Ask ChatGPT to recommend a product and it shows a carousel with eight options. Where do they come from?
In a March 2026 study, Tom Wells examined the origin. Out of more than 43,000 products, 83% matched Google's top 40 organic Shopping results. For Bing, only 11% matched, and almost all of those were also found on Google.
The products AI shoppers see don't come from the open web, your product detail pages (PDPs), or your reviews. They're pulled from a single file most brands haven't checked since setting up paid Shopping: your Google Merchant Center feed.
ChatGPT builds its product carousel using shopping query fan-outs — separate from the search queries that generate the answer text.
Wells found one of these queries often pulls a single page of Google Shopping results to populate an eight-product carousel, and 60% of strong matches come from the top 10 Shopping results. The order of products in the carousel matches their ranking in Google Shopping.
Profound reviewed more than 1 million ChatGPT shopping offers in June:
| Metric | Figure |
|---|---|
| Product citations pulled directly from merchant feeds that appeared as the top product offer | about 99.9% |
| Feed-sourced share of all ChatGPT shopping retrievals (six weeks) | 4.3% → about 20% |
The reason is completeness.
| Field | Feed-sourced offers | Page-scraped offers |
|---|---|---|
| Brand, product image and merchant details populated | 100% | 0% |
| Carried ChatGPT's "best price" tag | 100% | 21% |
The feed provides the LLM with clean, structured fields rather than forcing it to infer that information from the page.
Malte Landwehr of Peec AI, whose data supported the Wells study, added a new shop to Merchant Center and saw it appear in Google Shopping the next day, then in ChatGPT. Connect your feed and your products can show up; skip it and you might go invisible.
Most advice about optimizing for AI leaves out an important detail.
Per Profound's analysis, about 88% of ChatGPT product offers still come from web product detail pages rather than feeds. Even for merchants already using feeds, around 76% of offers still came from the page.
The feed doesn't replace the PDP. The feed helps with ranking, but most offers still come from the PDP.
Your catalog serves two purposes:
But coverage alone doesn't guarantee selection. When Lily Ray studied brands that ranked themselves No. 1 in their own listicles, about 69% were cited but not recommended — the top spot often went to a larger competitor on the same list.
An analysis of 11,400 AI shopping answers across ChatGPT, Perplexity and Gemini in June 2026 found category structure didn't affect whether AI recommended a brand on any platform.
Optimize one surface using the other's playbook, and you'll underperform on both.
Agencies are reorganizing already. Andre de Gaye, sales director at Charle, says his team is "moving away from treating SEO and feed management as separate silos."
An AI shopping agent doesn't browse your site the way a person does. It reads structured attributes and product details.
OpenAI explains product results are organic and unsponsored, ranked by relevance using signals such as "availability, price, quality, and whether a merchant is the primary seller." These signals come from catalogs and feeds, not on-page copy.
The catalogs include Target, Sephora, Nordstrom, Best Buy, The Home Depot, and millions of Shopify merchants.
The non-negotiable core is the data Google Shopping has always rewarded: a valid GTIN, an accurate title, price and availability that match your live site, a clean image, brand, and the correct product category. Get those wrong, and nothing downstream matters.
In July Google began supporting the product category property in merchant listing structured data, covering both Google's taxonomy and your product types, plus sale duration fields.
Google added conversational attributes to the Merchant Center product data specification — optional fields meant for AI features like AI Mode and Gemini:
| Attribute | Role |
|---|---|
| Question and answer | Structured Q&A pairs — lets you answer "Does this work for air travel?" before the shopper asks |
| Related product | Defines relationships using often_bought_with, required_part, accessory, substitute — helps an agent suggest a complete solution instead of a single product |
| Document link | Supporting PDFs such as manuals, spec sheets, or sizing guides |
| Item group title and variant option | Connects variants to a product family and matches "show me this in black, medium" |
| Popularity rank | A score showing performance against the rest of your catalog, answering "what's your best-selling running shoe?" |
None of these affect whether your product is approved — they're optional enhancements letting you provide information that previously lived only on the page or in a PDF a model couldn't always read.
For most brands, the main challenge isn't strategy. It's keeping everything updated and well maintained.
Common Merchant Center problems: missing or incorrect GTINs, image issues, and shipping or price details that don't match your website. If a product is disapproved, it won't just rank lower in AI shopping results — it won't show up at all.
The most harmful gaps are often the ones that don't trigger any errors.
Per Adobe's Q2 AI Traffic report, product detail pages scored only 63.5 for AI citation readability even at top retailers — far below their homepages and buying guides in the low 80s.
