AI Shopping ChatGPT Recommends: Products, Retailers, and Sources (Data Study)

If you sell a physical product, a growing share of your customers are no longer typing a query into Google and scanning ten links. They are asking ChatGPT “what’s the best running watch under $400?” and buying whatever it names. So we measured exactly what AI shopping ChatGPT recommends — which products, from which retailers, and where the picks come from.

Across 3,312 product-intent prompts run through six AI engines in cloro’s monitoring corpus, ChatGPT returned a structured product card on 87% of them, and Google AI Mode on 91% — while Perplexity (1%) and Gemini (0%) returned almost none. The brands and retailers those cards surface are not random, and where the engines get their picks from is the most actionable finding of all: AI shopping answers are sourced overwhelmingly from Reddit, YouTube, and review sites like RTINGS — not from brand websites.

This is the commerce companion to our ChatGPT ads penetration study, built on the same corpus and the same parsing pipeline. Where that post measured paid placements, this one measures the organic recommendations — the products ChatGPT names when nobody paid it to. It’s also one leg of a vertical-study family: we ran the same measurement for hotels and travel and restaurants, and the LLM citations study shows the sourcing pattern across all six verticals at once.

Which engines actually show product cards

Not every AI engine is a shopping surface. Two are, one is halfway, and three are effectively prose-only. Measured as the share of shopping-intent prompts on which each engine returned at least one product card:

Engine	Product-card rate	Avg. products / answer	Avg. sources cited
Google AI Mode	91%	4.7	17.1
ChatGPT	87%	4.0	13.1
Microsoft Copilot	33%	1.8	5.2
Perplexity	1%	0.0	2.3
Gemini	0%	0.0	6.8
Google AI Overview	0%	0.0	6.4

The split is clean. ChatGPT and Google AI Mode are genuine shopping surfaces — they return four or five specific products, with buy links, on the overwhelming majority of product-research prompts. Copilot does it a third of the time. Perplexity and Gemini answer the same questions in prose, naming products in the text but not rendering the structured cards that carry price and merchant. If your AI-shopping strategy assumes “the AI engines,” it is really a strategy about two of them.

Which retailers get named

Here is the first counter-intuitive result. In AI shopping answers, the retailers named are dominated by big-box and category stores, and Amazon is only fourth.

#	Merchant	Share of responses
1	Best Buy	12%
2	Walmart	11%
3	Target	5%
4	Amazon	4%
5	B&H Photo	2%
6	Newegg	1%
7	Home Depot	1%
8	Williams Sonoma	1%
9	Sony	1%
10	Apple	1%

Why Amazon lands only fourth

Best Buy and Walmart together are named in nearly a quarter of all responses. Amazon takes roughly 40% of US e-commerce, yet it appears in just 4%. The likely reason ties directly to the sourcing data below. AI engines assemble shopping answers from review content and structured retailer catalogs. Big-box retailers with clean, well-marked-up product pages and deep third-party review coverage surface more readily than a marketplace of third-party sellers.

606 merchants, no clear winner

It is also a fragmented field. 606 distinct merchants were named across the study (concentration HHI 0.046). No retailer owns AI shopping. If you are a mid-size retailer with good product data and review presence, the door is not closed.

Which products ChatGPT recommends

If retailers are moderately concentrated, the products AI shopping ChatGPT recommends are barely concentrated at all. The engines named 3,481 distinct products across 3,300 prompts — a nearly flat distribution (HHI 0.0005). Even the single most-named product appeared in only about 2% of responses.

#	Product	Category
1	Garmin Forerunner 265	Wearables
2	13-inch MacBook Air M5	Laptops
3	Apple Watch Series 11	Wearables
4	Fitbit Charge 6	Wearables
5	Helix Midnight Luxe	Mattresses
6	Amazon Kindle Paperwhite	E-readers
7	Ninja Nutri-Blender Pro	Kitchen
8	Sony WH-1000XM6	Headphones
9	Dyson V15 Detect	Home
10	GoPro HERO13 Black	Cameras

There is no “rank #1” for a product

The takeaway for brands is blunt. There is no “rank #1 in ChatGPT” for a product the way there is for a keyword. The engine assembles a fresh shortlist per query from whatever the current review consensus is. Winning is less about a single position. It is about being consistently present in the sources the engine reads — which is the next section.

