Agentic commerce is shopping in which an AI assistant does part of the buyer’s job: it searches, compares, shortlists and, in some cases, completes the purchase. Instead of typing a query and opening ten tabs, the shopper asks ChatGPT, Gemini, Copilot or Amazon’s assistant, and the assistant decides which products to show.
For a store, the question shifts from «how do I rank on Google?» to «how do I get chosen by an agent?». The short answer is that agents choose products whose data they can read and trust, which is the same discipline behind a well-run Google Merchant Center account. If you want that foundation built for you, that is what our Google Merchant Center agency work is about. This guide explains how the choice is made and how to check your own store.
What is agentic commerce, and how is it different from AI search?
Agentic commerce goes beyond answering a question: the assistant takes steps on the shopper’s behalf. It helps to think of three levels, because each asks something different from your store. This is our own way of framing it, not an official industry scale.
| Level | What the assistant does | What your store needs |
|---|---|---|
| Answers | Recommends products in a chat reply and links to the seller | Readable product data and pages that crawlers can access |
| Assists | Compares options, narrows a shortlist, hands the shopper to your checkout | Price and stock in the feed that match the live product page |
| Acts | Starts or completes the purchase inside the chat | Participation in a checkout program or protocol, on top of everything above |
Most US stores are dealing with the first two levels today. The third depends on platform and protocol decisions that are still moving, which is why we recommend getting levels one and two right before worrying about it.
Where agentic commerce is happening
- ChatGPT: OpenAI launched Instant Checkout in September 2025, built on the Agentic Commerce Protocol it open-sourced with Stripe.
- Google: Google and Shopify announced the Universal Commerce Protocol (UCP) in January 2026, an open standard for agents to query products, build carts and start checkout.
- Microsoft: Copilot Checkout launched in January 2026, with PayPal, Shopify and Stripe as partners.
- Amazon: its shopping assistant was renamed Alexa for Shopping in May 2026, with features that can buy on the shopper’s behalf.
The landscape moves fast. Some in-chat checkout plans have already been scaled back, as we explain in Discover in AI, buy on your site. What has not changed is where the agents get their information: product data. If you want to see how the underlying protocols differ, read Google UCP vs OpenAI ACP.
How does an AI agent choose which products to show?
It filters on whether it can match the request to your data, then on whether the offer looks real and trustworthy. Four checks do most of the work.
1. It needs structured data it can trust
Agents rely on structured product feeds and catalogs: titles, prices, stock, identifiers, attributes, shipping and returns. Data scraped from web pages is often out of date, so feeds that are complete and current have an advantage. In practice that means a product has a clear title, a description that matches the item, a price, an availability status, images, a brand, a GTIN where one exists, and shipping and return information. We break down the fields that matter in the product data AI shopping agents read.
2. It matches the request, not a keyword
A shopper asks for «a waterproof hiking jacket for women under $200 that ships this week». The agent needs product type, gender, material or feature, price and delivery time to answer. Every constraint in the request is a field it has to find in your data. A product with those fields filled is a candidate; a product without them is invisible for that request, however good the jacket is.
3. It checks that the offer is real
Price, availability and delivery must be consistent between the feed and the store. Inconsistent data is a reason to skip a product, because an agent that sends a shopper to a different price or an out-of-stock page has made a bad recommendation.
4. It weighs trust signals
Clear return policies, business information and reviews help an agent recommend a store with confidence. These are the signals a careful human shopper would look for too, which is a useful test: if a new customer could not find your return terms in under a minute, an agent probably cannot either.
What does this look like for a real catalog? (A hypothetical example)
Imagine a store with a few thousand SKUs selling outdoor apparel, and a shopper who asks for a women’s waterproof hiking jacket under $200 that ships this week. This is an illustration of the method, not a real account.
When we audit a catalog like this, we read the feed the way an agent would, one constraint at a time. The typical findings look like this:
- Titles such as «Alpine Pro Jacket – Black» say nothing about gender or waterproofing, so the product fails two of the shopper’s constraints before price is even considered.
- The gender attribute is filled for some products and blank for others, so only part of the range can be matched.
- The word «waterproof» appears in the description but not in the title or in a material or feature attribute.
- Delivery time is set once for the whole account, so the agent cannot tell which jackets really ship this week.
- The sale price on the product page changes at midnight, but the feed updates once a day in the morning, so for hours the two disagree.
