How AI agents "see" your products: no feed, no existence
- pipeline
- product
- observability
Target reader: E-commerce founders / heads of operations · non-technical readers who care about "why this matters to me"
The habit of searching for products is shifting. ChatGPT Instant Checkout shipped, Perplexity Shopping went public beta, Google AI Overview reached the product layer. A growing share of purchase intent now skips the search results page entirely and lands on "I recommend you buy X from Y", straight from an AI agent.
One fact is easy to miss: the AI agent isn't crawling the whole web. It picks results from a structured feed. That feed is maintained by OpenAI Commerce, Google Merchant Center, and Perplexity Shopping. No feed = doesn't exist in the agent's view. No amount of SEO ranking saves you.
The structured feed: agents read product cards, not pages
Traditional SEO solves "how does my site get found by Google." The AI agent isn't Googlebot, and it has no use for HTML. It wants a card it can drop straight into the conversation: product name, price, purchasable URL, shipping range, stock status. If those fields aren't delivered structured, the agent recommends a competitor whose data it already holds structured.
For small and mid-sized merchants this means three new things: (1) learn a fresh submission standard (OpenAI Commerce, Google Merchant UCP, and Perplexity each have their own schema); (2) maintain it continuously: SKU changes, stock changes, the feed has to sync; (3) know whether you're actually being cited: submitting isn't being seen, and being seen isn't being cited. None of the three is a one-time job.
Products with clean structured data get named
This comes from continuous live testing. Across the leading AI shopping engines (Doubao, Qwen and DeepSeek included), SoldByAI runs the same observation: pose the real "buy + product line + size" questions a shopper would ask, capture each engine's answer, and parse exactly which domains and product URLs get cited. The pattern is consistent: products with clean, structured data get named; the ones without simply don't appear. No amount of SEO ranking changes that.
Three ways buyers ask, one selection logic
Buyers phrase their needs differently. The agent's selection works the same way: narrow the category by need, then pick a short list from machine-readable product data. Three question patterns that real buyers use show what an answer is built from.
| Scenario | What the buyer asks | What the answer is made of |
|---|---|---|
| Digital-native accessories | By budget and destination: an earbud case under thirty dollars that ships locally | Answers usually keep two or three candidates, with price, material, and pickup options; products with incomplete data do not enter the comparison |
| Corporate gifting | By purpose and quantity: locally customizable notebooks for a fifty-person team | Answers group options by gifting scenario and surface the few that are customizable with local pickup; products missing specifications or minimum-order terms are hard to place |
| Home fragrance | By region and fulfilment: fragrance brands with local pickup | Answers list brands beside individual products; items that state pickup coverage and stock status get named first |
What the named products share across the three scenarios is not a category. It is data: identity, price, availability, and delivery coverage complete enough for the agent to justify the pick.
How SoldByAI runs it · the weekly loop
That observation is productized into a weekly loop. Connecting takes minutes, through whichever of three intake paths fits: point SoldByAI at a storefront URL for a direct read, authorize Shopify or WooCommerce in one click, or upload an Excel/CSV template. SoldByAI structures the catalog into an agent-readable draft and distributes it to the three endpoints agents buy from: Google Merchant Center, OpenAI's ACP, and Google's UCP. Nothing goes live until you approve.
From there the weekly loop runs: real shopping questions probed across the leading AI engines, every error traced to its source field, fixes proposed line by line, automatic re-testing after each fix. The results land in your dashboard: which SKUs are clean and which are missing fields, distribution status across the three endpoints, which queries cited you this week and which didn't, and the fields to fill that are projected to lift next week's hit rate. A weekly report mirrors the same digest to your inbox.
After connect, four jobs move to SoldByAI
- You don't have to learn each endpoint's submission standard yourself. SoldByAI distributes your catalog to the three endpoints agents buy from: Google Merchant Center, OpenAI ACP, and Google UCP.
- You don't have to hire someone to maintain the feed. When your catalog changes, the pack regenerates automatically; SKU retirement, stock updates, price changes all follow.
- You don't have to guess how AI treats your products. Each week's probe results show exactly which queries hit, which went to competitors, and why.
- You don't have to build your own monitoring or train an operations team. The dashboard holds this week's changes on one screen, the weekly report mirrors it to your inbox, and nothing goes live without your approval.
Product-data requirements in agent commerce are still taking shape. The earlier a catalog is structured, the earlier the brand establishes retrievable product data and its first measurement baseline. The complete timing case appears in "The Great Entry Migration."
