For years, most people in SEO assumed ChatGPT worked the same simple way: it sends a query out to Bing, gets some results back, and generates an answer from there. Peec AI, a company dealing with both geo and AI search analytics, has released a new report stating that such assumption was wrong already. It seems ChatGPT had been quietly developing its own retrieval index internally called Labrador, and it was years since they started.
This research has come from Tomek Rudzki of Peec AI, and has put together how ChatGPT actually finds information by a combination of leaked technical signals, Openai’s job postings, together with sworn court testimony. The result which comes out from this investigation is a lot more complex and ambitious than it could be generally expected.
Labrador Is Not One Index, It’s a Family of Them
Between May 21 and July 21 2026 ChatGPT’s server-side events contained the result_source field which gave exactly the source of the information used in generating an answer. It only took four possible values: Labrador Bright Oxylabs, and SERP. Labrador is owned by OpenAI while the other three are third-party scraping providers. Researchers say Labrador isn’t only one index type. It is divided in different kinds just likeGoogle did for its search infrastructure over twenty years with the main web index plus verticals that were added on top as the separate layers. The categories found were general web PDF YouTube, news (split into last-day, last-7-days, and older versions) arXiv Wikipedia local finance (with its own PDFspecific index) legal medical, shopping, and images. All of them have the same type of index data, like page text, dates of web crawls, and publication dates.
Job Postings Point to a Serious, Long-Term Investment
Peec AI also cites various OpenAI job ads as proof of what its researchers had been saying. A job posting for a Software Engineer working on Foundations Search says that they will evaluate the candidate’s experience in building and supporting “indexing systems, retrieval pipelines, and serving layers. ” Another Engineering Manager role in the Online Data Systems team at OpenAI outlines the tasks of constructing and maintaining the main database and indexing services that support OpenAI’s apps such as ChatGPT at what the ad calls exabyte scale. An exabyte amounts to around a billion gigabytes and such a kind of infrastructure cannot be built by a company which only uses someone else’s index to pass queries through it.
There is one more listing, which focuses mainly on embedding retrieval, needs the candidate to have a proven background in creating new embedding training objectives, vector store architectures, and dynamic indexing, besides work in dense, sparse, and hybrid representation methods. Sparse representation is actually a lexical ranking technique, like TF-IDF and BM25, which Google’s original search engine was based on. Pair the sparse search with dense vector embeddings, you get a method known as Reciprocal Rank Fusion, which combines the two types of results, something Peec AI researcher Metehan Yesilyurt had noted previously in ChatGPT’s mechanisms.
The Story Traces Back to a 2024 Antitrust Hearing
Peec AI relates this story to Google’s antitrust case in which Nick Turley, the leader of ChatGPT, gave a sworn testimony that OpenAI’s search strategy was one of their main talking points. Turley was reported to have stated that OpenAI faces considerable quality issues in the use of search data from non-Google API partners, and that OpenAI was willing to partner directly with Google even during the period that Google unfortunately refused the proposal.
As the court evidence given to the media describes, OpenAI’s switch to using its own index was motivated partially by that turn-down. Turley indicated that OpenAI started to work on its own search index in 2023 with an ambitious aim at first only to solve 80% of customer queries via that index by the end of that year. Also, Turley pointed out that if OpenAI had the opportunity of using the whole Google index data, it would still take them five years at least just to find out whether answering all queries from their own index is possible. In short, becoming fully independent in data sources is a slow journey. It is a gradual step-by-step process of testing on query categories one by one.
Shopping Is Where the Index Is Most Visible Right Now
Peec AI points out that the most evident live demonstration of an ChatGPT’s indexing system occurring naturally is in product search outcomes. As the research team, a mid-August 2026 A/B test named “prefer-index-over-serp-v3” proved to be operating on approximately 8% of conversational chats. As of September 2, 2026, at least five separate shopping experiments were still running, each involving testing different combinations of OpenAI’s internal data resources versus search from outside sources.
A study in detail that could be accessed by experts integrates three major steps: a lexicon search stage using the BM25 algorithm that reduces the set of items to ten sources; a ranking step across 400 candidate products; and an approximate nearest neighbor vector search that runs at two separate dimensions. A multi-stage ranking system like this one subjected to live shopping data implies that OpenAI team regularly carries out a validation process to ensure whether the index generated by itself is ready to be used at least at a trial production level.
