.org
Unique identifiers Recognise the same product across different shops gtin13, mpn, sku
Brand and exact name Match to the request, no confusion between similar models brand, name
Price and currency Filter by budget offers.price, offers.priceCurrency
Availability Only products in stock qualify for a purchase offers.availability
Shipping and delivery time A common hard criterion (“by Friday”) offers.shippingDetails
Return policy Weighing up the risk before buying offers.hasMerchantReturnPolicy
Condition Distinguish new, used and refurbished goods offers.itemCondition
Technical attributes Hard filters such as dimensions, material, compatibility additionalProperty, material, size
Variants Select the right size or colour ProductGroup, hasVariant, variesBy

Besides the structured fields, the visible text matters. Descriptions should name attributes explicitly: “waterproof to IPX7” instead of “perfect for any weather”, “compatible with USB-C devices” instead of “universally usable”. An agent can only compare what is stated as a fact. The guide Optimise product data shows how to prepare product data systematically.

Reviews: genuine only, with a notice

User reviews feed into the selection of many systems. That is exactly why only genuine reviews should be used. Under EU consumer law (in Germany § 5b(3) UWG), anyone who displays reviews must state whether and how they ensure that the reviews come from people who actually bought or used the product. Fake reviews are not permitted, including in structured markup.

How structured data supports discoverability

Structured data based on Schema.org, embedded as JSON-LD, is the machine-readable version of a product page. A crawler does not have to interpret the layout to find price, availability or identifiers. The fields are named and unambiguous.

Today, product data reaches AI agents in three ways:

What matters is that all three channels deliver the same facts. If price or availability in the feed differs from the product page, an agent cannot use the information reliably. Automatic updates are therefore more important than one-off perfection. The guide Keep Schema.org current describes typical mistakes and how to avoid them.

JSON-LD example: Product with offer (sample values)
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Example Trail 2 running shoe, wide fit",
  "brand": { "@type": "Brand", "name": "Example Brand" },
  "gtin13": "4000000000000",
  "mpn": "BT2-W",
  "sku": "BT2-W-43",
  "material": "Mesh, recycled polyester",
  "additionalProperty": [
    { "@type": "PropertyValue", "name": "Weight", "value": "280 g" },
    { "@type": "PropertyValue", "name": "Heel-to-toe drop", "value": "6 mm" }
  ],
  "offers": {
    "@type": "Offer",
    "price": "129.90",
    "priceCurrency": "EUR",
    "availability": "https://schema.org/InStock",
    "itemCondition": "https://schema.org/NewCondition",
    "shippingDetails": {
      "@type": "OfferShippingDetails",
      "shippingDestination": { "@type": "DefinedRegion", "addressCountry": "DE" },
      "deliveryTime": {
        "@type": "ShippingDeliveryTime",
        "handlingTime": { "@type": "QuantitativeValue", "minValue": 0, "maxValue": 1, "unitCode": "DAY" },
        "transitTime": { "@type": "QuantitativeValue", "minValue": 1, "maxValue": 2, "unitCode": "DAY" }
      }
    },
    "hasMerchantReturnPolicy": {
      "@type": "MerchantReturnPolicy",
      "applicableCountry": "DE",
      "returnPolicyCategory": "https://schema.org/MerchantReturnFiniteReturnWindow",
      "merchantReturnDays": 30
    }
  }
}

Two technical prerequisites are often overlooked. First, the robots.txt must allow the crawlers of AI search services, such as OpenAI's OAI-SearchBot and GPTBot or PerplexityBot. Second, product data should be part of the delivered HTML rather than loaded later via JavaScript, because not every crawler executes JavaScript. The guide llms.txt for AI visibility explains what role an llms.txt can play in addition.

Important

Structured data is no guarantee of a recommendation. It ensures that a system can read the facts about your product correctly. Whether an agent selects it also depends on the offer, price, availability and external sources.

How online shops can prepare their product pages

The following steps can be implemented regardless of which protocol prevails. At the same time, they improve how your products appear in traditional Google search.

1

Take stock

Check a sample of your product pages: are GTIN, brand, price, availability, shipping and returns visible and present in the markup? Where are technical attributes missing that customers typically ask about?

2

Complete and validate the Product markup

Add the missing fields to the JSON-LD and check it with Google's Rich Results Test or the Schema Markup Validator. The markup must match the visible content.

3

Keep feed and page in sync

Price and availability should come from the same source and update automatically: on the page, in the markup and in the feed.

