The Evolution of AI Search and the Shopping Data Mystery
For the past year, the tech world has watched closely as OpenAI has attempted to pivot ChatGPT from a conversational chatbot into a full-fledged search engine. With the rollout of integrated search features, the question on every SEO professional and digital marketer’s mind has been: where is this data coming from? While OpenAI’s historical partnership with Microsoft suggested a heavy reliance on Bing, recent technical investigations have uncovered a surprising reality. It appears that when it comes to e-commerce, ChatGPT is looking toward Mountain View, not Redmond.
A comprehensive new study has revealed that ChatGPT sources a staggering 83% of its carousel products directly from Google Shopping. This discovery was made by analyzing “query fan-outs” (QFOs), the internal search queries ChatGPT generates to fetch real-time data. The findings suggest that despite OpenAI’s move toward independence, the platform has developed a significant, perhaps even systemic, reliance on Google’s product index to power its shopping recommendations.
Understanding the Technical Link: The id_to_token_map Discovery
The investigation into ChatGPT’s sourcing began in late 2025, when AI researchers identified a specific field within ChatGPT’s source code labeled id_to_token_map. While the field initially appeared to be a string of gibberish, it was actually base64 encoded. Once decoded, the data revealed a treasure trove of parameters that are synonymous with the Google Shopping ecosystem.
Researchers found specific identifiers such as productid and offerid, alongside locale and language parameters. Most tellingly, the decoded field contained the exact query used to trigger the product lookup. By extracting these parameters, researchers were able to reconstruct full Google Shopping URLs. When these URLs were tested, they led directly to the same products displayed within the ChatGPT interface.
This technical “smoking gun” proved that ChatGPT isn’t just “finding” products on the web through general crawling; it is actively querying Google’s structured shopping data to populate its interactive carousels. This raises vital questions about the architecture of AI search and how much of the “AI answer” is simply a re-ranking of existing search engine results.
What Are Shopping Query Fan-Outs?
To understand how ChatGPT retrieves information, we have to look at “query fan-outs.” When a user types a prompt like “best budget mechanical keyboards,” ChatGPT doesn’t just look at its training data. Instead, it “fans out” the request into multiple secondary search queries to find current web results. The study categorized these into two types: normal search fan-outs and shopping query fan-outs (QFOs).
The data shows that these two processes are fundamentally different and operate on separate tracks. After analyzing 1.1 million shopping QFOs, researchers found that shopping fan-outs are unique to the user prompt 99.7% of the time. More importantly, they are distinct from the general search fan-outs 98.3% of the time. This suggests that ChatGPT knows when a user is in a “buying” mindset and switches to a specific retrieval pipeline designed for products.
The Differences in Query Structure
The study found a clear divergence in how these queries are constructed:
- Search Fan-Outs: These queries average 12 words in length. They are designed to be descriptive and contextual, aiming to retrieve web pages, articles, and reviews that can be used to synthesize a written response.
- Shopping Fan-Outs: These queries are much shorter, averaging only seven words. Their primary goal is to hit a specific shopping index and return a list of products. They act more like a traditional search bar entry than a conversational prompt.
Furthermore, the frequency of these queries differs. On average, a single user prompt triggers 2.4 search fan-outs but only 1.16 shopping fan-outs. This indicates that while ChatGPT needs multiple sources to write a detailed answer, it only needs a single, efficient query to Google Shopping to fill a product carousel with eight items.
The Data Breakdown: Google Shopping vs. Bing Shopping
To quantify the extent of this reliance, the study compared 43,000 products found in ChatGPT carousels against 200,000 organic shopping results from both Google and Bing. The methodology involved choosing diverse prompts across 10 industry verticals and using a sophisticated matching algorithm to identify product overlaps.
The Google Dominance
The results were conclusive. Approximately 45.8% of ChatGPT carousel products had an exact title match within the top 40 organic results of Google Shopping. When the criteria were expanded to “strong matches” (products that are clearly the same brand and model but may have slight title variations), the number jumped to over 83%.
The Bing Discrepancy
In contrast, Bing’s influence on the shopping carousel was almost non-existent. Only 0.48% of products were an exact match for Bing’s top 40 results. While 11% of products showed some level of similarity to Bing results, nearly all of those products were also found on Google. In fact, across the entire dataset of 43,000 products, only 70 items (a negligible 0.16%) were found exclusively on Bing. This proves that ChatGPT is essentially ignoring Bing Shopping in favor of Google’s more robust index.
The Impact of Positional Bias
For retailers and e-commerce managers, one of the most critical findings of this study is the correlation between Google Shopping rank and ChatGPT carousel placement. The study found a clear “sloping trendline,” meaning that products ranking higher on Google are significantly more likely to appear—and appear earlier—in ChatGPT.
Key statistics regarding positioning include:
- The Top 10 Rule: 60% of the strong product matches in ChatGPT come from the top 10 results in Google Shopping.
- The Top 20 Rule: Nearly 84% of matches come from the top 20 Google Shopping results.
- Carousel Ranking: The first position in a ChatGPT carousel typically corresponds to a product found in the top 5 of Google Shopping organic results.
This suggests that ChatGPT is not just sourcing from Google; it is largely trusting Google’s existing ranking algorithm to determine which products are most relevant to the user. If you are not ranking on the first page of Google Shopping, your chances of appearing in a ChatGPT product recommendation are statistically slim.
