65.4 %
web cards in the US
868 of 1,327 cards have no defined source marker; OpenAI feed: 30.7 %
71.4 %
web cards in first position in the US
300 of 420 first positions; OpenAI feed: 26.7 %
23/31
cheapest confirmed model card first
Mixed-source US answers; four apparent premiums, no confirmed price disadvantage
20.5 %
Google Shopping cards in Germany
246 of 1,201 cards; none with the defined OpenAI feed marker
Own measurements, October 22–24, 2026, signed-out ChatGPT. Source shares refer to cards; the price comparison covers mixed-source answers with a confirmed product model.
Data and method
The study measured signed-out ChatGPT for the United States and Germany from October 22 to 24, 2026. The curated set contains 168 questions across twelve product categories, covering category searches, specific products and comparisons. Each question was asked twice per day. That produced 1,008 answers with no reported collection error, 504 per market. Every expected question and repetition pair is present, with no duplicates.
Answers were collected automatically through a search results service. The analysis counted the product cards supplied and their order. These results may differ from what you see in your own browser. They describe signed-out cards, rather than every ChatGPT interface or every merchant.
Cards were assigned to sources using observable card and link markers, in this order:
- F – OpenAI feed: The card link contains
utm_medium=feed. - G – Google Shopping: The card contains a specific Google Shopping product identifier.
- H – merchant ID only: A merchant article number remains, without either of the preceding markers.
- W – web: None of these markers was found. The price’s actual origin remains unknown.
A web card does not prove that its merchant has no feed. Ten US cards contain the alternative marker utm_medium=openaifeed. Under the fixed classification, they are assigned to H or W depending on whether a merchant ID is present, rather than to F. These source shares therefore do not measure every merchant’s actual feed participation.
US: Most cards are classified as web
420 of 504 US answers contain cards (83.3 %). Of 1,327 cards, 408 carry the defined OpenAI feed marker (30.7 %), while 868 are classified as web (65.4 %). In 239 of the 420 answers containing cards, there is no F card at all.
October 22–24, 2026 · ChatGPT, signed out
| Source | United States | Germany |
|---|---|---|
| OpenAI feed | 408 (30.7 %) | 0 (0.0 %) |
| Google Shopping | 2 (0.2 %) | 246 (20.5 %) |
| Merchant ID only | 49 (3.7 %) | 4 (0.3 %) |
| Web (no source marker) | 868 (65.4 %) | 951 (79.2 %) |
Shares of cards, not shares of answers, distinct products or merchants. “Web” means none of the defined source markers was found; it does not prove that the merchant has no feed.
Cards without the defined source markers are common in this sample; the actual visibility of products without a feed was not established. The unit matters: these are shares of cards, not shares of answers, distinct products or merchants. The same models and merchants can appear repeatedly.
Web cards also lead in first position
In the US, a web card appears first 300 times (71.4 %), compared with 112 first positions for feed cards (26.7 %). Most first positions in this sample therefore lack the defined source markers. That does not establish how the ranking mechanism works.
October 22–24, 2026 · ChatGPT, signed out
| Source | United States | Germany |
|---|---|---|
| OpenAI feed | 112 (26.7 %) | 0 (0.0 %) |
| Google Shopping | 0 (0.0 %) | 54 (12.3 %) |
| Merchant ID only | 8 (1.9 %) | 0 (0.0 %) |
| Web (no source marker) | 300 (71.4 %) | 384 (87.7 %) |
Shares of first card positions in answers containing cards. This is the first card overall, before any product matching for the price analysis. Source markers describe the cards, not a merchant’s actual feed participation.
Prices: Four apparent premiums, no established price disadvantage
The direct price comparison uses only specific-product questions, positive prices in the local currency, and cards whose title or link confirms the requested model. Accessories, used items and recognisably different models are excluded. Within each answer, F cards are compared with every other source. Additional offers mentioned only in the answer text do not count as cards.
For the price analysis, “first” means the first confirmed card for the requested model. It need not be the first card in the answer overall. Matching a model also does not guarantee the same size, colour, equipment or purchasable variant.
