How AI assistants decide which products to recommend
What ChatGPT, Perplexity and Google's AI features look at when they build a shortlist of products, and what consumer brands can do about it.
Key takeaways
- AI assistants build product recommendations from two sources: what the model already learned and what it retrieves live from the web and product data.
- Assistants match products to the attributes in a shopper’s question. If your specs, price and use cases aren’t stated clearly and consistently, you’re hard to match.
- Independent sources such as reviews, press, creators and retailers help an assistant feel confident naming you.
- Nobody outside these companies knows the exact ranking logic. The practical move is to make your brand easy to understand, verify and quote.
The shopper’s question is the new search box
Shoppers used to type two or three words into a search bar and scroll. Now many of them describe what they want in full sentences: “best clean sunscreen for sensitive skin that doesn’t leave a white cast.” “Sustainable sneakers under $150 that work for standing all day.”
That shift matters. A long, specific question gives an AI assistant a lot to work with, and it usually responds with a short list: a handful of products, each with a sentence or two on why it fits. There is no page two. You’re either in that answer or you’re not.
So the question every consumer brand should be asking is simple: when an assistant builds that short list, what is it actually looking at?
Two ways an assistant “knows” your product
Broadly, assistants like ChatGPT, Perplexity, Gemini, Copilot and Google’s AI Overviews and AI Mode draw on two kinds of information.
What the model already learned
Large language models are trained on huge amounts of text. If your brand was discussed widely and consistently before a model’s training cutoff, in reviews, articles, forums and retailer listings, the model may already associate you with certain categories and attributes. If you launched recently or barely appear outside your own site, it may know little or nothing about you.
What it retrieves right now
Many assistants also search the web or query product data in real time, then summarize what they find. Perplexity leans heavily on live retrieval and shows its sources. ChatGPT can search the web and display shopping results. Google’s AI features sit on top of Google’s search index and shopping data. In those moments, the assistant is reading pages, yours and everyone else’s, and deciding what to cite.
Live retrieval is good news for challenger brands. You don’t have to wait for the next model to be trained. Pages that are clear, current and crawlable can show up in answers much sooner.
Shopping features run on structured product data
When an assistant shows product cards with images, prices and links, it is typically drawing on structured product information, not just prose. That can come from merchant feeds, retailer catalogs and the structured data on your product pages.
Google is the clearest example: product data submitted through Google Merchant Center and Product markup on your pages feed its shopping surfaces. Other assistants have described pulling from structured product metadata and third-party sources. Inside its own store, Amazon’s Rufus works from Amazon’s catalog, listings, reviews and customer Q&A.
The common thread: machines prefer facts they can parse. Price, availability, size, material, ingredients, certifications, ratings. If those live only in an image or a vague paragraph, you’re making the assistant guess.
The signals that tend to matter
We won’t pretend to know the weighting inside any of these systems. The companies don’t publish it, and it changes. But looking at what assistants actually cite and say, a few patterns show up again and again.
- Attribute match. The shopper asked for “fragrance-free,” “under $150” or “vegan leather.” Products whose pages state those attributes plainly are easier to match than ones that merely imply them.
- Clarity about who it’s for. “For sensitive, acne-prone skin” or “built for wide feet” gives the assistant a reason to recommend you for a specific need.
- Third-party corroboration. Reviews, editorial roundups, creator content and retailer listings that say similar things about you make a recommendation feel safe. We go deeper on this in our guide to the third-party proof AI leans on.
- Consistency. If your site says $48, a marketplace listing says $52 and an old article says the product is discontinued, you’ve created doubt.
- Freshness and availability. An out-of-stock product or a dead link makes a poor recommendation. Assistants with live data tend to favor what someone can actually buy.
Why the same question gets different answers
Ask the same question twice and you may get different lists. Ask it in a different assistant and the gap can be wider. That’s normal. Responses are generated, not looked up, and they can vary with wording, location, conversation history, personalization and whichever sources were retrieved at that moment.
Two implications. First, a single screenshot of your brand appearing (or not) proves very little. Second, the goal isn’t to “rank #1” for one prompt. It’s to show up reliably across the many ways real shoppers phrase their needs. We cover how to track that honestly in how to measure your visibility in ChatGPT and Perplexity.
Practical tip: Write down ten questions a real customer might ask before buying in your category, covering budget, skin type or size, use case and values. Those questions are your new keyword list.
What this means for your brand
Say you sell a mineral sunscreen out of Venice. A shopper asks for a reef-conscious, fragrance-free SPF for sensitive skin. To make that list, an assistant needs to find, on your site and elsewhere, that your product is mineral, fragrance-free, suited to sensitive skin, priced where the shopper expects, in stock, and well regarded by people who aren’t you.
That breaks down into a short, unglamorous checklist:
- Make product pages explicit about specs, ingredients or materials, sizing, price and use cases. Our guide to product pages built for AI shopping walks through the details.
- Add accurate structured data (Product, Offer, AggregateRating, Review) and keep merchant feeds in sync with your site.
- Keep product names and attributes identical across your store, marketplaces and retailers.
- Earn genuine third-party coverage: real reviews, honest creator partnerships with proper disclosure, and press.
- Track a consistent set of prompts over time instead of reacting to one-off results.
None of this is a trick. It’s the same work that makes your brand easier for a human to trust, done with machine readers in mind.
Where to start
You can’t control what an assistant says. You can control how easy you make it to understand your product, verify your claims and match you to the right shopper. Brands that do that well give themselves better odds. No guarantees, but a real edge.
If you want a clear read on where your brand stands today and what to fix first, explore our services or book a free strategy call.
