Knowledge base›Sources›Tom Wells (Vaer AI)
AI retailer bias study: how AI shopping answers favour chains over small shops
Key findings
Wells has just under half a million shopping answers from ChatGPT, Gemini, Google AI Mode and AI Overview evaluated for 20,000 size-neutral questions across four North American cities and ten categories, with and without web search. Without search the AI names a chain 63-70 percent of the time, with search still 46-58 percent; a small shop lands in roughly one answer in four. The bias is already in the search queries the model writes itself: when it types a store name, it is almost always a chain. Small shops appear in the source links but not in the recommendation. The more specific the product, the bigger the retailer. The word "independent" flips the picture, "local" does not.
Why the source matters for the project
It is the first large output statistic that traces the retailer bias back to the model's own search queries, that is, to what our fast and product lines show. It also provides the test idea for the German market: classify domains in Domain filters per question and sources_footnote by retailer size and count the merchant slot of the product (product search) mode separately. What remains open is whether the 97 percent finding is a US effect or simply not visible with the German prompt set.
Checked against our data
The core of the study, that the model decides before searching whom it will check, is what we see every day in the fast and product lines: "idealo" in every Nike query, dyson.de as a domain filter, five shoe models with version numbers. The direction "specific product -> big retailer" also holds for us: the vague advice question cites Munich specialist stores, the Nike and Dyson questions Amazon, Zalando, MediaMarkt and Saturn.
Two things look different for us. First, no chain name appears in any of our queries; what ChatGPT types for German questions are manufacturers, price comparison sites and, once, two independent stores. Second, the merchant slot in the product widget is not a chain slot for us: for the running shoes it lists Sport Hopfmann, Sport Weber, Il Corridore, Sport Conrad, Sport Forster. That slot is not filled from web search but from the product provider with Google Shopping IDs. Whoever is there with a feed makes it into the card, regardless of size.
The half of the study without web search we cannot reproduce: our shopping prompts always search. For us those numbers describe the model knowledge that flows into the fan-outs, not the answer.
| Claim | Status | Evidence |
|---|---|---|
| ChatGPT writes retailer and product names into its own search queries before it searches: in about a third of runs for vague questions, in two thirds for branded products | confirmed | Same mechanism in our fast/product lines: K3-01 (Nike Pegasus 41) carries 'idealo' in the query text plus the domain filter idealo.de on 5 of 5 days, K3-02 (Dyson V15) the filter dyson.de on 3 of 5 days, K2-02 names five shoe models with version numbers. For the vague local question K1-01, store names appear in only 1 of 8 runs (20-25 Aug 2026). |
| When ChatGPT types a store into the query, it is a chain 97% of the time; small and local shops stay at 1-4% | unverifiable | Not testable with three shopping prompts and one local prompt. What we see differs: no chain name appears in any of our queries. ChatGPT types manufacturers (nike.com, dyson.de), price comparison sites (idealo) and, for K1-01 on 20 Aug, two independent Munich stores (Lauf-bar, Sport Schuster). German prompt set, US study: not a contradiction, an open check. |
| The more specific the product question, the bigger the recommended retailer: for 'toys for a six-year-old' a small shop gets one in three slots, for 'LEGO Technic Ferrari' one in ten | confirmed | Same direction for us, n = 3 prompts: K1-01 (advice, vague) cites independent Munich stores in 8 of 8 runs (lauf-bar.de, sport-schuster.de, sportforster.de). K3-01 (Nike Pegasus 41) cites amazon.de, zalando.de, intersport.de, jdsports.com, footlocker; K3-02 (Dyson V15) mediamarkt.de, saturn.de, amazon.de. Electronics is the chain category for us too. |
| Funnel: in the sources, large and small retailers are level (about 38% each); by the top recommendation the small ones drop away. For ChatGPT, 41% of citations come from large chains but 58% of recommendations | unverifiable | The stages are measurable for us (grouped_webpages -> sources_footnote -> merchant slot), but we do not classify domains by retailer size. Notable: the merchant slot in the product widget shows almost only mid-sized sports retailers for K2-02 (Sport Hopfmann, Sport Weber, Il Corridore, Sport Conrad, Sport Forster), and Footlocker, SPORT 2000, Keller Sports, Handballcompany for K3-01. That slot is filled by the product provider with Google Shopping IDs (e-commerce finding of 26 Aug 2026), not by web search. |
| With web search off, the AI recommends a chain 63-70% of the time and a small shop about 10%; in a neutral head-to-head the bigger store wins 90-94% | unverifiable | Not reproducible for us: all six shopping and local prompts searched in 44 of 44 runs (20-25 Aug 2026, instant and thinking). The no-search numbers describe model knowledge, not the answer a Plus user gets. They show what the model writes into its fan-outs; see claim 1. |
| 'local' and 'near me' change almost nothing because the AI counts a chain branch as a local store; 'independent' lifts small shops from a third to four fifths of the picks | unverifiable | We do not measure modifier variants. K1-01 ('in Munich ... good advice') triggers the business mode with location Munich and returns independent stores in 8 of 8 runs, no chain branches. Whether 'advice' or the location causes that is open. |
Related
wells-2026-02-12 mohanadasan-2026-08-10 konitzny-2025-11-19 blyskal-2025-12-08