ChatGPT ran its shopping answers almost entirely on web search through early 2026, per Profound’s research report published 3 September. That changed fast: reliance on feed-integrated retrieval “exponentially accelerated” from May, then jumped from 8.26% to 61.54% in a single day, 10 July, what the report calls a roughly 6.5x increase, which it ties to OpenAI’s release of GPT-5.6 the day before. The shift moved where product answers come from, and it redistributed visibility across the brands Profound tracks unevenly enough that most of them lost ground.
Why Did ChatGPT’s Shopping Answers Change in July?
Profound’s report, authored by Allen Wu and titled “ChatGPT 5.6 has transformed Shopping mode,” found that feed-integrated retrieval for shopping prompts jumped from 8.26% to 61.54% on 10 July, and that roughly 65% of product recommendations since then require a merchant’s feed to be connected. OpenAI released GPT-5.6 on 9 July. The report calls the release “the only event-based factor on this exact date that has such explanatory power,” but stops short of confirming the two are causally linked.
The Losses Outnumbered the Gains
Profound tracked 687 of its own customers and compared their visibility for 7-9 July with 10-12 July, the jump day and the two that followed. Of that group, 517 swung by at least a third: 450 lost visibility, 67 gained. A regression model built on web-search loss and feed-retrieval gain together, per the report, “explains 83% of real, observed visibility swings.” Both variables in that model describe the retrieval path rather than anything a merchant publishes.
Fewer Merchants Are Getting Referenced
The report’s consolidation analysis, drawn from 33,167 prompt runs sampled at 10% between 7 and 12 July, found the top 10 merchants’ share of all product references rose from 22.5% to 41.8%, while the number of unique merchants referenced at all fell more than 20%, from 13,524 to 10,607. That concentration, not the headline retrieval-share number, is the exposure a merchant actually carries: getting referenced at all now depends on a shorter list.
| Metric | Before | After | Compared periods |
|---|---|---|---|
| Feed-integrated retrieval share | 8.26% | 61.54% | across 10 July |
| Top-10 merchant reference share | 22.5% | 41.8% | 7-9 July vs 10-12 July |
| Unique merchants referenced | 13,524 | 10,607 | 7-9 July vs 10-12 July |
Separately, the report puts Shopify’s share of all feed-integrated retrieval at about 35%. The report also runs an unsampled 281,356 prompt runs from 10 July through 24 August for its merchant analysis. In August, per the report, feed-integrated sources overtook web search overall as ChatGPT’s shopping source.
Feed Access Runs Through a Waitlist, Not a Ranking
Shopify and Etsy catalogs connect to ChatGPT automatically, according to Search Engine Journal’s 9 September coverage of the report. Other retailers apply through OpenAI’s merchant application page and wait. No source in this story reports the terms, cost, timeline or acceptance rate of that process. None of the sources describe any route into feed-integrated retrieval other than a connected feed, and that path now supplies roughly two-thirds of ChatGPT’s product recommendations.
Measuring It From the Merchant’s Side
Profound’s numbers measure retrieval and reference shares. The nearest thing a merchant has in its own reporting is the “AI Assistant” default channel GA4 made available across properties on 7 June, which breaks ChatGPT, Gemini, Copilot, Grok and DeepSeek sessions out of generic referral. It does not distinguish a session that began in a shopping answer from any other assistant session. Google, separately, is testing a report inside Google Ads that shows advertisers a sample of the product titles its AI generated for Shopping ads. That is a different mechanism on a different platform, and no source connects it to anything OpenAI did.
What’s Established, and What Isn’t
What’s measured: a one-day jump in retrieval share, a lopsided redistribution across Profound’s own customer base, and a shrinking, more concentrated list of referenced merchants. What isn’t: SEJ reports that, as of its publication date, OpenAI had not confirmed any connection between GPT-5.6 and the shopping shift, and that the model’s release notes contain no shopping-related updates. Profound’s own framing stays correlational, an explanatory regression rather than a confirmed mechanism, built on one quoted line about explanatory power. The feed connection is what changed. Why the retrieval path moved on that exact date is something OpenAI has not publicly linked to GPT-5.6.