The fix order matters more than the fix list. We start with the products that earn the most revenue, rewrite titles and fill attributes there, then correct shipping, then close the price-sync gap. Doing the top 10% of the range properly beats doing the whole catalog badly.
How can I check whether my store is ready?
Run this audit yourself in an afternoon, using a feed export and your own product pages.
- Export your product feed or catalog and sort it by product type, so you can see each category as a group.
- Pick the 20 products that earn the most revenue. Read each title and attributes cold, as a stranger would: can you tell what it is, who it is for and what it is made of?
- For the same 20 products, compare price and availability in the feed, on the product page and in the page’s structured data. Write down every mismatch.
- Check identifiers. Every product that has a GTIN should carry it, and brand should be filled everywhere.
- Check shipping and return information at product level, not just account level, and confirm it matches the policies on your site.
- Check that your robots.txt and any bot-protection rules do not block the crawlers of the assistants you care about.
- Measure how long a price or stock change in your system takes to reach the feed. That delay is your window of inconsistency.
- Rank the gaps by the revenue of the products they affect and fix from the top down.
If you want the full version, the agentic commerce readiness checklist goes through each area in detail.
What mistakes keep products out of AI recommendations?
Most failures are data problems that show up as a pattern in the feed. Use the symptom column to spot them.
| Mistake | Symptom in your data | Why an agent skips it |
|---|---|---|
| Keyword-stuffed or vague titles | Titles repeat a phrase, or lack product type, brand or key attribute | It cannot match the product to a specific request |
| Details only in the description | Attribute columns empty while the description mentions size, color or material | Structured fields are what gets matched; free text is a weaker signal |
| Stale price or stock | Feed value differs from the live page for the same item | The offer looks unreliable |
| Variants not grouped | Each size or color is a separate, unlinked item | The product looks like dozens of near-duplicates |
| Missing identifiers | GTIN blank on products that have one | It is harder to confirm what the product is |
| Generic policies | One shipping and return rule for every product | It cannot answer delivery or return questions per item |
What this means for your store
Most of the work is the same work that makes Google Merchant Center perform: a clean, complete, up-to-date catalog. That is why we treat Merchant Center as the foundation and build every other channel on the same data. Our guide to one product feed for every AI channel shows how the pieces fit together. To see how each assistant differs for shoppers, read ChatGPT vs Gemini vs Copilot.
When is agentic commerce not a priority for you?
It is not your first job if any of these apply:
- Your Google Merchant Center feed still has many disapproved or incomplete products. Fix that first, because every other channel inherits the same problems.
- Your products have no public price or are sold by quote only. An agent cannot compare what it cannot price.
- Most of your sales run through a marketplace that controls the listing. In that case the data that matters most is the data on the marketplace listing.
- Your catalog is a handful of products and changes rarely. A careful manual review may be all you need, with no tooling or agency involved.
How Avafa works
Avafa is a consultancy based in Spain that works remotely with ecommerce brands, manufacturers, distributors and retailers selling to the US. We start with an audit of your catalog against the checks above, build or repair the feed in Google Merchant Center, and then extend the same data to Microsoft Merchant Center and AI channels. We do not promise placement in ChatGPT, Copilot or any AI answer, because no one controls that. What we control is the quality and consistency of the data. You can see the scope of that work on our Google Merchant Center agency page, and you can start on your own with the readiness checklist.
FAQs
Is agentic commerce only for big retailers?
No. Platforms such as Shopify have made catalogs from millions of stores discoverable in AI assistants. What matters is the quality of your product data.
Do I need to sell inside ChatGPT to benefit?
No. Many purchases start with AI research and finish on the store’s own website. Being recommended is the first goal.
Where should I start?
With your Google Merchant Center feed. If it is complete and accurate, most AI channels can use the same data.
Is agentic commerce just SEO under a new name?
Not exactly. Classic SEO optimizes pages for a ranked list of links, while agents also read structured product data and compare it against a request. The two overlap, since clear pages and crawl access help both, but a strong website alone does not fix a weak feed.
Do I need a separate product feed for each assistant?
Usually not. One well-built catalog can be distributed to several channels, with small adjustments for each channel’s requirements.
How do I know if an agent is sending me traffic?
Start with your analytics referral sources and landing pages, and tag what you can. Attribution from AI assistants is still imperfect, so treat the numbers as directional and read how to measure AI as a discovery channel.
