ChatGPT Shopping does not select its products in the same way in the United States and Europe, according to Peec AI, a developer of visibility‑tracking tools for generative AIs. According to the company, the U.S. system relies on an OpenAI‑specific index, whereas the European system still depends on Google Shopping.
- OpenAI uses its own Labrador index for U.S. shopping, while Europe still relies on Google Shopping.
- In the United States, product feeds and seller verification determine which products appear.
- Descriptions are cut at about 50 tokens, so the opening text carries the most weight.
- In Europe, rankings mirror Google Shopping results, and Labrador may arrive there later.
In a blog post published on October 16, 2026, Metehan Yesilyurt notes that the team gained access to a ChatGPT debug API. For each prompt, it returns roughly 150,000 rows of data on product retrieval, including the score assigned to each. Malte Landwehr, CPO and CMO of Peec AI, presented these findings at brightonSEO 2026. These observations have not been publicly confirmed by OpenAI.
According to this data, OpenAI switched U.S. shopping to its own system a few weeks before that conference. In the United States, the Labrador index, built from product feeds, is enabled, while Google Shopping scraping via a SERP API—a third‑party service that extracts search engine result pages—is disabled. European sessions show the opposite configuration.
In the United States, the product feed determines visibility
At the time of the conference, U.S. shopping relied on 669 feeds, of which 606 were active. About 70 came from direct partnerships, such as a brand like Nike, which supplies its own feed to OpenAI. Nearly 500 were built by OpenAI by scraping merchants’ sites via Bright Data, with the remainder routed through Shopify.
To observe the selection, Peec AI tested a prompt about running shoes. ChatGPT retrieved 41 candidate products, turned 32 into detailed “prompt blocks,” and then displayed 5. The data show several signals, including the embedding scores, which measure semantic proximity between the query and the product, and the BM25 score, a ranking based on the frequency of query terms in the text.
Each candidate carries an is_site_verified_seller attribute. Without it, the store is excluded. In the test, one merchant held 25 of the 32 candidates yet placed none of its products in the final results. Peec AI acknowledges not having identified what determines this attribution. Dick’s Sporting Goods, on the other hand, had five candidates, four of which appeared among the five results. Peec AI concludes that it is better to submit each variant, by model or by color, as a separate entry.
Descriptions are also truncated to exactly 50 tokens, roughly 40 words or 250 characters in English. In the test, one of them stopped mid‑sentence, “Pattern offers a smooth heel to…”. The remainder of the text is not considered in the scoring. Peec AI notes that this equivalence applies to English and that the readable portion could be shorter for a French listing, “but that remains to be measured.”
In Europe, a simpler and slower operation
For a test prompt, ChatGPT sent five fan‑out requests to Google Shopping via a SERP API. Each returned between 1 and 40 products. The results were then merged sequentially, yielding a list of 113 products, of which 18 were displayed. The entire step took about four seconds, versus roughly 0.6 seconds in the United States.
No embedding, no BM25 and no GTIN matching are applied in Europe, and no seller verification is performed. According to Peec AI, a good ranking in ChatGPT Shopping in Europe therefore directly mirrors that achieved in Google Shopping, and may not withstand the arrival of Labrador in the region. No date has been announced for this rollout.
