By Tech & SEO Industry Desk
Published: August 2026
Executive Summary: Main Facts
Recent technical reverse-engineering investigations into ChatGPT’s network traffic have revealed surprising insights regarding how OpenAI sources, indexes, and serves web content to its millions of users. According to extensive data compiled by French SEO consultancy Resoneo—building upon foundational research originally published by SEO expert Suganthan Mohanadasan—OpenAI’s proprietary in-house search index treats content creators without formal licensing agreements almost identically to those with multi-million-dollar partnerships.
For months, the digital publishing ecosystem has operated under the assumption that high-profile content deals between OpenAI and major legacy media organizations (such as Reuters, The Guardian, and The Wall Street Journal) granted those partners exclusive visibility or privileged treatment within ChatGPT’s native retrieval systems. However, network-level data captured from July operations tells a different story.
The core findings indicate that OpenAI’s in-house search index—internally tagged as "labrador"—handles the vast majority of search results for free-tier ChatGPT users. Crucially, pages belonging to independent publishers, niche blogs, and small businesses with zero commercial ties to OpenAI are surfaced through this index using the exact same formatting, length, and freshness standards applied to licensed partner domains.
While OpenAI continues to ink lucrative data-sharing and syndication agreements with global media giants, these revelations force a re-evaluation of what those partnerships actually buy. If non-partner sites are already being served through OpenAI’s primary proprietary index on free accounts, publishers must ask hard questions about the true value proposition of these content deals.
The Investigation Timeline and Chronology
To understand how this narrative shifted from an assumption of walled-garden exclusivity to one of open indexing, it is necessary to examine the chronological progression of the research conducted by Mohanadasan and Resoneo.
June 2025–July 2025: The Initial "Allowlist" Hypothesis
The mystery of how ChatGPT selects its sources began to unravel in mid-2025 when SEO strategist Suganthan Mohanadasan decided to bypass standard output observation and instead read the raw network traffic flowing between ChatGPT and its servers.
Upon inspecting the pipeline tags attached to incoming search responses, Mohanadasan identified the "labrador" pipeline. In his initial June publication, he observed that this proprietary index appeared heavily weighted toward established, marquee publishers. Because domains like Reuters, The Guardian, The Wall Street Journal, and Wikipedia dominated the pipeline outputs in his test environment, Mohanadasan concluded that "labrador" operated as a strict, curated allowlist—effectively a closed, licensed tier reserved exclusively for corporate partners.
Mid-July 2025: The Italian Reader Correction
The narrative shifted dramatically on July 14, when Mohanadasan formally retracted his initial assertion. A reader from Italy, utilizing a free ChatGPT account, reached out to Mohanadasan with data captures demonstrating that small, independent Italian publishing sites—which unquestionably lacked any commercial relationship with OpenAI—were being routed through the exact same "labrador" pipeline.
Prompted by this counter-evidence, Mohanadasan re-ran his tests across broader parameters. In a follow-up blog post, he openly acknowledged that he had "over-reached" with his tier claim. He clarified that while OpenAI’s licensing deals with major media companies are entirely genuine, his initial deduction of an exclusive tier was flawed because it relied too heavily on the perspective of a single account configuration.
Late July 2025: Resoneo’s Deep Dive and Traffic Obfuscation
Seizing upon Mohanadasan’s foundational discoveries, the French SEO consultancy Resoneo launched a much larger, multi-faceted empirical study. Utilizing a specialized Chrome extension distributed to users, Resoneo analyzed 1,249 ChatGPT answers captured throughout July across a diverse matrix of free accounts, paid accounts, logged-out sessions, and international geographies.
Resoneo’s rigorous cross-account testing confirmed that non-partner sites were routinely served by the "labrador" index with identical formats and freshness metrics as licensed content.
However, this window into OpenAI’s architecture was short-lived. According to Resoneo, around July 21—shortly after the research gained traction within the SEO community—OpenAI seemingly adjusted its backend telemetry, abruptly stopping the practice of tagging each search result with the specific name of the fetching pipeline. This effectively closed the door on easy, real-time network traffic auditing.
Supporting Data: Dissecting the "Labrador" Pipeline and Paid vs. Free Dynamics
Resoneo’s data provides a granular look at how OpenAI balances its proprietary index against external scraping mechanisms, particularly when comparing free-tier users to paid subscribers using advanced reasoning ("thinking") modes.
