A buyer types “best wool runners for rainy commutes under $150” into ChatGPT Shopping, Perplexity, Claude, or Gemini. The assistant returns a short list of named products, often with merchant links, sometimes with a citation back to the product page. The buyer clicks through. The merchant whose product appeared gets a session that did not exist on the open web before 2024.
That session is what people usually mean when they say “AI commerce” now — and it is a much narrower phenomenon than the phrase suggests. This guide is the definitional read on what AI commerce is, what it is not, and why the distinction matters for any catalog operator trying to decide what to optimize for.
The end-to-end mechanic that makes the narrow definition concrete:
Every surface has its own implementation of this pipeline, and the indexes differ in what they ingest, how they rank, and where the buyer ends up. But the shape is consistent enough that operators can treat the surfaces as a category rather than as five unrelated problems.
Two definitions, only one of them useful
“AI commerce” is in heavy use across vendor decks, conference talks, analyst reports, and ecommerce trade press. Inside that usage, two distinct meanings live side by side, and they get conflated constantly.
The broad definition. Anywhere AI shows up in the shopping funnel. Chatbots on a product detail page. Recommendation engines that personalize a homepage. Dynamic pricing. AI-generated product photography. Computer-vision-driven search (“show me sneakers like this”). Customer-support copilots. AI-written product descriptions. Loss-prevention models in fulfillment. The broad definition is roughly “the retail tech stack, now with LLMs sprinkled in.” It has been growing since 2017 and is large and important. It is also operationally useless as a category, because none of those surfaces share a distribution mechanic, an optimization discipline, or a data dependency.
The narrow definition. An AI assistant fulfills a shopping intent on behalf of a buyer, by retrieving and citing specific products from a queryable index. The buyer addresses the AI surface in natural language; the AI surface returns named products. The products did not get there by being paid for, by being highest- ranked on a static SERP, or by being on a homepage merchandising shelf — they got there because the assistant retrieved them from an index and ranked them above their substitutes.
The narrow definition is the operationally useful one. It is a discrete surface category, with shared distribution mechanics, a shared data dependency, and a shared optimization discipline. When the rest of this guide says “AI commerce,” it means the narrow definition.
The two columns share almost no operational discipline. A recommendation-engine vendor and a product-data team optimizing for ChatGPT Shopping have approximately nothing to say to each other. That is the case for using “AI commerce” only in the narrow sense and using more specific labels (AI-assisted on-site experience, agentic shopping, etc.) for the rest.
The narrow definition, broken into parts
Four properties together make a surface “AI commerce” in the narrow sense. A surface that has only two or three of them is something else — useful, sometimes, but not the same category.
1. Natural-language buyer intent
The buyer addresses the surface conversationally. Not “wool + runners + men’s + waterproof,” but “wool runners that hold up in rainy commutes.” The surface has to interpret intent, not match keywords. This is what rules out classic search engines (which take keyword strings) and homepage merchandising (which takes no buyer input at all).
2. Retrieval from a queryable index
The surface does not generate products from scratch. It retrieves real, named products from an index it has built or licensed. The index is the surface’s view of the merchant world. A surface that hallucinates products instead of retrieving them is not in this category — it is a generative experience masquerading as commerce, and it tends to break trust fast.
3. Citation or merchant link
The response names specific products and, in most cases, links to the merchant page or surfaces a “buy” action inside the assistant. The citation is what makes the surface a distribution channel rather than a brand-awareness channel. A surface that summarizes a category without naming products may be useful to the buyer but it does not move sessions to the merchant.
4. An open-internet ingest path
The merchant did not need a private API or a vendor relationship to get into the index. The merchant published their catalog to the open web with structured markup (or to a feed surface like Google Merchant Center that the AI surface ingests), and the AI surface picked it up. This is what makes AI commerce a discipline closer to search than to retail-media partnerships.
Surfaces that fail property 4 — for example, a closed marketplace that returns AI-curated recommendations only from sellers who pay to be there — are real, but they belong to retail media, not AI commerce in the open-internet sense this guide is using.
The surfaces, by name
Six surfaces fit the narrow definition cleanly as of 2026. Each has its own quirks, but the shared mechanic is identical.
ChatGPT Shopping — OpenAI’s in-product shopping experience. Ingests a product index built from open-web catalogs and merchant feeds. Returns cited product cards inside the conversation; clicks go to the merchant page.
Perplexity Shopping — Returns shopping cards inline with answers. Treats merchant pages with structured markup as the canonical product reference. Includes an in-surface “buy” path for some merchants and a click-through path for others.
Google AI Overviews — Generative summaries above the classic search results. Commercial-intent queries can surface product cards inside the overview. AI Overviews uses Google’s existing crawl + Merchant Center pipeline, so the catalog-side levers are familiar to anyone doing search SEO and Shopping feed work.
Claude — Anthropic’s assistant. Returns named products with citations when the query is commercial; the underlying retrieval draws on Anthropic’s web search integration. Less of a discrete “shopping” product than ChatGPT or Perplexity, but the citation pattern is the same.
Gemini — Google’s assistant. Overlaps with AI Overviews for many shopping queries but also surfaces products through standalone Gemini sessions; the retrieval taps Google’s product graph.
Microsoft Copilot Shopping — Built on Bing’s product index, fed by Microsoft Merchant Center. Surfaces products inline in Copilot conversations.
The list is roughly stable now; new entrants are likely as more assistants add shopping behavior, but the mechanic — index, retrieval, citation — is converging across the category.
