How AI shopping assistants find your products (and how to be found)

More shoppers now start with a question instead of a search: "a soy candle that smells like a cabin, under $40", "a gift for a coffee lover who has everything". An AI assistant answers with a few products, sometimes with a link to buy.

If you run a small store, the obvious question is how to be one of those products. This guide explains what these assistants read, what you can control, and where the honest answer is "nobody knows yet". We build Waise, so we also say what our stores do by default, but almost everything here applies to any platform.

How an assistant finds a product

There isn't one mechanism. Today an assistant usually combines several:

1. Web search. The assistant searches the web and reads the pages it finds, much like a search engine. Your product pages need to be reachable and readable.
2. Structured product data. Pages that describe a product in a machine-readable format (schema.org Product markup, with price, availability and reviews) are easier to read correctly than a page of loose text.
3. Product feeds. Some platforms accept a feed of your catalogue directly: Google Merchant Center has done this for years, and OpenAI publishes a product feed specification for merchants selling in ChatGPT.
4. Direct connections. Newer protocols let an assistant query a store's catalogue and even start a checkout, instead of reading web pages.

You don't need all four to be found. But each one you have makes it more likely an assistant gets your product right.

Step 1: let the right crawlers in

An assistant can only recommend a page it's allowed to read. Check your store's robots.txt (at yourstore.com/robots.txt).

OpenAI, for example, uses separate crawlers for separate jobs:

• OAI-SearchBot surfaces websites in ChatGPT's search features. A site that blocks it won't appear in ChatGPT search answers.
• GPTBot crawls content that may be used to train OpenAI's models. Blocking it does not affect whether you appear in ChatGPT search.
• ChatGPT-User visits a page when a user asks ChatGPT about it; it isn't automatic crawling.

(Source: OpenAI, "Overview of OpenAI crawlers", developers.openai.com/api/docs/bots, checked November 3, 2026.)

So you can allow search and refuse training if you prefer. Many stores accidentally block both with a blanket rule. Other assistants have their own crawlers with their own names; check each one's documentation rather than guessing.

Step 2: make every product page answer the question

Assistants match products to questions. A product page that says "Cabin. $34." gives them nothing to match. One that says "Cabin: a hand-poured soy candle with cedar, smoke and vanilla notes, 8 oz, about 45 hours of burn time, in an amber glass jar" can be matched to a dozen different questions.

For each product:

• A descriptive title. Product type plus the thing that makes it yours: "Cedar and smoke soy candle, 8 oz".
• The facts in words. Materials, size, weight, scent notes, burn time, care. If a shopper would ask it, write it.
• Variants as variants. Sizes and colours as options on one product, not separate pages that compete with each other.
• Real photos with descriptive alt text.
• Price and availability that are true right now. An assistant that recommends a sold-out product to a shopper does neither of you any favours.
• Shipping and returns stated plainly. "Ships in 2 business days, free over $50, 30-day returns" answers questions shoppers ask assistants all the time.

Step 3: give the data a structure

Structured data is a block of machine-readable information in each page, usually schema.org markup in JSON-LD. For a product it carries the name, image, price, currency, availability and, where you have them, reviews, shipping and return policy. Google describes two kinds of product results this supports: product snippets, for pages that describe a product, and merchant listings, for pages where shoppers can buy it. (Source: Google Search Central, "Product structured data", checked November 3, 2026.)

Most store platforms write this markup for you. Check yours with Google's Rich Results Test on one of your product pages.

Step 4: an llms.txt file

llms.txt is a proposal, published by Jeremy Howard on September 3, 2024: a plain Markdown file at /llms.txt that gives AI tools a short overview of a site and links to the important pages. It is not an official standard, and no assistant has promised to rank sites by it, but it is cheap to have and it is read by some tools. (Source: llmstxt.org, checked November 3, 2026.)

For a store, a useful llms.txt says what you sell, who you are, your shipping and returns, and links to your main collections.

Step 5: feeds and buying inside the assistant

Google Merchant Center takes a product feed and uses it across Google's shopping surfaces. If you sell in the US and haven't connected it, it's one of the highest-value hours you can spend.

OpenAI's Agentic Commerce Protocol (ACP) is an open protocol, developed with Stripe, for selling inside ChatGPT: the merchant provides a product feed and checkout endpoints, keeps its own payment processor and stays the merchant of record. OpenAI's developer documentation says Instant Checkout is currently available to approved partners, who apply through OpenAI. (Source: OpenAI developer docs, developers.openai.com/commerce, checked November 3, 2026. Details on eligibility, regions and fees come from secondary sources we could not confirm with OpenAI; check the current terms before you plan around it.)

MCP (Model Context Protocol) lets an AI tool connect to a service and use its tools. A store can offer a read-only shopping server that lets an assistant search the catalogue, read a product and create a link to checkout.

What a Waise store does by default

So you can see what "done" looks like, here is what every Waise store has, and where you control it:

• Structured product data on product pages, and a sitemap.
• Its own /llms.txt describing the store, its policies and its catalogue.
• A read-only shopping server for assistants (MCP) and a catalogue API: search, product details, and a signed checkout link. The shopper always pays on the store's own checkout. It answers only once the store is launched and the owner has opened the catalogue to assistants.
• An ACP product feed and checkout sessions for ChatGPT, behind a switch that is off by default. It works with Stripe only, and nothing is sold until OpenAI has approved the merchant and the owner has pasted OpenAI's signing secret.
• A GEO screen in the admin, which checks a store's readiness for AI search, lists each page's status, and lets the owner choose which AI crawlers may read the site.
• Never shared with assistants: customers, orders, costs, and exact stock counts beyond an "only a few left" line.

What nobody can promise you

Be wary of anyone who guarantees placement in ChatGPT, Perplexity or Google's AI results. The assistants don't publish how they choose products, the systems change often, and a small store has no special access. What you can do is remove the reasons to be skipped: blocked crawlers, thin descriptions, missing prices, unclear shipping.

We don't yet have data from Waise stores on how many orders come through AI assistants, so we won't quote one. When we do, it'll be in this blog with how we counted it.

A one-hour checklist

• Open your robots.txt and check which AI crawlers you allow
• Rewrite your five best-selling product titles and descriptions to answer real questions
• Make sure every product page shows price, stock, shipping and returns
• Run one product page through Google's Rich Results Test
• Check you have an /llms.txt, or write one
• Connect Google Merchant Center if you sell in the US
• Ask an assistant about your own product category and read what it recommends, and why