AI chatbots find products in five steps. They turn the buyer's question into constraints, retrieve candidate pages, read the attributes on each page, compare the candidates and recommend a few by name. Many stores drop out at the second step, retrieval, before the chatbot has judged anything about the product.
A store the chatbot never fetched does not get judged at all, however good its products are.
What are the five steps from question to recommendation?
Each step discards candidates, so a store that fails early never benefits from the work it did on later steps.
| Step | What the chatbot does | How stores fail at this step |
|---|---|---|
| Interpret | Turns the question into constraints such as category, budget and use. | This step belongs to the assistant, so a store cannot fail here. |
| Retrieve | Gathers candidate pages from a search index, its own crawler or a product feed. | Crawlers are blocked, pages are not indexed or the content appears only after scripts run. |
| Read | Extracts attributes such as size, material, price and stock from each page. | Attributes live in images or in lifestyle prose. |
| Compare | Lines the candidates up against the buyer's constraints. | The page leaves out the attribute the buyer asked about. |
| Recommend | Names a few products and explains why. | The page holds nothing specific enough to justify the choice. |
Why do most stores drop out at retrieval?
Retrieval happens before any judgment about quality. OpenAI's crawler documentation says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers. Anthropic and Perplexity document search crawlers of their own that play the same role.
In our June 2026 study of 216 Shopify storefronts, 41% failed Front Door, the most basic test of whether an AI crawler can read the page at all. Front Door was the strongest of our five layers, and two in five stores still failed it. Our guide to sitemap health covers the part of retrieval that lives in your sitemap.
What makes a product easy for a chatbot to read and compare?
A chatbot compares on the attributes it can see. A size chart that exists only as an image hides the size. A material mentioned once in a paragraph about a weekend in the mountains may not be extracted. Stock that appears only after a script runs may read as unknown.
Product identifiers such as the barcode belong here too, since they let the same product be matched across sources. Metafields give each attribute its own field, which is easier to read than a sentence.
How do I check each step on my Shopify store?
Work through the steps in order, since a fix on a later step is wasted while an earlier one fails.
- Open yourstore.com/robots.txt in a browser and check that it does not disallow the whole store for AI crawlers such as OAI-SearchBot, PerplexityBot or Claude-SearchBot.
- In Google Search Console, inspect one product URL with the URL Inspection tool and confirm that it is indexed.
- Open a product page, right click, choose View page source and search for the price and the main material. If they are not in the source, a crawler that does not run scripts cannot see them.
- In your Shopify admin, open Products, open the product and check the Variants section. Each variant needs its own SKU and, where you have one, its Barcode.
- Go to Settings > Custom data > Products and add a metafield definition for each attribute buyers compare on, such as material or dimensions. Fill the values in on each product page and show them on the product template through Online Store > Themes > Customize.
- Rewrite the first sentence of the product description so it states what the product is, who it suits and the one fact that sets it apart.
At which step does a running shoe store that stocks the right shoe drop out of a chatbot answer?
Picture a running shoe store on Shopify with 30 models. A buyer asks a chatbot for a cushioned road shoe with a low heel drop for wide feet. The store has the right shoe.
At retrieval, the store passes. Its robots.txt is the Shopify default and its product pages are indexed. At the reading step, it fails. The heel drop appears only in a spec image, and "wide" is a variant option without its own text on the page. The chatbot cannot match two of the buyer's three constraints, so it is more likely to name shoes from stores that wrote those facts down.
The owner adds a heel drop metafield and a width metafield, shows both on the product template and writes one sentence in each description that states them. The shoe can now be compared on the facts the buyer asked about.
Does the path differ between chatbots?
Each assistant gets its pages a little differently, and the five steps stay the same. One assistant relies on its own crawler, another on a search index you already optimize for, and another fetches pages live and cites them. Our guide to how each AI finds your store has the differences, and how AI chooses between two similar products covers the final step, when two candidates are close.
The free scan checks the retrieval step for eight AI crawler names and reads the structured data and text of five key pages. You will see which step your store fails first.
Questions owners ask
How do AI chatbots find products to recommend?
They interpret the buyer's question, retrieve candidate pages from a search index, their own crawler or a product feed, read the attributes on each page, compare the candidates and name a few. A store has to pass retrieval before anything about its products counts.
Do I need a product feed?
A feed helps on shopping surfaces that read feeds, and it states attributes in a clean way. It does not replace product pages that a crawler can read, since many assistants still fetch and read the page itself.
Do product images help AI chatbots?
Images help a buyer and some assistants can read them, but do not rely on an image for a fact. Anything that decides a purchase, such as size, material or compatibility, should also exist as text or as a metafield.
How do I know where my store drops out?
Work backward. If your pages are not indexed or crawlers are blocked, the problem is retrieval. If the pages are indexed but never named, the problem is reading or comparison, which usually means missing attributes.
Sources
- Overview of OpenAI crawlers, OpenAI, read November 9, 2026.
- Perplexity crawlers, Perplexity, read November 9, 2026.
- Does Anthropic crawl data from the web?, Claude Help Center, read November 9, 2026.
- The State of AI Visibility 2026, Visibility Mesh, 216 Shopify storefronts scored in June 2026.