The pages containing the product data AI needs are often the hardest for machines to read. Each missing detail acts as a silent filter: your product might be approved and indexed and still be invisible in a natural language search.
| Metric | Figure |
|---|---|
| AI-source traffic to U.S. retail sites (Q1, YoY) | +393% |
| By December | up more than 1,150% |
| AI-referred traffic conversion vs non-AI (by March) | 42% better (about half that a year earlier) |
| Global online holiday sales linked to AI and agents (Salesforce) | about 20%, roughly $262 billion |
| AI-referred conversion vs social traffic | about 8x |
AI platforms are now bringing shoppers who are ready to buy to retail sites. The key question is: are your products being recommended?
1. Eligibility — check Merchant Center for disapprovals and diagnostics; fix GTIN errors and price mismatches first. Any disapproved product can't be shown by AI agents.
2. Coverage and specificity — review the titles and descriptions of your top 50 revenue products as if you were an AI agent. If the description just repeats the title, it doesn't add value.
3. Conversational attributes — begin with question-and-answer sections on your best sellers, then link related products and include popularity rank. Focus on SKUs that already generate revenue; you don't need to update all 40,000 items.
4. Freshness — keep prices and availability continuously in sync. AI shopping surfaces refresh constantly, so a feed that updates once a day is already behind.
Google's Shopping Graph now holds 60 billion product listings, up from 50 billion earlier in the year.
Google worked with Shopify, Etsy, Wayfair, Target and Walmart to create the Universal Commerce Protocol, which uses your existing Merchant Center feed for agentic checkout. To opt in, merchants add a new native_commerce attribute and keep product, offer and review schema in sync. Miss it, and products are ineligible for AI-powered checkout.
As Jason Tabeling noted in July, Merchant Center is "no longer simply for Shopping ads. It's becoming the primary source of product data for AI discovery."
Meanwhile presence on AI surfaces is decoupling from classic rankings. A July SE Ranking study found only about 2.32% of advertisers appearing in Google's AI Mode also ranked organically for the same search, and around 85% didn't appear in organic results at all.
"The last decade rewarded marketing arbitrage. Agentic commerce rewards product truth." — Kevin Indig
How people complete purchases is still evolving — Walmart tested checkout inside ChatGPT but it converted at about one-third the rate of its own site, and OpenAI has stopped offering hosted checkout. Still, shoppers are already turning to AI for discovery, and that shift relies on the product feed.
Reclassify the feed from paid-Shopping config to discovery asset. The article in one line: the file most brands haven't touched since setup is now the primary source for AI discovery.
Manage feed and PDP as separate jobs. 88% still coming from PDPs matters — fixing only one is not the answer.
Eliminate "contact us for pricing." The most immediately actionable finding from 6.77 million sessions: give AI no number to compare and you leave the comparison.
Start with the top 50 revenue products. Attempting 40,000 SKUs achieves nothing. The author's order: Q&A → related products → popularity rank.
Raise feed refresh above once daily — "a feed that updates once a day is already behind" is a concrete bar.
Measure cited-but-not-recommended separately. Lily Ray's 69% is the same structure as Ghost Citations.
Decide on agentic checkout readiness now. A missing native_commerce attribute means ineligibility for AI checkout — background in When the AI Assistant Fills the Cart.
Read it alongside ad-side automation — the same feed generates ad copy: AI Max Spotted in Standard Shopping Campaigns and Google Tests AI-Generated Descriptions in Shopping Ads.
Tom Wells' March 2026 study found 83% of more than 43,000 products matched Google's top 40 organic Shopping results. Carousel order matches Google Shopping ranking, and 60% of strong matches come from the top 10 Shopping results.
In Profound's analysis, feed-sourced offers populated brand, image and merchant details 100% of the time and carried the 'best price' tag 100% of the time, versus 0% and 21% for page-scraped offers. About 99.9% of feed-sourced citations appeared as the top offer.
No. About 88% of ChatGPT product offers still come from PDPs, and even for merchants using feeds around 76% come from the page. The feed governs inclusion and ranking; the PDP does the persuading and gathers reviews and coverage.
Question and answer pairs, related product relationships (often_bought_with, required_part, accessory, substitute), document links for manuals and sizing guides, item group title and variant option, and popularity rank. All optional and none affect product approval.
Eligibility (disapprovals, GTIN errors, price mismatches), specificity of titles and descriptions on the top 50 revenue products, conversational attributes starting with best sellers, and refresh frequency — a feed updating once a day is already behind.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Naver’s shopping AI agent lifted transaction value 82% since June. Here’s how to prepare for agentic commerce: structured product attributes, agent-readable product pages, clear reasons to buy and trust signals.
Shipping delays, order cutoffs and fulfillment updates all need different timing and audiences. Here’s a simple framework for e-commerce customer messaging — manual, recurring and automated sends — plus SMS fallback.
Korean D2C brand Wiselly now sells every product at cost for a ₩2,990 monthly membership. How cutting distribution, marketing and packaging costs built its zero-margin commerce model.