Does Prompt Branding Change the Results?
One might wonder if branded queries (e.g., “Best Nike running shoes”) behave differently than non-branded queries (e.g., “Best running shoes for marathons”). The study explored this and found that the sourcing behavior remains remarkably consistent. Both branded and non-branded prompts showed a high level of reliance on Google Shopping, with non-branded prompts actually showing a slightly higher match rate. This indicates that the reliance on Google is a systemic architectural choice by OpenAI, not an artifact of specific types of queries.
The Methodology Behind the Study
To ensure the findings were robust and not limited to a single niche, the study utilized Peec AI data across 10 distinct industry verticals:
- Apparel & Footwear
- Baby & Kids
- Beauty & Personal Care
- Electronics
- Home Improvement
- Home & Kitchen
- Office Supplies
- Pet Supplies
- Sports & Outdoors
- Toys & Games
The Three-Stage Matching Algorithm
Because product titles can vary between platforms (due to different character limits or SEO optimizations), the researchers used a conservative three-stage cascade approach to identify matches:
- Stage 1: Exact Match: Case-insensitive string equality. This captured nearly half of the results.
- Stage 2: Near-Exact Match: Used a SequenceMatcher ratio to catch minor differences in punctuation, spacing, or unicode characters (e.g., different types of dashes). A threshold of 0.95 was required.
- Stage 3: Hybrid Match: A weighted average combining character-level similarity (40%) and token/word overlap (60%). This was used to identify products where the brand and name were identical but the word order or descriptors varied slightly.
The researchers set a “good match” threshold at 0.8. This ensured that only products of the same brand and model were counted, while excluding “similar” products that were actually different models or competitors.
Why Does ChatGPT Rely on Google?
The shift away from Bing for shopping data is a significant development in the AI landscape. While OpenAI has not officially commented on the specific data sources for its shopping carousels, several logical reasons exist for this architectural choice:
1. Index Depth and Accuracy
Google Shopping remains the most comprehensive product index in the world. For an AI that prides itself on accuracy and relevance, Google’s data is often more up-to-date regarding pricing, availability, and product variations than Bing’s index.
2. Integration Ease
The discovery of Google-specific parameters (productid, offerid) suggests that OpenAI may be using specific APIs or structured data formats that are natively compatible with Google’s Merchant Center output. This allows for a more “plug-and-play” integration for carousels.
3. User Intent Matching
Google has decades of experience matching vague search queries to specific product intent. By leveraging Google’s organic shopping results, ChatGPT can provide high-quality recommendations without having to build its own e-commerce ranking engine from scratch.
Strategic Implications for SEO and E-commerce
This finding completely changes the playbook for “AI Engine Optimization” (AEO) in the e-commerce sector. Previously, marketers were unsure whether they should focus on Bing SEO, schema markup, or brand mentions to get into ChatGPT.
The answer is now much clearer: Google Shopping optimization is the primary lever for ChatGPT visibility.
How to Optimize for ChatGPT Carousels
- Focus on Organic Google Shopping: Since ChatGPT sources from the top 40 organic results, and heavily favors the top 10, your primary goal should be improving your organic rank within Google’s shopping tab.
- Maintain High Feed Quality: Ensure your Google Merchant Center feed is impeccable. Use clear, high-quality images and standardized titles, as these are the elements ChatGPT pulls directly into its interface.
- Monitor Product Sentiment: While Google Shopping provides the “selection set,” ChatGPT likely uses search fan-outs to check sentiment. If your product ranks #1 on Google but has negative reviews across the web, ChatGPT might choose the #2 or #3 option instead.
- Brand Authority: The study notes that while Google Shopping rank is the strongest predictor, overall product mentions in “context sources” still matter. Building a strong brand presence through traditional PR and content marketing will help ensure that when ChatGPT “checks” your product, it finds positive context.
The Future of AI Search Relationships
The fact that ChatGPT is sourcing 83% of its products from Google Shopping highlights a strange irony in the tech world. Microsoft has invested billions into OpenAI to challenge Google’s search dominance, yet the flagship product of that partnership is currently relying on Google’s infrastructure to provide value to its users.
This behavior has been consistent for at least four months, suggesting it is not a temporary test but a core part of the current ChatGPT Search architecture. However, the AI landscape is famously volatile. OpenAI could transition to its own index or move back toward Bing at any time.
For now, the data is undeniable. If you want your products to show up when a user asks ChatGPT for recommendations, your journey starts—and largely ends—with Google Shopping. Understanding this retrieval pipeline is no longer optional; it is a fundamental requirement for anyone looking to compete in the new era of generative search.
Summary of Key Findings
- 83% Overlap: The vast majority of ChatGPT carousel products match Google Shopping’s top 40 results.
- Bing is Negligible: Less than 1% of products are exclusive to Bing Shopping; Bing’s overall match rate is only 11%.
- Ranking Matters: 60% of ChatGPT products come from the top 10 results on Google.
- Unique Pipelines: ChatGPT uses separate, shorter queries for shopping than it does for general web search.
- Universal Behavior: This sourcing pattern is consistent across all major retail categories and both branded and non-branded searches.
As AI search continues to evolve, the line between traditional search engines and LLMs continues to blur. ChatGPT is not replacing Google Shopping; rather, it is acting as a new, conversational layer on top of it. For brands, this means that the traditional pillars of SEO and feed management are more important than ever—they are simply being viewed through a different lens.