Where the recommendations come from

This is the finding that should reshape how brands think about AI-shopping visibility. We tallied every domain the engines cited alongside their product answers. The top sources are not retailers and are not brands — they are video, community, and independent review sites.

#	Cited domain	Share of responses	Type
1	youtube.com	19%	Video / reviews
2	reddit.com	19%	Community
3	rtings.com	16%	Independent lab reviews
4	google.com	12%	Search / Shopping
5	forbes.com	9%	Editorial reviews
6	pcmag.com	8%	Editorial reviews
7	cnet.com	8%	Editorial reviews
8	tomsguide.com	6%	Editorial reviews
9	techradar.com	6%	Editorial reviews
10	amazon.com	5%	Retail / reviews

Reddit, YouTube, and RTINGS lead

The pattern is unmistakable. Reddit and YouTube — user-generated opinion — sit at the top, tied at 19%. RTINGS publishes standardized lab measurements. It ranks third at 16% and punches far above its traffic weight, because its data is exactly what a model wants to cite. Then come the editorial review desks — PCMag, CNET, Tom’s Guide, TechRadar, Wired, and the audio and cooking specialists (SoundGuys, Serious Eats) deeper in the tail.

Brand sites barely register

Brand-owned domains are conspicuously absent from the top of this list. An AI shopping engine decides what to recommend by reading what independent reviewers and communities say about a product, then naming a retailer that stocks it. For a brand, the lever is not your product page. It is your presence and reputation across Reddit threads, YouTube reviews, and the RTINGS/PCMag/CNET review circuit. This is the concrete, measurable version of the advice that generative-engine optimization writers have given in the abstract. It maps closely to what we found in our study of how AI engines cite sources.

Voice commerce runs on the same engines

“Voice commerce” used to mean shouting a reorder at a smart speaker. In 2026 it means something more consequential: the assistants people speak to are increasingly powered by the same large models measured in this study. When a next-generation Alexa, a Gemini-backed Assistant, or a ChatGPT voice session answers “order me a good pair of running headphones,” it is drawing on the same recommendation layer — the same Reddit-and-RTINGS-sourced shortlist — that produces the product cards above.

That collapses a distinction brands used to make. Optimizing for “voice search” and optimizing for “AI shopping” are now the same project, because the voice surface and the text surface share a brain. The practical implication: the visibility work that gets your product named in a ChatGPT product card is the same work that gets it spoken aloud by an assistant. We cover the search side of that convergence in our guide to what AI SEO means in 2026.

The AI shopping assistant landscape

Where each surface stands as a place to buy, based on the data above and each platform’s public direction:

• ChatGPT — the most developed shopping surface. Product cards on 87% of prompts, and OpenAI has been rolling out in-chat “Instant Checkout” so the transaction can complete without leaving the conversation. This is the surface most worth monitoring for a consumer brand.
• Google AI Mode — the highest card rate (91%) and the deepest sourcing (17 sources per answer on average), tightly wired into Google’s existing Shopping Graph. If Google’s product data already lists you well, AI Mode reflects it.
• Microsoft Copilot — a partial shopping surface (33%), leaning on Bing’s merchant data.
• Perplexity and Gemini — answer product questions in prose and name products in the text, but rarely render structured, buyable cards today. Worth tracking for mentions, not yet for card placement.

Instant Checkout: when a recommendation becomes a purchase

Getting named in a product card is no longer the end of the funnel. In September 2025, OpenAI and Stripe launched Instant Checkout in ChatGPT. The shopper can buy a recommended item without leaving the chat. Stripe powers the payment. The two companies co-developed the underlying Agentic Commerce Protocol (ACP) to standardize how agents and merchants transact (Stripe).

What the Agentic Commerce Protocol changes

ACP is an open standard, released under the Apache 2.0 license, and ChatGPT was the first platform to implement it (agenticcommerce.dev). The merchant stays in control. They accept or decline each transaction, and they keep the customer relationship and fulfillment. That separation matters for everything above. The recommendation layer and the checkout layer are distinct. Being recommended is still won upstream, in the sources the engine reads — not at the payment step.

