How AI Shopping Agents Choose Products (and How Brands Get Picked)

Ask ChatGPT for the best quiet vacuum for a small apartment, or tell Google's AI Mode to find a sofa that ships before the holidays, and you get something that did not exist a year ago: a short, confident list of specific products, often with prices, ratings, and a button to buy. The agent did not scroll your homepage or admire your photography. It pulled structured data, ranked it against narrow sub-queries, checked it against what other sources say, and named a handful of winners. As of 2026 these agents also increasingly transact, so being on that list is no longer a vanity metric. Understanding how AI shopping agents work, and what actually produces AI product recommendations, is now the difference between being in the answer and being invisible. This piece walks through the mechanism, the signals that decide which products get picked, and the concrete moves that raise a brand's odds.

How AI shopping agents actually work The most useful thing to internalise is that an AI shopping agent does not experience your site as a shopper does. It retrieves structured records, ranks them, and corroborates them. The clearest evidence for where those records come from arrived in a March 2026 analysis of more than 43,000 ChatGPT carousel products, reported by [Search Engine Land](https://searchengineland.com/new-finding-chatgpt-sources-83-of-its-carousel-products-from-google-shopping-via-shopping-query-fan-outs-470723), which found that 83% of ChatGPT's product recommendations match Google Shopping's top 40 organic listings, with around 60% drawn from the top 10. ChatGPT reaches them through what the study calls shopping query fan-outs: it decomposes a request into several short, intent-specific sub-queries, averaging about seven words each, and pulls candidate products from Google Shopping's organic index. The practical takeaway is blunt. For ChatGPT, getting picked starts with your Google Shopping feed, not your homepage copy. The same retrieve-rank-corroborate pattern holds across engines, even where the underlying index differs.

How AI agents choose products: the five signals Once candidates are retrieved, a handful of signals decide which ones survive into the answer. The first is semantic query relevance. The agent is matching intent, so a product described only as a vacuum loses to one whose copy names its noise level and dimensions when the query is quiet vacuum for a small apartment. The second is structured product data: Product, Offer, and AggregateRating schema on the page, plus complete feed attributes such as GTIN, MPN, brand, color, size, material, and the correct Google product taxonomy. The third is price and availability accuracy, where current pricing and an in-stock status rank higher and any feed-versus-landing-page mismatch erodes trust. The fourth, and the one brands most underrate, is third-party authority, the roundups, buyer guides, and review platforms an agent uses to verify a claim. The fifth is contextual weighting: budget-led queries are scored more on price, quality-led queries more on ratings. These factors echo the broader pattern in [how AI models choose which brands to recommend](/blog/how-ai-models-choose-which-brands-to-recommend), now applied at the level of individual SKUs.

Structured data and feeds: the new storefront If retrieval runs on structured records, your feed is your storefront in a way your designed pages no longer are. OpenAI's own commerce specification makes the shift explicit. Merchants submit product feeds in CSV, TSV, XML, or JSON to a secure OpenAI endpoint rather than waiting to be crawled, and the spec supports refreshes as often as every 15 minutes so price and stock stay close to real time, as documented in the [OpenAI commerce file-upload spec](https://developers.openai.com/commerce/specs/file-upload/products). Google's side of the equation is larger still: its Shopping Graph now holds more than 50 billion product listings, and Merchant Center has added conversational-era attributes such as answers to common questions and lists of compatible accessories, described in [Google's agentic shopping announcement](https://blog.google/products-and-platforms/products/products/shopping/agentic-checkout-holiday-ai-shopping/). A thin or stale feed does not just rank lower, it gets skipped or misrepresented. This is the same discipline that underpins [AI visibility for e-commerce brands](/blog/ai-visibility-for-ecommerce) and the case for [schema markup as an AI visibility signal](/blog/schema-markup-ai-visibility): give the machine clean, complete, current attributes or accept that it will fill the gaps with someone else's product.

Why third-party signals outweigh your own site Here is the part that frustrates marketing teams. You can ship a flawless feed and still lose the slot, because agents corroborate before they recommend, and most of that corroboration happens off your domain. Vendor analysis from [Alhena](https://alhena.ai/blog/chatgpt-shopping-product-recommendations/) reports that roughly 91% of AI citations trace to off-site sources, the buyer guides, forum threads, and review platforms an engine trusts more than your marketing copy. That figure comes from a vendor blog rather than primary research, so treat it as directional, but the direction is well established elsewhere. Brands with active profiles on the major review platforms are markedly more likely to be cited, a pattern we examined in [how Trustpilot and G2 shape AI recommendations](/blog/review-sites-ai-recommendations), and community discussion carries similar weight, which is [why AI models cite Reddit](/blog/why-ai-models-cite-reddit) more often than the brand's own pages. The lesson for product teams is that off-site presence is not a nice-to-have layered on top of a good feed. It is the corroboration step the agent runs before it will name you at all.

