Has OpenAI’s increasing independence from Microsoft and, by extension, Bing, become an overly dependent relationship with Google?

Our study comparing shopping query fan-outs (QFOs) in ChatGPT from both Google and Bing carousels appears to have provided at least a partial answer to that question. Let’s take a look at how this study was conceived and what we found.

Brief shopping fan-out background and technical explainer

In November 2025, a few researchers in the AI research space, including myself, detected a mysterious field in ChatGPT’s source code: id_to_token_map. But what that field revealed when decoded was even more intriguing.

This field is what’s called base64 encoded, but when we decoded it, it revealed what looked to be Google Shopping parameters, such as productid, and offerid, but also language/locale parameters. Even more interesting? This field revealed a query used to look up that particular product. 

To categorically prove this was indeed a Google Shopping link, we would have to be able to reconstruct the shopping URL solely from the extracted parameters. 

Let’s look at an example of what this looks like using the ChatGPT product carousel for the prompt “best smartphones under $500.”

If we decode the relevant field, we can recreate the Google Shopping link from the extracted parameters.

The big question was: Would this link correspond to the exact product in the ChatGPT product carousel? So we tried it:

It turns out that, in fact, yes it does!

But this decoding technique alone doesn’t answer any of these important questions:

• Is this retrieval process uniform across diverse product categories?
• Does ChatGPT select from a certain number of Google product positions?
• Does ChatGPT favor higher Google Shopping product positions?
• How common is this process at scale?
• Was this just a fluke or, given a large enough dataset, could we match these products with any online retailer or even Bing Shopping results?

Using Peec AI data, the following study aimed to robustly prove once and for all that ChatGPT does indeed mainly source from Google Shopping. 

To do this we analyzed more than 40,000 carousel products and 200,000 organic products from each Google and Bing. By comparing the similarity of the products, we got a very clear picture of what was really happening behind the scenes. Let’s dig into our findings.

Are shopping query fan-outs really that different from normal search query fan-outs?

To answer whether shopping query fan-outs are different from normal search query fan-outs, we analyzed 1.1M shopping query fan-outs from Peec AI data and compared them to the normal search query fan-outs for the same user prompt. We found that they are almost always different:

Shopping QFO unique to user prompt	99.70%
Shopping QFO unique to normal query search fan-out	98.31%

To dive deeper, we explored the average word counts of both of these query fan-out types by calendar week. 

The chart below clearly shows that normal fan-outs are significantly longer — 12 vs. seven words. That makes sense since search query fan-outs are used to retrieve contextual information. This means they need to be long enough to retrieve web results that are specific to the user prompt. Vector search (or comparing embeddings) works best with more context. 

Shopping fan-outs, on the other hand, typically target a specific shopping results page and therefore do not need to be as long. It appears the main goal is to retrieve products based on the shopping fan-out. Rather than compare chunks of text, the data in this study supports the hypothesis that ChatGPT relies heavily on Google organic shopping results to populate its carousel.

Further evidence of the distinct nature of the shopping fan-outs surfaces when we look at how many are used per prompt. On average, 2.4 search fan-outs are used per prompt vs. just 1.16 for shopping fan-outs. For reasons similar to above, retrieving more contextual information often requires more search fan-outs vs. simply retrieving products. To populate an eight product carousel in ChatGPT, it seems that, for the most part, one page of Google Shopping results is enough.

How similar are ChatGPT Carousel products to Google Shopping products?

To answer this question in the fairest possible way, we extracted around 5,000 ChatGPT carousels comprising 43,000 products from the Peec AI dataset. Prompts were chosen to be as diverse as possible (see Methodology for the creation process).

We then extracted the organic shopping pages and retrieved the top 40 organic products for both Google and Bing shopping results. Paid ads and sponsored products were excluded from the analysis. 

We used a three-step matching algorithm (see Methodology for exact details) to attain a similarity score between the ChatGPT product title and the title found in organic shopping results. This is because not only is ChatGPT probabilistic, but so is, to a certain extent, Google Shopping. Product titles can be rewritten with or without certain product features and results are very sensitive to the exact proxy location where the results are retrieved. 

We counted a product as matching if it reached a threshold of 0.8 or above, effectively, if it was the same brand and product name and exhibited a very high degree of similarity.

The results are summarized in the chart below.

Impressively, across 43,000 highly diverse ChatGPT carousel products, 45.8% were found to have an exact title match in the corresponding Google top 40 organic shopping products for that exact shopping fan-out. 

For Bing, this exact match rate was just 0.48%. 

