The era of assuming that a strong presence in Google and Bing automatically guarantees visibility in ChatGPT is over. Compelling evidence now suggests that OpenAI has been operating its own dedicated search index, potentially for years, fundamentally altering the landscape of AI-driven discovery and rendering traditional SEO metrics insufficient. This development, hidden in plain sight within data streams, job postings, and court documents, demands a complete reevaluation of how content is discovered and surfaced in an age where conversational AI is becoming the primary gateway to information.
What Is the Labrador Index? The Evidence for ChatGPT’s Own Search Infrastructure
For a long time, the prevailing wisdom held that ChatGPT’s web results were largely sourced from Bing or Google, either directly or through data derived from their search results. This was convenient for publishers and marketers: a top ranking in traditional search engines almost certainly meant inclusion in ChatGPT’s responses. That comfortable assumption no longer holds water. An investigation by Tomek Rudzki at Peec AI, combined with data analysis from the French agency Resoneo and published by Search Engine Land, provides clear and compelling evidence for the existence of a proprietary search index operated by OpenAI, reportedly codenamed Labrador.
Between May 21 and July 21, 2026, a data field named result_source appeared in ChatGPT’s data stream. This field carried only four possible values: labrador, bright, oxylabs, and serp. The last three names refer to external providers that scrape search results from Google. Labrador, by contrast, is OpenAI’s own index. The field’s subsequent disappearance from the stream in July 2026 has only deepened the intrigue, suggesting a deliberate move to obscure the infrastructure powering the model’s answers.
The Architecture: A Network of Specialized Sub-Indexes
Labrador is not a monolithic database. Behind the single name lies a sophisticated system of multiple, distinct indexes, each tailored to a specific type of content. This architectural choice reveals a deliberate strategy to organize and retrieve information with high precision across verticals. The identified sub-indexes include:
- General Web Index
- YouTube
- News (separated by recency: last 24 hours, last seven days, and everything older)
- arXiv (preprint repository)
- Wikipedia
- Local Results
- Finance (with its own dedicated PDF index)
- Legal
- Medical
- Shopping
- Images
For each document stored within these indexes, ChatGPT captures the same core data points as Google: the full page content, the date the page was crawled, and the original publication date. This level of detail is essential for providing accurate, context-rich responses.
How Job Listings and Court Testimony Confirm the Index Build
Beyond observable data traffic, the most compelling evidence for the construction of a proprietary search index comes from OpenAI’s own hiring practices and legal testimony. A job posting for an Engineering Manager, Online Data Systems describes a team responsible for building and operating the core database and indexing services behind ChatGPT, operating across multiple regions and clouds at the exabyte scale. To put that into perspective: an exabyte is 1,000 petabytes, and a petabyte is 1,000 terabytes. This scale is not for a simple cache; it is for a massive, world-class search index.
A second job posting specifically targets software engineers for indexing systems, retrieval pipelines, and serving layers, explicitly mentioning work on dense, sparse, and hybrid retrieval methods. Sparse representations are the technical language of TF-IDF and BM25, the lexical evaluation methods on which Google was built. Combined with dense embeddings, this precisely describes the architecture required for Reciprocal Rank Fusion (RRF), a technique that Merges and ranks results from multiple sources. Metehan Yesilyurt was among the first to document this RRF process in action within ChatGPT, confirming the technical necessity of such an index.
The timeline for this effort is outlined in sworn testimony from the United States v. Google antitrust trial. Nick Turley, OpenAI’s executive responsible for ChatGPT, testified that data from non-Google search partners suffered from significant quality problems. OpenAI sought a licensing agreement with Google, but Google declined. As a direct result, OpenAI began building its own index in 2023 with the stated goal of answering 80 percent of all user queries from its own data by the end of the year. Turley also revealed a staggering figure in the same testimony: even with full access to the Google index, it would take OpenAI at least five years to determine whether 100 percent self-sufficiency is technically achievable.