4

Write factual descriptions

State attributes, dimensions, materials, compatibility and intended use explicitly. Promotional wording without facts does not help an agent compare.

5

Answer typical customer questions

Questions about fit, compatibility, care or what is in the box belong on the product page as visible text. Agents readily use such answers in their reasoning.

6

Ensure crawler access

Check robots.txt, deliver product data in the HTML, keep loading times short. A page that a crawler cannot retrieve does not exist for the agent.

7

Measure visibility instead of guessing

Check regularly whether AI systems mention your product for typical purchase queries, and record changes. Only then can you see whether changes to the data are working.

8

Keep an eye on checkout integration

Ask your shop system or payment provider whether and when a connection to commerce protocols such as ACP or UCP will be offered, and for which countries.

For purchase queries in ChatGPT specifically, the guide ChatGPT Shopping optimisation is also worth reading; for discoverability in Gemini and Google AI Overviews, see Optimise AI product search.

Where Feed-AI comes in

Feed-AI turns your details into a structured, crawlable listing page with Schema.org markup, without any programming. The self-learning field audit checks every week which information AI search newly requires, adds the fields and notifies you. From the Pro plan onwards, the AI visibility check measures whether AI systems such as ChatGPT, Perplexity and Gemini mention your product.

Feed-AI is not a checkout provider and does not process purchases. Its contribution is discoverability: your product data should be readable for agents before the purchase comes into play.

Opportunities and limits of agentic commerce

Agentic commerce is neither automatic success nor a distant future. Shops face concrete opportunities, but also limits worth knowing about.

Opportunities

  • An agent that filters by specific criteria can also select providers that rank far down in traditional search results, provided their data clearly meets the criteria.
  • Clean product data works twice: for AI agents and for how products appear in traditional search.
  • New sales channels emerge where customers already ask questions: in conversation with an AI system.

Limits

  • Providers do not fully disclose their selection criteria. Impact can only be measured, not predicted.
  • Many purchase features launched in the US first. Shops in other markets should follow the platforms' announcements.
  • The platform moves between shop and customer. Brand loyalty and direct customer contact may weaken.
  • Trust and liability are not yet settled everywhere. Protocols such as AP2 therefore rely on verifiable proof of user approval.

Legal obligations remain in place even when an agent reads the data: prices must be stated correctly, and incorrect information in the feed or markup is just as misleading as on the visible page. Shops that keep their data clean and current are well positioned for agents and customers alike.

Frequently asked questions

What is agentic commerce? +

Agentic commerce describes retail in which AI agents take over buying tasks on a person's behalf: translating a need into criteria, researching offers, comparing them and, once the person approves or has set rules in advance, placing the order. The agent works with tools such as web search, product feeds and shop interfaces.

How does agentic commerce differ from traditional e-commerce? +

In traditional e-commerce, people search, filter and compare themselves. In agentic commerce, an AI agent handles these steps and only presents a short selection or a ready-made basket. The product page therefore has to be convincing for people and unambiguous for machines.

What product data do AI agents need? +

The most important items are unique identifiers such as GTIN and manufacturer part number, the brand and exact product name, price with currency, availability, shipping costs and delivery time, return policy, technical attributes and the available variants. This information should appear as visible text and additionally as Schema.org markup.

Do online shops need a dedicated interface for agentic commerce? +

Not necessarily for being found. The foundation is complete, crawlable product pages with Schema.org markup and well-maintained product feeds. For purchases directly inside an AI system there are open protocols such as the Agentic Commerce Protocol by OpenAI and Stripe or the Universal Commerce Protocol, which Google introduced with retailers in early 2026. The connection is usually made through the shop system or the payment provider.

Does structured data guarantee that an AI agent will recommend my product? +

No. Structured data ensures that a system can read and compare the facts about your product correctly. Whether an agent recommends it also depends on the offer, price, availability, external sources and the undisclosed selection criteria of each provider. That is why it pays to measure actual visibility regularly.

What are ACP, AP2 and UCP? +

They are open protocols for agentic commerce. The Agentic Commerce Protocol (ACP) was developed by OpenAI and Stripe and handles checkout between agent and merchant. The Agent Payments Protocol (AP2) was introduced by Google in September 2025 and uses signed mandates to prove what the user has approved. According to its project site, the Universal Commerce Protocol (UCP) covers the journey from catalogue search through basket and checkout to order management.

Make your product data readable for AI agents

Feed-AI turns your details into structured, crawlable listing pages with Schema.org markup and, from the Pro plan onwards, measures whether AI systems mention your product.

Get started →

from €19/month, cancel monthly