Key takeaways
- ChatGPT's shopping carousel is not an independent product database. Research from Peec AI, confirmed by decoding hidden parameters in ChatGPT's own page source, found that 100% of products shown in the carousel could be traced back to the top 40 organic results in Google Shopping for that query.
- ChatGPT applies its own re-ranking on top of what it pulls from Google: product attributes (comfort, durability, price range), star ratings (anything under three stars is effectively filtered out), and terms injected during "shopping fanout" queries all shift what gets shown and in what order.
- Paid advertising inside ChatGPT is real now and growing fast. Promptwatch's tracking found ChatGPT served zero ads before May 27, 2026, then climbed past 30% of web-search responses by July 1, and sat around a 29.6% 30-day average by mid-August. Ads sit below the answer as a separate, labeled unit, they don't replace the organic carousel.
- Instant Checkout, the feature that let people buy directly inside a ChatGPT conversation, was scaled back by OpenAI in March 2026 after weak adoption and messy inventory sync with retailers. The model in 2026 is "discover in chat, buy on the retailer's site," not "buy in chat."
- Structured, accurate product feeds (Google Merchant Center-compatible) remain the single highest-leverage lever a retailer controls. Brands with 9+ structured product facts saw roughly 78% average AI coverage versus 9% for brands with two or fewer, according to Erlin's 2026 benchmark of 500+ brands.
Why this question matters more than it did a year ago
I'll be honest: a year ago, "how does ChatGPT pick products" was a fringe question mostly asked by curious marketers. It isn't fringe anymore. ChatGPT now sits between over a billion weekly users and the open web, and a meaningful share of those users are asking some version of "what should I buy." OpenAI itself says shopping queries have grown enormously, and Adobe for Business found LLM-referred traffic converts 42% higher than non-AI traffic, for the simple reason that someone chatting with an AI about a specific purchase is further along than someone typing a vague query into Google.
So the mechanics matter. If you run an e-commerce brand, a DTC store, or you're the person responsible for product visibility, you need to know exactly which levers move the needle and which ones are theater.
Where ChatGPT actually gets its product data
This is the part that surprised a lot of people in the industry, myself included. The working assumption for most of 2025 was that ChatGPT had built its own product index, maybe licensed data from retailers, maybe crawled product pages directly. The reality, confirmed by researchers at Peec AI, is far less exotic.
When a user asks something that triggers a product carousel, say "best espresso machine under $300," ChatGPT isn't searching a database it built. Researchers found a hidden, encoded field in ChatGPT's page source that, once decoded, contained unmistakable Google Shopping parameters: product IDs, catalogue IDs, offer document IDs, and the exact query used to retrieve the product. They reconstructed a working Google Shopping URL from those parameters and it matched the carousel product exactly.
Across more than a million shopping fanout queries, compared against Google Shopping results across many industries and markets, the finding held: 100% of the products seen in ChatGPT Shopping could be explained by the top 40 organic products in Google Shopping for the corresponding query. Bing Shopping, by comparison, explained only 11% of what showed up, and almost none of that exclusively.
The practical consequence is simple and a little freeing. If you already maintain a Google Merchant Center feed, you're most of the way to being eligible for ChatGPT's carousel. If your products never crack Google Shopping's top 40 for a given query, they don't exist as far as ChatGPT's carousel is concerned, no matter how good your website is.
The re-ranking layer: what happens after Google hands over the candidates
ChatGPT doesn't just copy Google Shopping's ranking and call it a day. It re-ranks what it pulls, and that re-ranking is where the actual competition happens in 2026.
Product attributes carry weight
When ChatGPT builds a comparison table, it scores products on attributes like comfort, durability, versatility, or price range. These aren't arbitrary; they reflect what the model considers the defining dimensions of that product category. If your product copy doesn't use those words, backed by evidence, you're starting from behind in the re-rank, even if your Google Shopping listing is technically eligible.
Star ratings act as a gate, not just a signal
Peec AI's research found that products scoring below three stars on the attributes ChatGPT evaluates are effectively invisible in the carousel. This isn't a minor ranking nudge, it's closer to a pass/fail threshold. Getting genuine reviews onto your product pages isn't optional polish anymore.