Which merchants are live

Instant Checkout launched US-only, starting with US-based Etsy sellers. Over a million Shopify merchants — including Glossier, Vuori, Spanx, and SKIMS — were slated to follow (Stripe). The pattern echoes our merchant leaderboard. The buy step is consolidating around a few large catalogs and platforms. Meanwhile the products AI shopping ChatGPT recommends stay fragmented across thousands of SKUs.

There is an attribution wrinkle here that brands should plan for. When a shopper buys inside the chat, the sale never touches your site analytics. You may never see the click that a review earned you. The recommendation happened, the purchase happened, and both were invisible to your funnel. That is one more reason to measure the recommendation layer directly rather than infer it from web traffic. What AI shopping ChatGPT recommends is now the top of a funnel you do not fully own.

How to get your products recommended by AI shopping

The sourcing data points to a clear playbook. AI shopping ChatGPT recommends what the review consensus already endorses. So the work is to earn that consensus, then make it machine-readable.

Get reviewed where the engines read

Reddit, YouTube, and RTINGS lead the citations. Editorial desks like PCMag, CNET, and Tom’s Guide follow close behind. Pitch those reviewers directly. Seed genuine discussion in the relevant subreddits. A single well-ranked YouTube review can surface your product across many prompts at once.

Keep clean, structured product data

Big-box retailers win partly on data hygiene. Well-marked-up product pages, accurate specs, and clear pricing are easier for an engine to parse into a card. Ship valid product schema. Keep your specs consistent across every retailer that carries you. Inconsistent data is a reason to get skipped.

Track what actually gets named

You cannot optimize what you cannot see. Monitor which prompts name your product. Watch which competitors show up beside it. Note which sources the engine cited to make the pick. That last signal tells you exactly which review to chase next.

How to monitor whether AI recommends your products

The reason this is measurable at all is that these product cards are structured data, not screenshots. cloro parses them directly. For any prompt you care about — “best [category] under $X”, “[your brand] vs [competitor]”, a category term — you can pull:

• the shoppingCards and inlineProducts arrays: which products were named, at which merchants, at what price, with which buy links;
• the entities the answer named, so you can track your brand and your competitors’ brands across responses;
• the cited sources, so you can see which Reddit thread or review page drove a recommendation — and therefore where to invest;
• all of it by country, since AI shopping answers localize, and with alerting when your presence changes.

That is the AI visibility tracking workflow applied to commerce, and it runs on the same ChatGPT API surface the rest of the platform uses. The AI brand visibility tracker walks through the mention-and-citation side end to end.

Methodology

The measurement is drawn from cloro’s production monitoring corpus: 3,312 successful responses (552 per engine across ChatGPT, Google AI Mode, Google AI Overview, Copilot, Gemini, and Perplexity) to a stratified set of product-intent prompts spanning electronics, wearables, home, kitchen, audio, and outdoor categories. “Product-card rate” is the share of an engine’s responses that contained at least one parsed product card (shoppingCards[] or inlineProducts[]). Merchant, product, and citation leaderboards aggregate the parsed entities and cited domains across all responses; concentration is reported as a Herfindahl-Hirschman Index (HHI), where lower means more fragmented.

Prompt-mix caveat: this corpus is weighted toward US-market, English-language, commercial-research prompts, so the absolute card rates likely overstate what a uniform sample of all shopping queries would show. The relative findings are robust to that weighting: the engine split (ChatGPT/AI Mode as card surfaces, Perplexity/Gemini as prose), the retailer ordering (big-box over marketplace), and above all the sourcing pattern (UGC and review sites over brand sites) hold regardless of prompt mix. As with all of our studies, this is one independent measurement from one corpus — treat the point estimates as cloro-corpus signals and the directional findings as the story.

We re-run this study quarterly, because the surface is moving fast — Instant Checkout, ChatGPT’s ad ramp (measured in the companion ads study), and Google’s Shopping Graph integration are all actively reshaping it. If you want the version scoped to your own catalog and competitors rather than our corpus, that is exactly what cloro is built to produce.