ChatGPT Also Appears to Cache Pages
Live retrieval means the system gets the latest version of the website whenever it is required. At ChatGPT’s level of scaling, retrieving every web page live each time is almost impossible since many pages take more than a second to load. Peec AI, which is an AI company, maintains that this is a very compelling argument for assuming that ChatGPT holds a cache of commonly visited web pages. The idea is very similar to one that Google put forward in their own patents.
The Peec AI researcher published evidence in the form of a test website consisting of 1 billion pages in particular to study what ChatGPT’s web spider was doing on these pages. By September 2026, ChatGPT had crawled around 6 million of them, which is a pretty small fraction considering the massive scale. But the spider is still at a speed of 35,000 pages per hour. Peec AI also tested one of ChatGPT’s settings, i. e. “lockdown mode, ” where you cannot surf live but can get cached or offline results. They surprisingly found out that ChatGPT could show up current-looking homepage content of some top SEO publishers even in lockdown mode, which indicates that the content probably was not freshly downloaded but rather served by a cache.
Google and Microsoft Still Play a Role
ChatGPT, though it constructs its own index, has not completely dissociated itself from external search engines for example, Peec AI’s CPO Malte Landwehr apparently did an experiment with a website that wasn’t generating any organic traffic where he interrogated ChatGPT about it for a few days through this website. As they expected from the Google Search Console they saw that on that days’ traffic of the site quite a bit increased – so they consider that a hard proof showing that ChatGPT queries Google directly to enhance its answers.
Meanwhile Bing, also appears, and more importantly during Deep Research sessions Bing is even mentioned there as a direct reference of the chatGPT’s data source at their server-side events. In addition Microsoft has their own Web IQ grounding platform that the company states delivers sub-165 millisecond latency, is said to have ChatGPT with products being powered already by the Web IQ like Microsoft Copilot and customers at the enterprise level like Nasdaq.
In total Peec AI claims that there are minimum eight additional data suppliers besides ChatGPT’s crawler that are contributing content to ChatGPT’s results and these include Bright Data Oxylabs a separate SERP-scraping channel Yelp TripAdvisor, two internal data pipelines used for business listings and Google Maps data, and Microsoft’s Web IQ.
What This Means for Website Owners
The main point of the study was clearly stating that using Bing’s visibility as a proxy for ChatGPT’s visibility is no longer safe, if it was ever completely reliable. ChatGPT is creating its own crawler and index while using signals from various outside sources; they then test different combinations with live traffic and over time shift more weight toward their own systems.
For online stores, where customers can see and compare the price, the most dramatic difference in this transition seems to show as that’s where OpenAI seems to run the most elaborate experiments. Meanwhile, for those who count business on their local presence, their listings on review sites like Yelp and TripAdvisor are directly affected and seem to matter the most since these reviews are treated as data licenses whereas Google Maps data enters ChatGPT regardless of what the businesses choose. Also, viewing how a website is presented through ChatGPT’s lockdown mode can give a preliminary idea about its visibility. Still, everything is happening at a fast pace. Peec AI itself has remarked on these facts and has been making changes on parts of this system weekly; so, some details here might be already outdated by the time you read this.
FAQs
Q 1: What is Labrador?
It is the internal name of ChatGPT’s own search index, which was developed by OpenAI to supplement its search capabilities, and serve as an alternative to relying on Google and Bing.
Q 2: Does ChatGPT rely on Google and Bing?
Yes, as revealed by Peec AI’s investigation, which discovered direct evidence of ChatGPT accessing Google and Bing, with the latter being used intensively during the Deep Research sessions.
Q 3: What data does Labrador index?
The data categories include general web pages, PDFs, YouTube videos, news, arXiv, Wikipedia, local listings, finance, legal, healthcare, shopping, and images.
Q 4: When was the date when OpenAI began building its own index?
Courthouse documents, which feature testimony from head of ChatGPT Nick Turley, confirm that the process began in 2023.
Q 5: Why did OpenAI feel the need to build its own search index?
The chatbot giant was frustrated by the quality of the search results provided by third-party platforms. Testimony revealed that they aimed to establish a direct partnership with Google, which refused the proposal.
Q 6: Where can ChatGPT’s self-hosted index be seen?
It can be seen in the shopping results, where multiple A/B tests were conducted to compare the results of external indexes with those generated by OpenAI.
Q 7: Does ChatGPT cache web pages instead of fetching them live?
Evidence from the research suggests yes, based on crawling behavior observed on a test website and results returned in ChatGPT’s lockdown mode.
Q 8: What other search data providers does ChatGPT use?
Apart from Google and Bing, there are a total of eight different data providers, including Bright Data, Oxylabs, Yelp, TripAdvisor, and Microsoft’s\Web IQ.