ChatGPT, Google, and Perplexity compared The three engines most buyers touch retrieve, structure, and transact differently, which is why a product can surface in one and vanish in another. The table below summarises where each pulls product data, what structured input it rewards, whether it can complete a purchase, and how fresh its data tends to be. None of these are static, and the checkout rows in particular are moving quickly through 2026, so treat the comparison as a snapshot rather than a fixed spec. | Engine | Product data source | Structured-data needs | In-answer checkout | Data freshness | |---|---|---|---|---| | ChatGPT | Google Shopping organic index plus direct merchant feeds | Product feed (GTIN, taxonomy) and on-page schema | Yes, Instant Checkout with participating merchants | Feeds refreshable every ~15 min | | Google AI Mode / Gemini | Shopping Graph (50B+ listings) | Merchant Center feed plus conversational attributes | Yes, agentic checkout rolling out to select merchants | Near-real-time via Merchant Center | | Perplexity | Live web retrieval plus merchant integrations | Crawlable pages, schema, strong third-party signals | Limited, expanding via partner integrations | Live at query time |

From recommended to bought: the agentic commerce layer The newer development is that the engines no longer stop at a recommendation. On 29 September 2025, OpenAI launched Instant Checkout in ChatGPT with Etsy and Shopify merchants, and alongside it OpenAI and Stripe open-sourced the [Agentic Commerce Protocol](https://stripe.com/newsroom/news/stripe-openai-instant-checkout), which uses a Shared Payment Token scoped to one merchant and one cart total, single use, so the agent never handles raw card credentials. Google is building the same capability into Search and AI Mode, with agentic checkout reaching merchants such as Wayfair, Chewy, and select Shopify stores, and is publishing its own Universal Commerce Protocol and the AP2 standard for agent payments per its [agentic commerce tools announcement](https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/). The card networks are moving too, with Mastercard Agent Pay and Visa Intelligent Commerce defining how agents authenticate and pay. The honest framing is that being picked and being bought are now two distinct stages. Brands optimise the first through feeds, schema, and reputation; protocols govern the second. We unpack the plumbing further in [what agentic commerce is](/blog/what-is-agentic-commerce) and the rival standards in our piece on [agentic commerce protocols](/blog/agentic-commerce-protocols).

How brands raise the odds of being picked None of this requires guesswork, because the inputs are largely controllable. Start with the feed, since for ChatGPT it is the front door: complete every attribute, include GTIN or MPN, map products to the correct Google taxonomy, and keep price and stock accurate to the hour rather than the week. Match the feed to the landing page exactly, because a price or availability mismatch is the fastest way to lose trust. Add Product, Offer, and AggregateRating schema so the on-page data corroborates the feed. Then work the off-site layer deliberately: maintain active, well-reviewed profiles on the platforms agents trust, and earn placement in the independent buyer guides and roundups they cite, the same off-site corroboration covered in [ChatGPT shopping brand strategy](/blog/chatgpt-shopping-brand-strategy). Finally, keep your brand entity consistent across the web, since agents resolve a product to a brand and a brand to a reputation. Do these well and you are not gaming the system, you are simply giving it the clean, corroborated signals it is built to reward.

You cannot improve what you cannot measure The uncomfortable truth underneath all of this is volatility. Agents re-roll their picks constantly, and our own monitoring shows the same shopping query can return a different set of products roughly 70% of the time, while the named list differs from one engine to the next. That means a brand can hold a near-perfect feed and accurate stock and still be absent from the answer a given buyer sees, with no error to point to. Spot-checking by typing your category into ChatGPT once tells you almost nothing, a problem we set out in [why spot-checking AI visibility fails](/blog/ai-visibility-monitoring-why-spot-checking-fails). The only reliable way to know whether your products are being named, where you sit against rivals, and which engines favour you is to measure it continuously across engines rather than sample it occasionally. Honeyb exists for exactly that: tracking which products and brands AI names, so the work you put into feeds, schema, and reputation can be checked against what agents actually recommend. Being chosen by AI, and now bought by it, is something you have to measure before you can improve it.