AI shopping uses an assistant to help a buyer research, compare or obtain products. Agentic commerce describes experiences in which software also carries out parts of the shopping task on the buyer's behalf. The practical question is which actions the particular service can take and what the buyer must approve.
A recommendation, a prepared cart and a completed purchase are different outcomes. Measuring them together can make a merchant appear successful while the buyer is still unable to get the right item.
Follow one buying task
Consider an illustrative request: “Find a dishwasher-safe bottle that holds at least 700 ml, fits a 7 cm cup holder and arrives before Friday.”
Stage | What the assistant must establish | Merchant information that helps |
|---|---|---|
Discovery | Which products might satisfy the request | Product type, capacity and intended use |
Comparison | Whether each candidate meets the constraints | Base diameter, care instructions and supporting evidence |
Offer selection | Which exact variant can be purchased | Variant, current price, availability and seller |
Cart preparation | Whether the right quantity and variant were selected | Pack size, item identifiers and cart contents |
Purchase | Whether the buyer authorized the actual terms | Delivery estimate, total cost, payment and confirmation |
An assistant that names a bottle but cannot verify its base diameter has not completed the comparison. One that creates the right cart but has no authorization to pay has not completed the purchase.
Where the information can come from
A model can draw on learned information. An assistant may also retrieve pages, use product feeds or call connected services. The available inputs depend on the product, mode and task.
OpenAI documents structured merchant feeds for product information. Google documents product data attributes for Merchant Center. These are provider-specific interfaces, not a universal file that gives every assistant the same access.
A third-party review may help explain a tradeoff; the merchant's current offer is a different source of evidence. For a time-sensitive fact such as stock or delivery, check the current source and timestamp instead of relying on an old descriptive passage.
What changes for the merchant
The product page still needs to serve a person who wants to inspect the recommendation. Clear dimensions, material, compatibility, care instructions and return terms help that person decide whether the item fits the task.
Keep visible content, structured data and feeds consistent. A page that says “in stock” while the selected variant is unavailable creates a practical failure regardless of whether the assistant mentioned the brand.
Check access by purpose. A training crawler, search crawler and user-initiated visit can be different mechanisms. OpenAI's crawler documentation explains its named agents; use the equivalent current documentation for each service you intend to support.
What SEO, AEO and GEO can usefully mean
SEO concerns discovery through search. AEO and GEO are commonly used for improving how content can answer questions or appear in generated responses. The labels do not establish a separate guaranteed ranking system.
For Google's AI search features, the existing search fundamentals apply. Google does not require a special AI text file or additional AI-specific schema. A useful answer, accessible page and accurate supporting information matter more than adopting a label.
Test the buyer outcome
Run a small set of real purchase questions and preserve the answers, cited pages and selected variants. Repeat under recorded conditions; do not use one favorable screenshot as the whole result.
For the bottle example, score each stage separately: identified, constraints verified, offer current, cart correct, and purchase confirmed if that stage was actually tested with authorization. Stop the test at the stage you intend to evaluate. A discovery study does not need to place an order.
Report the limits clearly. “The assistant identified a suitable bottle in four of six valid answers” is an inspectable observation. It does not establish how often all buyers will see the product or how many sales the recommendation caused.
The product-page guide gives a worked data example. The benchmark guide explains how to assess whether a comparison reflects your buyers and category.