How to Identify Results from the ChatGPT Index in the Wild
For SEO professionals and publishers, the immediate practical question is how to determine whether a website’s content is being sourced from the Labrador index or from a traditional search engine scrape. Several diagnostic markers have been identified through systematic testing.
Unusually Long Page Titles
Bing truncates page titles at approximately 75 characters. In an analysis of 1,200 ChatGPT responses yielding 88,000 search results across 26,900 individual pages, Resoneo found that 24 percent of titles from the Labrador index exceeded that 75-character limit. The longest title captured was 289 characters long. In contrast, titles scraped from Google results consistently end with ellipses after roughly 60 characters, and their snippets typically run to about 160 characters. The rule of thumb is simple: if you see an uncut, unusually long page title in a ChatGPT response, the result very likely came from OpenAI’s own index.
Static, Unchanging Snippets
Labrador delivers three data points per result: the URL, the full title, and a text excerpt of roughly 200 characters. Crucially, this excerpt is cut at the time of indexing and then served identically for every query that triggers that URL. If you ask the same URL about a product’s price on one query and its history on another, ChatGPT will display the same static text excerpt. Google and Bing, by contrast, have adapted their snippets to the specific user query for nearly two decades. A static snippet is a strong indicator of Labrador sourcing.
Extremely Low Overlap with Bing Top-20 Results
A practical test involves taking ten prompts from a relevant topic area, collecting the sources ChatGPT displays, and checking how many of those URLs appear in the top-20 Bing results for the same keywords. In Resoneo’s testing, the overlap was a mere 1.5 percent. Furthermore, not a single Labrador snippet matched a Bing snippet. This low correlation is a powerful diagnostic for identifying index sourcing in your own content niche.
The Lockdown Mode Test
ChatGPT offers a setting in which it performs no live fetch requests from the internet, but continues to serve content from its stored data. Peec AI tested this by querying ChatGPT about the homepages of several SEO publications. ChatGPT returned current content from those pages without making a live request. Anything that appears in this mode is stored within ChatGPT’s index or cache. Anything that is absent has not yet been ingested.
Server Log Analysis
In server logs, two distinct access patterns can be identified. One is a direct fetch from a ChatGPT user during response generation. The other is the OAI-SearchBot, which is the official crawler for the search index. Resoneo set up test pages with full logging and observed almost no OAI-SearchBot activity, even though the index continued to deliver results for those pages. The identity of the robot (or robots) that feed the Labrador index remains an open question. One experiment by Metehan Yesilyurt offers a clue to scale: he put a test site of one billion pages online. By early September 2026, ChatGPT had fetched six million of those pages, at a rate of approximately 35,000 requests per hour.
Impression Spikes in Google Search Console
SEO consultant Malte Landwehr conducted an experiment using a website with no organic traffic. On specific days, he exclusively asked ChatGPT questions about that website. On precisely those days, Google Search Console registered impressions for the domain. This indicates that ChatGPT is still querying Google for some data, even as its own index delivers the majority of results. This parallel activity can create ghost impressions in your analytics.
ChatGPT’s Caching System: How Pages Are Stored and Served
Beyond its proprietary index, OpenAI operates a separate read cache. This cache stores complete web pages that have been converted from HTML to Markdown format. The cache is keyed by URL and is shared across all users and subscription tiers. When a paying user triggers a page fetch, a free user on the other side of the world may receive a copy of that page a week later. The frequency with which a page is re-cached depends entirely on user demand; popular pages stay fresh, while neglected pages age.
Four technical details from this caching system are critical for content optimization:
- Cache-Control headers are ignored: The directives
no-storecodecodecodecodecode andnoindexcodecodecodecodecode are not respected. If content is requested and stored, it will be served regardless of these instructions. - JavaScript is not executed: Client-side rendered content does not exist for ChatGPT. Pages must be fully functional and readable without JavaScript.
- 4 MB size limit: Pages larger than 4 MB are rejected with an HTTP 400 error and are not read at all. There is no truncation; the page is simply excluded.
