How does AI shopping work for ecommerce brands?
A shopper describes what they want to ChatGPT or Gemini. The assistant may return products, prices, and links. We study which brands appear, which products are returned, and where those links lead.
This article explains the recorded stages. It does not say what every shopper sees or whether a purchase occurred.
AI shopping adds another product-discovery surface.
A shopping question can begin in a chat interface as well as in search or a marketplace.
“I’m shopping for a nice weighted blanket for a hot sleeper. Show me some good options.”
The assistant may return a product, a price, an image, and a way to reach a merchant. That output is a discovery surface. It may shape which product and destination appear in the recorded answer.
Which AI shopping surfaces should a store measure first?
ChatGPT and Gemini are a reasonable first comparison because they combine broad usage with product-oriented interfaces. Other surfaces can follow.
These are broad usage estimates, not shopping counts. They show why ChatGPT and Gemini are useful first surfaces to compare, but they are not a denominator for a Caeliai visibility study or a measure of merchant demand.
This comparison indicates a large potential discovery surface. It is not a count of shoppers, completed purchases, or merchant revenue.
The interfaces differ. Some assistants return product cards; others return text mentions. We record those outputs separately.
Many assistants can recommend a product in text. ChatGPT and Gemini also return product-card interfaces in the observed product surfaces. Their coverage and behavior should be measured separately by market and date.
What does an AI shopping query look like?
The query is the shopper’s sentence. It is a full request, not just a list of keywords:
“I want a weighted blanket that won’t make me overheat, under $200, good reviews. What should I get?”
This request includes a product type, a constraint (won’t overheat), a budget (under $200), and a quality bar (good reviews).
The output may include a product card with an image, name, price, and rating, a short explanation, and a link or button.
Two things we record in AI shopping tests.
AI answers can vary between runs, and account context can affect a result. Both matter when we measure visibility.
The same query can give different answers. Run “best weighted blanket for a hot sleeper” five times and the list may change. A single check is not enough to estimate a rate.
Visibility is a rate, not a yes or no.
The same request may produce different results across accounts or sessions. Shopping features can use context, but one account should not be treated as a universal view. We record the context used in the test.
Account context can affect a result. Record it before comparing runs. This is a measurement rule, not a guarantee about any shopper.
PDP, or Product Detail Page: the page for one specific item in one specific size and color.
Not your homepage. Not a category page. Every product on your store has one.
A product destination in an AI result can point to a brand-owned PDP, a retailer or marketplace listing, or no usable product-page path. The destination is separate from the recommendation itself.
The three output types we record.
ChatGPT provides a clear example of how a shopping query can produce different outputs. Gemini can produce similar product-oriented results.
Mode 01 · product cards
Product cards with a destination.
Product cards that link to a store, sometimes several stores per product, or to a direct product page. The card supplies a destination; it does not establish that a purchase occurred.
Mode 02 · info panel
Product or brand information without a link.
A summary about a brand or product, with no card and no product-page link. It is an output without a recorded product destination.
Mode 03 · plain text
A brand mention without a destination.
A brand named in passing, with no card, image, or product-page link in the answer. No destination was recorded.
These are different output modes. The first includes a product destination; the others do not. They should not be collapsed into a single visibility measure.
Traffic after an AI referral is a separate question.
Traffic that arrives from an AI product surface may differ from organic search. Published industry estimates compare some AI-referred cohorts with organic traffic, but those figures are directional and do not establish a completed purchase for a given store.
The rails underneath: ACP, UCP, and AP2.
Buying inside the chat can run on commerce protocols. Merchants do not need to implement every protocol to measure the resulting product path; the protocols explain how an assistant can connect to a destination.
The relevant distinction is the connection between an assistant, product data, and a merchant destination. Shopify and Stripe participate in parts of that ecosystem. Availability and merchant coverage vary by platform, market, and date.
Being named is not the same as owning the path.
A brand can be named without receiving a brand-owned product-page path. Every recorded recommendation is classified into one of three outcomes.
Outcome · owns
Owns the path
The assistant recommends the brand and the returned product path lands on a brand-owned product page.
Outcome · leaks
Leaks the path
The assistant recommends the brand, but the returned path goes to a retailer, reseller, or marketplace.
Outcome · no path
No path
The answer mentions or shows the brand, but there is no usable product-page path in the recorded output.
Caeliai measured this across real ChatGPT and Gemini shopping conversations in 11 ecommerce categories, with roughly 380 brand observations in total. The result:
The two platforms showed different routing patterns in this study. ChatGPT often named a product without a usable product-page path. Gemini attached more buy buttons, while a larger share of returned destinations went to third-party sellers. The brands appeared; the recorded destinations varied.
How to check your own store.
A brand appearing in an AI answer is not a measure of shopper demand. A store needs a recorded observation before it can choose a next step.
Results are not deterministic and may vary with session context. A useful check uses a spread of shopping queries, several runs of each, and the three path classes: owns the path, leaks the path, or no path. The resulting rate is more informative than a single screenshot.
Where did the returned link lead?
We run defined shopping prompts for your category and report where the returned destinations land: your store, a reseller, or nowhere. One form, no meeting.
Common questions.
What is AI shopping?
AI shopping is when a shopper uses an AI assistant to explore or compare products. The shopper describes what they want to ChatGPT or Gemini, and the assistant may return products, prices, explanations, and links.
What is AEO (Answer Engine Optimization) for ecommerce?
AEO is a term for making a brand and its products easier for answer systems to find and describe. Caeliai measures what appears in defined tests; it does not promise a universal position or a brand-owned path.
Which AI assistants have a real shopping interface?
ChatGPT and Gemini can render product-card interfaces with images, prices, and links in the observed product surfaces. Other assistants may differ by market, account, interface, and date, so we record them separately.
What is a PDP in ecommerce?
A PDP, or Product Detail Page, is the page for one specific item in one specific size and color. It is not the homepage or a category page. A returned product link may lead to the brand, a retailer, a marketplace, or nowhere.
How often do AI shopping recommendations resolve to the brand’s own store?
In Caeliai’s GENESIS-03 field study of real ChatGPT and Gemini shopping conversations across 11 ecommerce categories, about 37% of recorded brand recommendations resolved to the brand’s own product page. Around 18% resolved to a retailer or marketplace, and about 45% had no usable product-page path. These are study results, not a measure of every shopper.
AI assistant usage from SimilarWeb (2026, global estimates). Conversion figures from Semrush and Adobe (2025–26). Buy-path data from Caeliai GENESIS-03 (PDF), a field study of real ChatGPT and Gemini shopping conversations across 11 ecommerce categories.