Does ranking still matter for AI shopping agents? Feeds now fill most US shortlists, but list position still steers what agents inspect

A claim has circulated through agentic-commerce guides since September: "Rank is dead when AI agents shop. The agent reads your product data, not your ranking". Trade restates in October recast it as "feeds, not web pages". The verdict is Overstated. In US ChatGPT Shopping, feed-integrated retrieval rose from 8.26% to about 65% of tracked recommendations after 10 July 2026 Wu · Profound · 2026B. Product data has taken over much of that retrieval step. Rank still governs what the claim leaves out: the shortlist outside the US, and the order in which an agent reads any shortlist. Listing position still causally shifts which products agents inspect, four to ten times more weakly than for human shoppers Wadi · Not stated in the arXiv paper (authors Davood Wadi; Yu Ma) · 2026A.

US ChatGPT Shopping now retrieves most products from feeds

In the US, the door into ChatGPT's shopping shortlist has moved toward product data. Profound tracked 1,757,723 ChatGPT Shopping prompt runs in July 2026 through its customers' network logs Wu · Profound · 2026B. The feed-integrated share of product recommendations jumped from 8.26% to 61.54% on 10 July, and the vendor puts it at about 65% by 3 September.

The same vendor had found the opposite pattern in June. In a 30-day snapshot of about one million offers, 88.29% were drawn from web product pages Wu · Profound · 2026B. Feed-derived offers were already taking the first carousel slot almost every time.

The shift also concentrated the shortlist. Unique merchants referenced fell from 13,524 to 10,607, while the top ten merchants' share rose from 22.5% to 41.8% Wu · Profound · 2026B. Both Profound studies are single-vendor and observational, built on customers' tracked prompts rather than shopper sessions. OpenAI has not confirmed the change.

Product feeds move the ranking into OpenAI's own index

Feeds change which ranked list counts rather than abolishing ranking. Peec AI reports that ChatGPT runs its own family of vertical retrieval indexes, internally called Labrador, including one for shopping Rudzki · Peec AI · 2026B. A shopping test that preferred the index over search-engine results ran on 8% of chats in mid-August.

Peec AI reads the stack as keyword scoring plus vector search, though those labels are its interpretation of experiment names Rudzki · Peec AI · 2026B. A product in a feed is still scored against other products for each query. The ranker is OpenAI's rather than Google's.

OpenAI's own documentation says as much. Its Help Center states that merchant lists are ranked by availability, price, quality and whether the merchant is the maker or primary seller OpenAI · 2026A. The claim treats feed inclusion as the end of ranking. The evidence describes a new index with its own ranking rules.

Outside the US, Google Shopping rank is still the best-documented route

OpenAI's merchant page says shopping in ChatGPT is currently live only for US users, with more regions planned OpenAI · 2026A. The page does not say how shopping results in other markets are sourced. That gap matters for any brand selling in Europe.

The strongest published evidence on the older route is Peec AI's March 2026 analysis of more than 43,000 carousel products across 10 verticals Landwehr · Peec AI · 2026B. About 83% matched Google Shopping's top organic positions, and 60% came from the top ten. Bing matched 11%.

That study predates the July shift and did not split results by market. Until a market-split study is published, the 83% figure is the best available baseline for markets outside OpenAI's feed programme. There, the best available evidence ties ChatGPT's shopping shortlist to Google Shopping rank, the reverse of the claim.

Web pages and model knowledge still feed ChatGPT's product picks

OpenAI names more inputs than feeds. Its Help Center lists structured metadata from first- and third-party providers and other third-party content OpenAI · 2026A. It also lists the model's own response, generated before it considers new search results. Review summaries come from public websites.

Asked whether a merchant needs a feed if ChatGPT already crawls its site, OpenAI answers no, though feeds give more control over how products appear OpenAI · 2026A. Shopify and Etsy catalogs are integrated without an application. A brand can therefore sit in the feed through its platform, without an integration of its own.

On Profound's September figures, roughly a third of recommendations still came from outside the feeds Wu · Profound · 2026B. That share is the remainder of the 65% feed figure, not a separate measurement. It is a smaller door than in June, and it is still open.

Inside the shortlist, list position still moves AI agents

Two randomised studies test whether agents ignore position once products are in front of them. Davood Wadi and Yu Ma randomised the order of 100 hotel listings across 2,000 agent sessions on four LLMs Wadi · Not stated in the arXiv paper (authors Davood Wadi; Yu Ma) · 2026A. Position predicted which listings were inspected for every model, four to ten times more weakly than for humans. Because order was randomised per session, the effect is causal by construction.

The effect was non-monotonic. Inspection bottomed out near ranks 68 to 74, so the weakest place on a long agent-read list is the middle. Position reached the booking stage for only two of the four models.

A second audit finds a firmer effect at the choice stage. A pre-specified randomised conjoint across twelve models found that list position, a content-free artifact, shifted recommendations by roughly twelve dollars a night Baig et al. · Fandaqah (Al Khobar, Saudi Arabia); Hamdard University (Karachi, Pakistan) · 2026A. That audit offered five hotels against 100 in the first study. List length is the obvious moderator, and nobody has tested it.

Displayed attributes decide the pick, when the agent reads them

The two studies agree on what decides the final choice. In the conjoint audit, a top guest rating raised selection by 31.6 percentage points and a high price lowered it by 30.0 Baig et al. · Fandaqah (Al Khobar, Saudi Arabia); Hamdard University (Karachi, Pakistan) · 2026A. In the Wadi and Ma study, all four models converged on the same undominated listing, which took 78.2% of bookings Wadi · Not stated in the arXiv paper (authors Davood Wadi; Yu Ma) · 2026A.

Attributes count only when the agent reads them. In a follow-up, the same authors made eight commercial LLM agents pay for each attribute through a tool call Wadi · Desautels Faculty of Management, McGill University · 2026A. With a vague "best deal" goal, agents chose the cheaper-per-ounce coffee 29% of the time, against 89% when inspection was free. The authors place the weakness in storefront information architecture rather than in the models.

Both agent studies ran on supplied listing pages, not on ChatGPT Shopping or another live surface. Live agents also meet closed doors: TechCrunch reports that Amazon blocks Meta's Muse shopping agent and that Walmart's human-verification checks block agents unintentionally Perez · TechCrunch · 2026B. Evidence on how widespread such blocking is remains anecdotal.