For twenty years, digital commerce has been designed for people looking at screens. Hero images, persuasive copy, urgency badges and carefully tuned page layouts. All of it assumes a human is looking.
That assumption is starting to break. A growing share of product discovery begins inside an AI assistant, and the assistants are learning to buy. The shopper still decides what she wants. Increasingly, a machine does the looking, the comparing and sometimes the paying.
The buyer changes, so the storefront changes
When a shopper asks an assistant to "find a waterproof hiking jacket under $300 that can arrive by Friday", the agent doesn't admire your lookbook. It reads structured data: materials, sizes, waterproof rating, price, stock by location, delivery promise and return policy. Then it compares.
If your catalog says "great for the outdoors" where a competitor's says "20,000 mm waterproof, taped seams, ships tomorrow from Mississauga", the agent picks the competitor. It isn't persuaded. It is matching.
Look at what OpenAI asks of a merchant who wants products to appear in ChatGPT shopping. The product feed specification requires nine fields: a stable item ID, title, description, product URL, brand, seller name, image, availability and price. It recommends many more: GTIN and MPN identifiers, material, color, size, shipping price, return deadline and policy, review count and star rating. Nothing on that list is about brand story or photography. Every item is a fact a machine can read and check.
What is real today
Open standards for agent checkout and payment authorization exist, and major platforms have experimented with in-assistant checkout. They are also moving quickly. Early in-chat checkouts have launched, expanded and been withdrawn within months, and retailers are testing branded apps inside assistants instead of handing over a generic checkout. Treat every specific fee, date and program as provisional. Our short guide to what each of the main protocols does sets out the moving parts.
The traffic is real. Adobe measured visits from generative AI tools to US retail sites rising 693% across a holiday season, and those visitors converted better than other traffic. (Adobe) The open question is which stores the agents can buy from.
Where stores fall short
When we look at catalogs through an agent's eyes, the same gaps appear again and again.
- Attributes live in free text. The waterproof rating is somewhere in a paragraph, and no field holds it.
- Suppliers send data in different units and formats, so one product says "L" and its twin says "Large" or "42 cm".
- Stock and delivery are only visible after a customer enters a postal code.
- Return terms sit in a PDF, and the product page says something slightly different.
- Prices differ between the site, the feed and the marketplace listing.
None of these gaps is dramatic. Together they make a store hard to quote, and an agent that can't quote a store moves on to one it can.
The no-regret investment
Whichever protocol or channel wins, agents need the same thing: complete, accurate, machine-readable product data and live price and availability. That work also improves on-site search, filtering and conversion for human shoppers now. If agents never arrive in your category, you still have a better catalog.
Start with three questions.
- What share of your SKUs have complete, structured attributes for the way customers actually search?
- Can an outside system get your live price, stock and delivery promise without scraping a page?
- Are your shipping, returns and warranty policies explicit enough for a machine to quote back?
If any answer is "not really", that is the roadmap.
A first month
A small team can do a useful amount in four weeks.
In the first week, pick a category of a few hundred products and score each against the fields above. Count the blanks. The blanks are your backlog, and the count is your baseline.
In the second week, fix the highest-traffic products by hand and look for patterns in what is missing. Attributes that are empty for a whole supplier point to a supplier problem. Attributes that are empty across the board point to a data model problem.
In the third week, use a language model to extract attributes from existing descriptions and images into the missing fields, with a person reviewing a sample. This is where the largest gains usually come from, and where catalog enrichment pays off.
In the fourth week, publish a clean feed for the category and put the questions your customers ask to a few assistants. Record who gets named and what is said. Repeat monthly. The agent-readiness checklist gives you twenty questions to score the rest of the store the same way.
Who owns it
This work crosses merchandising, engineering, marketing and legal, and it stalls when nobody is named. Pick one owner, give them the baseline, and review the number every month. A catalog that agents can read is an asset that builds. Every attribute filled in stays filled in.