Fanout queries add words you never typed
This is the most actionable and most overlooked finding. A user asking for a "cozy, heavyweight, oversized hoodie" triggers a fanout query that ChatGPT expands on its own, adding terms like "premium cotton," "structured," and "men." If those exact words don't appear on your product page or in your feed, your product is less likely to be retrieved in the first place, regardless of how good a match it actually is. Across e-commerce categories, the terms ChatGPT most commonly adds to fanout queries include "best," "review," "2026," "top," "comparison," and "vs," which tells you plainly what kind of content needs to exist around your products: comparison pages, not just product pages.
Promptwatch's own tracking of query fanout behavior backs up how mechanical this has become: average fanout query length dropped from roughly 117 characters in late 2025 to around 53 characters by April 2026, less than half, meaning ChatGPT increasingly searches like someone typing short keyword strings rather than full sentences. That shift rewards pages that match tight, specific phrasing over pages written for a human reading a full question.
How often does the shopping carousel even show up?
Shopping triggers are not constant. One tracking study of 260 million ChatGPT prompts found shopping features surface on roughly 9% of queries overall, concentrated heavily in fashion (62% trigger rate), other physical goods (56%), and groceries (over 30%). Promptwatch's own data on this shows the rate moves in visible steps rather than staying flat, it roughly doubled overnight in late May 2026 and then fell back weeks later, which is a strong hint that OpenAI is actively tuning when and how often shopping cards appear rather than letting it run purely organically.
The ad layer: the part that's changing fastest
For most of 2025 and early 2026, the clean story was "ChatGPT shopping is 100% free, there's no auction, nothing to buy." That story is getting dated fast. Promptwatch's 90-day tracking (May 20 to August 17, 2026) found ChatGPT served zero ads before May 27, 2026. That day, ads appeared in 1.69% of responses. By July 1, the share had passed 30%, and the most recent 7-day average in that window was 32.4%, up 12.5% week over week.
Of those ad-bearing responses, organic (non-branded) prompts carried roughly 73% of the ads, brand-specific prompts about 18%, and competitor-comparison prompts around 9%. That matters because it tells you ads aren't just chasing people who already searched a brand name, they're showing up heavily on generic "best X for Y" questions, exactly the kind of prompt that also triggers the organic carousel.
OpenAI is explicit that ads and organic carousel results are structurally separate: ads sit below the answer, always labeled, and the company states advertising "does not influence the answers ChatGPT provides." So the carousel itself is still not for sale. But a parallel paid channel now exists alongside it, with OpenAI's new Product Feed Ads letting retailers upload a full catalog and auto-generate sponsored units, using essentially the same feed spec as Google Merchant Center. Recommended starting bids sit around $3 to $5 CPC, with some reports of CPMs near $25.
| Channel | Who decides placement | Cost | What controls it |
|---|---|---|---|
| Organic shopping carousel | ChatGPT's re-ranking of Google Shopping's top 40 organic results | Free | Feed completeness, attributes, ratings, fanout term coverage |
| ChatGPT Ads (standard) | Relevance-weighted second-price auction | CPC/CPM/oCPC bidding | Ad campaign setup, bid, landing page |
| ChatGPT Product Feed Ads | Catalog-driven auto-generated sponsored units | Same auction, catalog-based | Google Merchant Center-compatible product feed |
The checkout story: buy-in-chat didn't stick (yet)
It's worth clearing up a misconception that's still floating around: ChatGPT does not let most people complete a purchase inside the conversation anymore, at least not the way it looked in late 2025. OpenAI launched Instant Checkout in October 2025 with Walmart as a founding partner and Stripe handling payments, under something called the Agentic Commerce Protocol. Shopify had an exclusive integration letting its merchants process orders natively.
By March 2026, OpenAI scaled the feature back. The Information reported only around a dozen Shopify merchants were actually live with native checkout at the time of the pullback. Forrester's analysis of why is blunt: completing a purchase inside an answer engine was users' least-adopted use case, well behind asking questions and researching products, and inventory management was, in Forrester's words,