The mechanic, in one paragraph
Every AI commerce surface, regardless of vendor, runs some version of: a buyer issues a natural-language query; the surface embeds the query; it retrieves candidate products from an index of catalog content the surface has previously crawled or ingested; it ranks the candidates against the query and against signals (popularity, freshness, structured-data completeness, merchant trust); it returns a short list of named, cited products in the assistant’s response. The buyer either clicks to the merchant page or completes a purchase in-surface.
The implementation differs surface by surface. The optimization discipline does not, much: the catalog-side work that lifts ranking on one surface tends to lift ranking on the others, because the indexes are all consuming variations on the same Schema.org-tagged open web.
Why this matters now
Three concrete shifts are what made AI commerce a distinct category between roughly 2023 and 2026.
LLM-powered shopping assistants reached real distribution. The launch of ChatGPT Search in October 2024, the public expansion of Perplexity Shopping in late 2024, and the broadening of Google AI Overviews to commercial queries through 2024–2025 put a meaningful share of buyer sessions through LLM-mediated surfaces. Before this period, AI in shopping was mostly an on-site experience layer; from this period on, AI is a discovery channel.
The retrieval target became Schema.org-tagged open web. Each
AI surface chose to ingest the same kinds of structured catalog
data merchants were already publishing for Google. That decision —
to lean on Schema.org Product markup
rather than build a surface-specific ingestion pipeline per vendor —
made the optimization discipline portable across surfaces. The
catalog operator who got schema right for one surface usually got
it right for all of them.
The merchant-side discipline went from “ranking on Google” to “being citable by an assistant.” Citability is a different optimization target than ranking. Ranking is one signal: search-result position. Citability has more shape — the assistant has to find the product, trust the data enough to include it, and have enough specificity to match the buyer’s natural-language constraint. The work of making a catalog citable is the work this site refers to as AI readiness.
What changes for a catalog operator
The narrow definition has implications. A few worth naming:
The audience is an assistant, not a human. The first reader of a product page in AI commerce is an LLM-powered retrieval system, not a buyer. The page can be beautifully designed and still be invisible if the structured data underneath is thin. Conversely, a page with weak design but strong structured data and dense descriptions can outperform far more polished competitors on AI surfaces.
Distribution is not paid. No AI commerce surface in the narrow sense currently has a paid placement mechanism comparable to Google Ads. Citability is earned via catalog quality and structured markup. (Some surfaces, e.g., Microsoft Copilot Shopping, sit on top of an ads infrastructure, but the AI Overview-style citation is generally an organic placement.) This is plausibly the most underestimated property of the category — operators used to buying their way to visibility do not have that lever here.
Optimization is portable across surfaces, but not identical. Schema completeness lifts citability on every surface. But each surface has its own quirks: Perplexity is unusually strict about schema validation; ChatGPT Shopping draws heavily on the OpenAI product index, which has its own ingestion latency; Google AI Overviews still inherits classic Google ranking signals. See How AI agents discover and rank products for the surface-by-surface treatment.
Measurement is harder. Sessions arriving from an AI surface often present as direct traffic, branded search, or generic referrers. The traffic is real, but classical analytics tools attribute it incorrectly. Treat agentic-traffic attribution as an open problem and use multiple methods (server-log filtering on known assistant user agents, dedicated landing pages for the clusters that get cited most, periodic manual sampling).
Common conflations to push back on
Three usages of “AI commerce” creep into general usage and create confusion when an operator tries to plan against them. None of them fit the narrow definition.
“We added a chatbot, so we have AI commerce.” An on-site chatbot is an experience-layer feature. It does not change a catalog’s citability on an off-site AI surface. Useful, but not the same surface.
“Our recommendation engine uses an LLM under the hood.” Personalization is an experience-layer feature. The LLM is on the merchant’s infrastructure, not the buyer’s. This is AI in retail tech; it is not AI commerce in the discovery sense.
“We generate product descriptions with an LLM.” Generation is content production. It can lift citability if the generated content is structurally better than what it replaced, and it can hurt citability if the generated content is generic and crowds out attribute-dense detail. But the act of generation is not itself AI commerce. The surface that consumes the resulting content is.
Where the narrow definition itself breaks down
Even the narrow definition has edge cases. Three worth naming.
Closed-marketplace AI. Amazon’s Rufus assistant and similar in-marketplace shopping agents fit properties 1, 2, and 3 of the narrow definition but fail property 4 (open-internet ingest). The surface is real and valuable, but the optimization discipline for in-marketplace AI is closer to marketplace SEO than to the open-internet category this guide is mapping.
Image- and voice-led shopping. A buyer who shows a screenshot
of an outfit to a multi-modal assistant and asks “where can I buy
something like this” is in the AI commerce surface in the narrow
sense, but the retrieval signal is the image rather than a text
query. The catalog-side optimization is the same (structured data,
images with good alt, GTIN coverage), but the surface’s reader
of the catalog is a multi-modal model.
Agentic shopping (autonomous purchase). Assistants that complete the purchase on the buyer’s behalf — book a trip, place a recurring order, restock a pantry — are an emerging variant. They share the index + retrieval + citation mechanic but add a trust dimension (the assistant has to be confident enough to act, not just recommend). When this category matures it is likely worth a separate label; for now it sits inside AI commerce.
What this guide is part of
The rest of this guides section breaks the discipline down into practical work. The most useful next reads are:
- How AI agents discover and rank products — the mechanic in more depth, with the surface-specific differences.
- The 6 dimensions of AI readiness — Lumio’s framework for measuring how citable a catalog is.
- Traditional SEO vs. AI search optimization — what carries over from search work and what is genuinely new.
- AI commerce glossary — the terms that come up repeatedly in this material.