If we simply look at the percentage of strong product matches across all eight ChatGPT carousel positions, over 83% were found in the Google top 40 products, but that number drops to just under 11% for products found on Bing. This is very strong evidence that ChatGPT sources its carousel products from organic Google Shopping results.

We also see a very high number of weak matches in Bing at over 62%. This implies that the top 40 returned products for each shopping fan-out differ significantly across Google and Bing. This makes sense as there are many 1000s of possible combinations of brand and product that can be surfaced in shopping results. 

Even if Bing found around 11% of ChatGPT carousel products, how many of those products were only found by Bing? Across the 43,000 carousel products Bing only found 70 that were not found in Google Shopping, constituting just 0.16%. This means that in almost every case there was a match in Bing there was also a match in Google. 

It seems unlikely, then, that ChatGPT is also sourcing products from Bing Shopping in the vast majority of cases.

How does the ChatGPT carousel position affect the match rate?

Here we explore the most common positions (mean and median shown) of Google shopping product positions for each ChatGPT carousel position:

For example, for the first carousel position we can see that the average Google Shopping position is around five. Note that we see a sloping trendline for the carousel positions that correspond to higher Google Shopping positions. This implies that ChatGPT sources top carousel products from higher Google Shopping positions. 

Plotted another way, we can visualize the cumulative number of strong matches across organic Google Shopping positions. This chart allows us to see that 60% of the strong product matches are found in the top 10 Google shopping results alone. 

Comparing the top 20 vs. positions 21-40, ChatGPT’s favoritism for higher positions becomes clear, with an overwhelming majority of matches (almost 84%) coming from the top 20:

Finally, we explored whether the prompt being branded vs. non-branded made a difference to the product matching results.

The results show a similar high level of product matching for both branded and non-branded prompts, with only slightly higher match rates for non-branded:

Summary of findings

This study analyzed over 43,000 ChatGPT carousel products across 10 industry verticals and compared them against 200,000+ organic shopping results from both Google and Bing. The findings painted a clear picture.

ChatGPT sources its carousel products from Google Shopping, not Bing 

Over 83% of ChatGPT carousel products were found as strong matches in Google’s top 40 organic shopping results. For Bing, that figure was just 11%, and of those, only 70 products across the entire dataset (0.16%) were found exclusively in Bing. In almost every case where Bing returned a match, Google had already returned the same product.

Product retrieval and contextual retrieval are separate processes 

The data strongly supports this. Shopping query fan-outs are distinct from normal search fan-outs 98.3% of the time. They are significantly shorter (seven vs. 12 words), and ChatGPT uses far fewer of them per prompt (1.16 vs. 2.4 words). This makes sense; populating a product carousel is a fundamentally different task from gathering contextual information to construct a written answer. One is about retrieving structured product listings from a shopping index while the other is meant to retrieve web pages rich enough in context for vector search and re-ranking to work effectively.

ChatGPT favors higher Google Shopping positions 

The data shows a clear positional bias, with 60% of strong matches coming from the top 10 Google Shopping results and nearly 84% from the top 20. ChatGPT carousel position correlates with Google Shopping rank, meaning products that rank higher in Google Shopping are more likely to appear earlier in the ChatGPT carousel.

This points to systemic architectural behavior

Since these patterns hold across branded and non-branded prompts, and across all 10 verticals tested, this reinforces that this is a systematic architectural behavior rather than a category-specific or query-specific artifact.

What this means

For brands and retailers, the implication is straightforward: Your Google Shopping ranking strongly influences whether your products make it into ChatGPT’s carousel. These findings indicate that the selection set of carousel products in many cases is effectively the top 40 organic Google Shopping positions for the corresponding shopping fan-out query.

But while product ranking in Google Shopping plays a role, it doesn’t tell the full story. It is likely that other factors, such as overall product mentions and sentiment in the context sources retrieved, also factor into the final ChatGPT carousel selection and ranking. 

Understanding the full picture in terms of how your products are perceived across relevant sources, as well as how you show up on Google Shopping, could be the key to understanding ChatGPT product carousels.

For the AI research community, this study provides robust, large-scale evidence that ChatGPT’s product carousel operates as an independent retrieval pipeline for the selection set of products, separate from the contextual web search that powers the written portion of its responses. It is possible, and even likely, that for the final selection and ranking of products, ChatGPT uses contextual clues such as product sentiment from the sources retrieved by the normal search fan-outs.

As always, this represents a snapshot of current behavior. OpenAI could change its retrieval sources or methods at any time, but this behavior has been consistent in our findings for at least the last four months. 

