AI Shopping Starts With Your Product Feed, Not Your Product Page

When ChatGPT recommends products, it pulls from your Google Merchant Center feed, not from your product page. The feed is now the listing, and a feed that has not been touched since setup is a listing that has not been touched since setup.

The carousel comes from the feed

Ask ChatGPT to recommend a product and it shows a carousel of eight options. In a March 2026 study of more than 43,000 products, 83 percent matched Google top 40 organic Shopping results, and the order in the carousel matched the ranking in Google Shopping. The products do not come from the open web, from product detail pages, or from reviews. They come from the Google Merchant Center feed, the file most brands have not opened since they set up paid Shopping. The product page has taken a back seat. The feed is the listing now.

A second study, from Profound, reviewed more than one million ChatGPT shopping offers in June and found that feed-sourced offers populated brand, product image, and merchant details 100 percent of the time, compared with 0 percent for page-scraped offers, and carried the best price tag 100 percent of the time, compared with 21 percent. The share of feed-sourced retrievals grew from 4.3 percent to about 20 percent of all ChatGPT shopping retrievals in six weeks. The feed gives the model clean, structured fields instead of forcing it to infer brand, price, and availability from a page. That is why the feed wins the citation and the page does not.

The six new fields, and the two worth filling first

Google added six new Merchant Center attributes at Marketing Live 2026, all optional, all available in every country: question and answer, document link, related product, item group title, variant option, and popularity rank. The help pages for the first four open with the same line: the attribute is primarily intended for conversational experiences such as AI Mode. Shopping ads do not get a mention. The fields exist for AI shopping, and they are the cheapest way to give the model more to work with.

Fill question and answer first. Up to 30 pairs per product, capped at 1,000 characters each and 10,000 characters total, which is more text than most product pages carry. Base them on the questions customers actually ask: fit, compatibility, what is included, what it replaces. Then fill product highlight, the Amazon-style bullet points that tell the model what the product does in benefit language, not spec language. The other four fields matter, but those two are the ones that change whether the model recommends your product and what it says about it. Use TSV, not CSV, for the upload, because answers are full sentences and full sentences contain commas.

What still comes from the page

The feed decides inclusion and ranking, but most offers still come from the product page. In Profound data, about 88 percent of ChatGPT product offers still came from web product detail pages rather than feeds, and even for merchants already using feeds, around 76 percent of offers still came from the page. The feed helps with ranking and with the completeness of the offer, but the page is still where most offers originate, which means the page still matters, just differently than it used to.

What the page needs now is the structured data that lets the model trust it: Product JSON-LD with price, availability, item condition, shipping details, and return policy, kept current and matching the feed exactly. A price on the page that disagrees with the price in the feed is a contradiction the model resolves by trusting neither, or by trusting the feed and dropping the page. Keep the two in sync, and the page and the feed reinforce each other instead of competing.

The feed check to run this week

Open Merchant Center, open the feed diagnostics, and read the warnings, not just the errors. Errors block the product. Warnings degrade it, and degraded products are the ones the model skips. Check that every product has a current price, an availability value that matches the page, a GTIN where one exists, and an image that is not a placeholder. Then add the question and answer pairs for the ten best-selling products, and watch the AI referral traffic in the analytics for the next two weeks. The feed is a listing, and listings that are maintained get recommended. The ones that are not get skipped, quietly, by a model that never tells you why.

Frequently asked questions

Where do ChatGPT shopping recommendations come from?

Mostly from Google Merchant Center feeds and product pages, with the feed deciding inclusion and ranking. In one study, 83 percent of products in ChatGPT carousels matched Google top 40 Shopping results, and feed-sourced offers carried complete brand, image, and price details.

What are the six new Merchant Center attributes?

Question and answer, document link, related product, item group title, variant option, and popularity rank, launched at Google Marketing Live 2026. They are primarily intended for conversational experiences like AI Mode in Google Search.

Should I still optimize my product pages?

Yes. About 88 percent of ChatGPT product offers still come from product detail pages, so the page matters. Keep Product JSON-LD current and matching the feed exactly, because a price mismatch between page and feed costs the citation.

How many question and answer pairs can I add per product?

Up to 30 pairs, capped at 1,000 characters each and 10,000 characters total. Base them on the questions customers actually ask, and upload as TSV, because answers are full sentences and sentences contain commas.