- Structured data loss: Scripts, iframes, and JSON-LD blocks are stripped during the HTML-to-Markdown conversion. Structured data does not reach the model through this pathway. Image alt-text is preserved, as is text hidden via CSS.
Google and Microsoft Are Still in the Loop
It is important to understand that the proprietary index does not mean OpenAI has completely severed ties with traditional search. Rudzki’s analysis identifies at least eight external providers alongside the own crawler: Bright Data for Google Web and Google Maps, Oxylabs for Google and specifically for news, a third SERP channel, Yelp, TripAdvisor, two unidentified internal pipes, and Microsoft’s Grounding platform known as Web IQ. Microsoft explicitly lists ChatGPT as a customer on the Web IQ product page, alongside its own Copilot and Nasdaq. This means that data appearing to come from Bing may have actually been sourced through Web IQ, which is a different product with a different purpose.
The proportion of results coming from the proprietary index versus scraped sources depends heavily on the user’s mode. Resoneo ran the same prompts across all ChatGPT configurations in August 2026. In the free Think mode, 74.7 percent of results came from Labrador, with only 3.1 percent from scraped Google Web results. In the paid Thinking mode, the balance flipped dramatically: 75.3 percent of results came from Google scraping, and only 24.7 percent from Labrador. Since over 90 percent of ChatGPT users are on the free tier, and the free tier defaults to Instant mode, the overwhelming majority of responses are now powered by the Labrador index.
What Does This Mean for SEO and Content Strategy?
The rise of ChatGPT’s own index and caching infrastructure requires a fundamental shift in how content visibility is measured and achieved. The old assumption that Google rank equals AI visibility is no longer valid.
Stop Using Google and Bing Visibility as Your Only Benchmark
If your goal is to be found in ChatGPT, you must actively work to get your content into the Labrador index. Evaluating your performance solely through traditional search console data will give you an incomplete and potentially misleading picture of your AI-powered audience reach.
Write Longer, More Informative Page Titles
Because ChatGPT serves titles uncut, the 55-to-60-character limit that governed SEO for a decade is now a self-imposed restriction that may be costing you. Longer titles can include more relevant keywords and provide a stronger signal. This has been true for Google as a ranking factor for some time, but the truncation display limit no longer provides an excuse to omit important terms.
Lead with Your Core Argument in the First 200 Characters
In ChatGPT’s Instant mode, the snippet of roughly 200 characters taken from after the H1 heading is often the only text the model sees from your page. If your content buries its main point under category lines, author biographies, table of contents, or date stamps, you are wasting this critical space. The core message of the page must appear directly beneath the main heading.
Respect the Technical Boundaries
Ensure your pages are under 4 MB in size. Verify that all essential content is readable without JavaScript. Place informative alt-text on images high in the DOM. Bloated templates can easily exceed the 4 MB threshold, causing the entire page to be excluded from the index.
Acknowledge That Shopping and Local Results Have Different Rules
Product results in ChatGPT are sourced from merchant feeds. Local business results come from Yelp and TripAdvisor, even if the links displayed point to Google Maps. If you sell products or operate a physical storefront, your first priority should be ensuring your data is clean and complete on those specific platforms.
Leverage Specialized Vertical Indexes
PDFs, YouTube videos, news articles, arXiv papers, and Wikipedia entries each have their own dedicated index within Labrador. A technical white paper published as a PDF, a video with a clean transcript, or a well-maintained Wikipedia entry may find its way into ChatGPT through a much easier path than a standard web page. For specialized or niche topics, these verticals represent a significant strategic opportunity.
The development of a proprietary search index by OpenAI marks a decisive turning point. The comfortable dependency on Google and Bing for AI visibility is over. For publishers, marketers, and SEO professionals, the path forward requires understanding a new technical landscape, respecting new constraints, and optimizing content for a system that thinks, caches, and retrieves information on its own terms. Those who adapt to this reality will define the next era of content discovery. Those who ignore it will find themselves invisible in the conversation.