Methodology

Objective

Measure how much product overlap there is between ChatGPT Shopping (via product carousels) and Google Shopping organic results for the same queries, across 10 industry verticals. This was contrasted to Bing shopping results as a control using an identical pipeline.

Specifically, the study evaluated:

• How often ChatGPT recommends products that also appear in Google Shopping results
• Where those overlapping products rank in each system

PromptSet creation

Prompts were created with the purpose of triggering ChatGPT carousels. To maximize diversity, a mixture of branded and non-branded prompts were used, as well as prompts that explicitly included a price and ones that did not.

Additionally, a diverse selection of verticals were chosen to make the findings more robust. These were: Apparel & Footwear, Baby & Kids, Beauty & Personal Care, Electronics, Home Improvement, Home & Kitchen, Office Supplies, Pet Supplies, Sports & Outdoors, Toys & Games.

Product matching 

The product matching algorithm compared ChatGPT product titles against the top 40 Google Shopping titles using a three-stage cascade approach

The goal was to find the best match between a ChatGPT product title and the corresponding Google Shopping titles. A match was determined using a cascade of three stages:

• Stage 1: Exact match
  • Method: Case-insensitive string equality after removing whitespace
  • Score: 1.0
  • Label: exact
• Stage 2: Near-exact match
  • Method: Uses the Python SequenceMatcher ratio on lowercased strings
  • Trigger: Activated if the best ratio across all candidates is 0.95 or higher
  • Purpose: To catch minor, trivial differences like spacing, punctuation, or different types of dashes
  • Score: The SequenceMatcher ratio (rounded to three decimal places)
  • Label: near-exact
• Stage 3: Hybrid match
  • Method: A weighted average combining character-level similarity and token (word) overlap
  • Components and Weights:
    • SequenceMatcher Ratio (Character Similarity): 40% weight.
    • Token Overlap (Word Inclusion): 60% weight (fraction of tokens in the shorter title found in the longer one)
  • Selection: The candidate with the highest hybrid score is chosen, regardless of a specific threshold
  • Score: Calculated as (0.4 * SequenceMatcher Ratio) + (0.6 * Token Overlap) (rounded to 3 decimal places)
  • Label: hybrid

This approach was set to be fairly conservative, and 0.8 was determined as a reasonable threshold for a product match as this often corresponds very closely to the same brand and product. 

Real examples of matching thresholds from the data:

Match threshold	Description	ChatGPT product	Google Shopping	Differences observed
1.0	Exact string match, no differences	Hot Wheels RC 1:64 Mustang GTD	Hot Wheels RC 1:64 Mustang GTD	None
0.95	Near exact, minor differences such as hyphen, punctuation only	Learning Resources Snap-n-Learn Matching Dinos	Learning Resources Snap‑n‑Learn Matching Dinos	The hyphen character is different in unicode
0.9	Same brand and product, additional non-crucial words allowed	Block Tech 250 Piece Set	Block Tech 250 Piece Building Blocks Set	“Building” added to blocks, but product and brand are the same
.85	Same product and brand, potentially slightly different word order and additional, non-crucial words	LEGO Japanese Red Maple Bonsai Tree	Japanese Red Maple Bonsai Tree LEGO Botanicals	Different word order and one additional word “Botanicals,” same product and brand
.8 good match threshold Same brand, same product	Same brand and product, possibly additional descriptors	Cards Game Against FRIENDS – Limited Edition	Cards Game Against FRIENDS – Limited Edition – Party Card Games For Adults	Same brand and product with additional descriptors that don’t affect the match
.75	Same brand and product line, very minor product differences such as size or dimensions	My Sweet Love 14-inch My Cuddly Baby Doll	My Sweet Love 8-Inch MinWeBaby Doll	Same brand and product line but different size dimension
.7	Same brand, often slightly different product, but within same category	Adventure Force Ram Truck RC Car	Adventure Force McLaren 765LT RC Car	Same brand and product category but different individual product
.65	Same brand, often slightly different product but within same category	Mattel 300‑Piece Puzzle	Mattel 80th Anniversary Puzzle	Same brand and product category but different individual product
.6	Typically same product category, but often different brand and product line	Tell Me Without Telling Me Party Card Game	Elimino! Card Game	Different brand and product line, the same overall category of “card game”
.55	Similar product category but usually not either different brand and/or different product	Furby Interactive Plush Toy Interactive Digital Pet Toy	Interactive Digital Pet Toy	Different brand, similar product category but different specific product