In 23 of 31 mixed-source US answers (74.2 %), the first confirmed model card is also the cheapest confirmed card. In 4 of 31 answers (12.9 %), the extracted structured fields put a more expensive feed card ahead of a cheaper non-feed card. The descriptive 95 % Wilson interval is 5.13–28.85 %. These 31 answers cover just 17 distinct product questions or models, so the observations are not independent.
| Date and product | Extracted card prices and positions | What the review found |
|---|---|---|
| Oct 22 · UPPAbaby Vista V3 | Baby enRoute (F), position 2: 735.39; Capital Kidz (W), position 3: 719.99 (+2.1 %). | Position 1 is an open-box item and is excluded. Equipment cannot be reliably matched. No price disadvantage for the same offer is established. |
| Oct 23 · Sony WH-1000XM5 | Target (F), position 1: 399.99; Amazon/Best Buy (W), positions 2/3: 198.00 each (+102.0 %). | Extraction conflict: the title and answer text name Sony first at 198.00, while the structured fields name Target at 399.99. No confirmed premium in the displayed answer. |
| Oct 24 · Cuisinart DCC-3200 | Home Depot (F), position 1: 89.95; eBay (W), position 3: 78.00 (+15.3 %). | The answer text adds 19.50 shipping for eBay, giving 97.50 before tax. Home Depot’s listed item price is 89.95; its shipping costs are not specified. This does not establish a total-price disadvantage for the feed offer. |
| Oct 24 · ASICS Gel-Nimbus 26 | Pacers (F), position 1: 95.00; Amazon (W), position 2: 75.00 (+26.7 %). | Pacers: men’s US 11.5. Amazon: women’s, with only an approximate reported promotional price and a search link. No identical available variant is confirmed. |
Excluding configurations classified as non-comparable under the product rules leaves 27 mixed-source US answers. In 3 of them, the extracted fields put a more expensive feed card first (11.1 %; descriptive 95 % Wilson interval 3.85–28.06 %). This restriction still does not guarantee identical sizes, colours or gender variants. The sample is small.
Germany: Google Shopping data, no defined OpenAI feed marker
In Germany, 438 of 504 answers contain cards (86.9 %). None of the 1,201 cards carries the defined OpenAI feed marker. 246 cards (20.5 %) carry Google Shopping identifiers; 951 (79.2 %) have none of the structured source markers. A mixed F/non-F price comparison is therefore not possible here.
In this measurement, OpenAI feeds identified by the defined marker play no role in Germany, while Google Shopping data do. The actual origin of any individual web price remains open: it might come from a merchant offer, a comparison site or another source. This finding applies to the period and access method measured.
What you can take from this as a merchant
The case reviews show how easily prices can look comparable when they are not. These are practical implications for data quality, rather than demonstrated ranking factors:
- Keep product information consistent. If you supply data through several channels, the model, merchant and price should agree with your product page. The Sony case shows how conflicting fields can distort an analysis.
- Identify variants clearly. Size, gender variant, generation and included equipment determine whether offers can actually be compared.
- Make the total price, including shipping, clear. In the Cuisinart case, adding eBay’s stated shipping charge removes its apparent advantage over Home Depot’s listed item price; a complete comparison of final prices remains open. A low card price alone does not establish a price advantage.
This study does not answer whether adding a feed is worthwhile for your shop or improves your position.
Limitations
- Selection and timing: twelve curated categories, three measurement days, signed-out access in the US and Germany. This is not a representative sample of all shopping queries. Signed-in US buy boxes were not measured.
- Source markers: Missing or alternative markers limit classification. “Web” does not establish that a merchant has no feed.
- Small, dependent price sample: Repeated questions and the same products across days are dependent observations. The Wilson intervals describe the observed shares; they are not a causal test of ranking preference.
- Comparability and extraction: Sizes, equipment, gender variants, stock and final prices including shipping and taxes were not fully verified. Structured card fields may contradict the answer text.
- No market-wide price guarantee: “Cheapest” means the cheapest confirmed card within the same answer. It does not establish the lowest price in the market or a currently purchasable offer.
Sources
- Own measurements: signed-out ChatGPT, US and Germany, October 22–24, 2026; 1,008 answers with no reported collection error. Card markers, prices and order were analysed as described in the method.
- Review of all four apparent price premiums against saved answer texts, plus an additional manual sample checking model matching and source markers. A model match does not confirm an identical purchasable variant.
Related service: GEO for online shops (in German).
