Google Shopping, from the SERP to AI Mode: what a product must prove to be shown, then bought

Take the query Gnoise cancelling headphones

The catalog holds thousands of headphone offers; the product grid shows about twenty. The catalog filter judges offers on four scores. Two judge the text: the product’s relevance, and how close the words that carry weight in the query (“noise cancelling”, “wireless”) are to those in the title and attributes. The other two judge clicks.

Between the catalog and a product block, filters decide which offers Google shows; each block draws on a set of 82. Grouped by what they check, they first ask four questions; the others judge the image, build the rows, take the shopper into account or tune the filtering.

To write a shopping answer, magi, the AI of Google’s results pages (AI Overviews and AI Mode), gives its language model fact sheets in plain text. Five are about shopping: the product, the offer, the reviews, the brand and the store. When the model compares offers, the merchant’s rating, the return policy, shipping and price history sit next to the price.

A tracking tool that queries AI Mode without an account measures the baseline answer. A loyal customer sees an answer where the brand they already bought and their loyalty tier carry weight. Shopping is the area where Google keeps the most personal contexts: 56 types out of 392, far ahead of any other. That is where measured visibility drifts furthest from real visibility.

In the United States, a shopping question asked in AI Mode can close without opening the merchant’s site. Google logs cart actions and direct purchases.

1. The agent picks a product and an offerProductResult · OfferResult
2. The offer goes into the universal cartAddToUniversalCartResult
3. The cart gathers items from several merchantsUniversalCartResult
4. Google follows up: price drop, back in stock, reminder, promotionSHOPPING_UNIVERSAL_CART_*_NOTIF
5. Checkout via UCP, with Google PayUcpCheckoutResult

Outside the universal cart, the sale closes on the merchant’s site, and Google’s agents make the trip for the buyer.

Gemini in Chrome, the browser’s assistant, also calls by name the tools a page exposes through WebMCP, the standard Chrome proposes so a site can describe its actions to agents (add to cart, filter): adding to cart then takes a single call. Our demo RESONEO Shop exposes 135 such tools to an agent.

In France, the scores, the filters and the photo score already apply to free Shopping blocks, with one filter specific to the European Economic Area; the universal cart is open in the United States, agentic checkout in the United States and Australia.

1. Make the product page the reference for the offer

   The Merchant Center feed, Product structured data and the visible page give the same price and the same stock, kept current as sales move. Confidence filters drop an unreliable offer before any comparison. A brand that also sells through resellers should supply the reference data for its products itself (title, description, images, features) in Manufacturer Center: Google shows it on its product pages, whatever the reseller, and in return gives the brand the average price across its points of sale and the competing products shown next to its own.

2. Take care of the title and the main photo

   The title and the photo make the clicks measured by NavBoost, Google Search’s click memory, and anticipated by the predicted CTR. A title that picks up the words that carry weight in the query (type, use, feature) serves relevance and the click; a photo on a clean background, with the product filling 75 to 90% of the frame, serves its score and its visual cluster.

3. Prepare Black Friday by the rules

   Set a strikethrough price that follows Merchant Center’s rules (30 days of reference price within the last 200, a discount above 5% and below 90%), an active and dated promotion and the same price on the page. The five deals filters then judge the promotion’s freshness, whether it is active, how credible the price is, its check by a crawl and the right to show the reason for the deal.

4. Feed what the model reads

   The AI behind AI Overviews and AI Mode reads the return policy, the merchant rating, price history, Q&A, aspect-by-aspect reviews and press and video reviews. Filling in Merchant Center’s conversational attributes, publishing a clear return policy and getting your products reviewed all weigh in the comparison.

5. Add Q&A to the feed

   Merchant Center’s question_and_answer attribute lands in the offer sheet that the AI behind AI Overviews and AI Mode reads. The questions to put there are the ones customer service receives and the ones that keep coming up in reviews. Each pair is written "question":"answer", pairs separated by commas, preferably in a supplemental data source, with no effect on product approval. The manual, linked with document_link, arrives summarized in the same sheet.

6. Open the checkout to agents

   Verify Google-Agent by its published IP ranges, and by its Web Bot Auth signature when the request carries one, rather than by its user agent alone, tune the anti-bot so it does not block a verified request, keep guest checkout and accurate autocomplete fields, test cart transfer. A blocked agent leaves a trace, and the sale can go to a competitor.

7. Get ready for WebMCP

   WebMCP, the standard Chrome proposes, lets a site declare its actions to agents, through annotated HTML forms or tools written in JavaScript. Chrome has opened it for testing to sites that sign up (origin trial) since Chrome 149, and a flag is enough to try it locally (chrome://flags/#enable-webmcp-testing). The first actions to expose are the checkout path ones: search, filters, variant choice, add to cart. Our demo RESONEO Shop exposes 135 of them to an agent.

8. Count the sales analytics does not see

   In the United States, some sales that start in AI Mode close at Google, with Google Pay, or arrive through a transferred cart: the site receives the order without the question, the comparison or the cart that came before. UCP orders and the visits tagged with the utm_content of Chrome’s agent are tracked separately.