Inside ChatGPT's Shopping Connector: Query Fanouts Update
Our extension now captures product recommendations from ChatGPT, showing you the metadata behind each product listing.
Capture Product Recommendations and Their Metadata
A Quick Recap
In a previous blog, we introduced Query Fanouts, the sub-queries ChatGPT generates when researching your question. Our extension captures those, giving you visibility into what the AI is actually searching for.
For shopping-related questions like "best mechanical keyboard under 10000" or "wireless earbuds for running" etc, ChatGPT goes a step further. It doesn't just search the web; it fetches product data and displays recommendations directly in its response, complete with prices, ratings, and merchant links.
We wanted to know where that product data comes from and how it gets there. So we built a feature to capture it.
What's New: Product Recommendations Capture
Now our extension can also capture the products listed in ChatGPT's response, alongside the query fanouts we were already capturing. In addition, we extract and display the metadata behind every product being listed in GPT's response, data that isn't visible to users but tells us more about where these products come from. When you run a query that triggers ChatGPT's "Shopping Connector" tool to display products, you'll see a new section in your report.
The products are displayed in the same carousel format as ChatGPT shows them, with the product title, image, price, and merchants visible on each card. When you click on any product, a metadata panel appears below with additional data collected from the API response.
Product Metadata
When you click on a product in the carousel, the metadata panel appears below with three sections:
Overview
- AI Rerank - Shows the product's position in the final carousel (e.g., #1) and how much it moved from its original position (e.g., ↑2 means it moved up 2 spots)
- Source Pos - The original position and citation reference from the API response (like
turn0product2). Every product we captured has this indexing pattern, though what exactly the number represents isn't entirely clear. It could be the order in which products were fetched from the data provider, or something else. What we did observe is that the final carousel order doesn't always match this index, suggesting some rearrangement happens before display. - Rating - Star rating and review count when available
- Listing Type - A label we derive from the underlying identifiers: Multi-Seller, Single in Cluster, or Standalone (more on this below)
Search Context
- Market - Country and language codes (e.g., IN / EN)
- Location - User location if detected in the response (e.g., Haryana, India)
- Query - The search query that was sent to fetch products, which sometimes differs from what the user originally typed
Data Sources
- Provider - Where the product data appears to come from (e.g., Google Shopping p2)
- gpcid - A product cluster identifier we found in the decoded response
- Merchant ID - A merchant identifier when present
- Catalog ID - A catalog reference number
What We Observed
We ran a handful of product queries like earbuds, power banks, keyboards, gaming mice etc and looked at the captured data. A few patterns stood out.
The p2 Provider
Every product we captured had a providers: ["p2"] tag. Based on the structure of the decoded tokens, which contain fields like gpcid, mid, and Google-style query parameters, this appears to be Google Shopping data.
We can't say this definitively, but the field names and format are consistent with Google's product infrastructure.
Carousel Order vs Internal Index
Here's something we didn't expect. Each product has an internal reference like turn0product0, turn0product1, etc. You'd assume product0 appears first in the carousel. It doesn't always.
In one earbuds query, the product referenced as turn0product4 was displayed first, while turn0product0 showed up second. ChatGPT seems to reorder products after receiving them, possibly based on relevance to the specific question, or some ranking logic we can't see.
The MID Pattern
We noticed that some products have all three identifiers (catalogid, gpcid, mid), while others only have a product ID. When we clicked on products that had mid present, the detail view showed multiple merchant offers with price comparisons. Products without mid showed only a single seller.
This led us to create three labels:
| Label | What We Observed |
|---|---|
| Multi-Seller | Has mid, multiple merchants shown, price comparison available |
| Single in Cluster | Has gpcid but no mid, cluster exists but only one seller appeared |
| Standalone | Only productid, no cluster structure, single exclusive listing |
These labels are our interpretation based on a limited sample. They might not hold for every query or market.
Query: "best rated power bank 20000mah"
Location Context
Every product token we decoded contained a uule parameter, an encoded location string. In our tests (run from India), this decoded to "Haryana, India" with gl: in and hl: en.
This suggests the Shopping Connector is geo-targeted. The products you see may differ based on where you are, or where ChatGPT thinks you are.
What This Tool Helps With
We built this feature for SEO and e-commerce professionals who want to understand how products surface in AI responses. The extension captures what ChatGPT receives, so you can:
- See the raw data, not just what's displayed but the underlying payload
- Track carousel positions and know where products rank for specific queries
- Identify metadata patterns like cluster IDs, merchant tags, provider sources
- Export for analysis, copy JSON, compare across queries, build your own datasets
A Note on What We Don't Know
There's a lot we haven't figured out:
- Why does ChatGPT reorder products? We see that it does, but not the logic behind it.
- What are p1 and p3? These appear in some metadata fields but not as primary providers. Could be enrichment sources, internal systems, or something else entirely.
- How stable are these patterns? Our observations come from a small set of queries in one market. Results may vary.
We'll keep exploring. If you run queries and notice something interesting, we'd love to hear about it.
Try It Yourself
The Product Recommendations capture is live in the latest version of our extension. Run a shopping query on ChatGPT, and you'll see the data appear in your report.
Get the Extension →What's Next
This is Part 2 of our series on what happens inside ChatGPT when you ask a question. We've covered Query Fanouts and now the Shopping Connector.
There's more to explore, like how citations are structured, how news results differ from web results, and how the AI decides what to include in its response. We'll share what we find.