Free Accounts vs. Paid Accounts: A Tale of Two Pipelines
The research revealed a stark operational divergence depending on whether the user was operating a free or a paid ChatGPT account:
- Free-Tier Accounts: For users on free plans, the "labrador" in-house index handles the lion’s share of search results. When queries involve settled factual answers, local business lookouts, and product recommendations, this proprietary index is deployed almost every single time. Furthermore, news results on free accounts were found to be split fairly evenly between OpenAI’s internal index and traditional Google scraping methods.
- Paid Accounts (Thinking Mode): The dynamic shifts drastically when analyzing heavy, multi-step reasoning queries on paid accounts. In a sample of 16,407 search results recorded by Resoneo during thinking-mode sessions, external Google scraping accounted for approximately 75% of the retrieved sources. Meanwhile, OpenAI’s in-house "labrador" index made up the remaining 24%.
What is the "Labrador" Index?
Resoneo describes "labrador" as a self-contained, vertically integrated search index that OpenAI actively tops up using direct press feeds and open science archives. The primary architectural advantage of this system is that OpenAI can ingest, process, and retrieve this data directly without relying on or paying third-party search APIs or intermediaries.
Under the Hood: What OpenAI’s Index Sees of Your Page
Beyond pipeline distribution, Resoneo conducted a structural audit of 534 specific web pages cited by ChatGPT, comparing the live pages against the precise text snippets stored within OpenAI’s internal index:
- H1 Headings: Out of 463 pages featuring explicit H1 markup, a remarkable 83.6% (387 snippets) successfully included the H1 text within the stored snippet.
- The 200-Character Limit: OpenAI’s index typically cuts off snippets right after the 200-character mark. Crucially, these snippets are generally pulled from the very beginning of the page body content rather than from the page’s meta description (a practice that Google’s scraping pipeline still relies on roughly one out of three times).
- Template Pollution: Resoneo’s data revealed that structural page templates consume valuable real estate within that strict 200-character budget. For instance:
- Section Kickers: Appear before the H1 on 29% of pages, consuming about 18 characters.
- Publication Dates: Present on 11% of pages, taking up roughly 25 characters.
- Image Alt Text: Found in the first image on 9% of pages, occasionally devouring up to 50 characters on its own.
- Missing H1s: Roughly one out of every seven pages lacked H1 markup entirely; in those instances, the snippet defaults to whatever subheading the site template provides first.
Official Responses and Industry Context
As of the publication of these findings, OpenAI has maintained a guarded stance regarding the exact inner workings of its proprietary retrieval algorithms.
A review of OpenAI’s official developer documentation and bot crawler guidelines reveals minimal technical transparency regarding the "labrador" index, its inclusion criteria, or the explicit mechanics of its publisher agreements. While OpenAI publicly promotes its multi-million-dollar partnerships with elite news organizations—framing them as collaborative efforts to advance AI journalism and ensure accurate real-time reporting—the company has not explicitly detailed whether its crawler treats partner sites differently at the ingestion phase compared to non-partner sites.
Industry analysts note a vital technical distinction highlighted by Resoneo: while partner articles often reach OpenAI’s ecosystem via direct, high-speed programmatic feeds rather than traditional web crawling, the final indexing and serving layer ("labrador") appears agnostic to whether the content arrived via a paid licensing feed or standard web discovery.
Implications for Publishers, SEOs, and the AI Ecosystem
The revelation that non-partner websites enjoy parity within OpenAI’s primary free-tier search index carries profound strategic implications for digital publishers, content creators, and the broader search engine optimization industry.
1. Reassessing the Value of OpenAI Content Deals
For months, independent publishers have felt pressured to sign data-sharing agreements with AI giants out of fear that refusing to do so would result in digital oblivion—complete exclusion from AI-generated answers. Resoneo’s findings complicate this narrative.
If sites without a paid content deal are already fully integrated into the "labrador" index—the system powering the majority of free-tier ChatGPT search results—then mere visibility in AI answers is no longer an exclusive perk locked behind a corporate paywall. Publishers must now evaluate content deals not on the false premise of visibility, but on other commercial metrics, such as direct financial compensation, indemnification, API access, or preferential inclusion in specialized multimodal pipelines.
2. The Mechanics of AI-Driven SEO (GEO)
The technical details uncovered by Resoneo provide actionable, albeit sobering, insights for site owners looking to optimize their content for Large Language Model retrieval systems.
Because OpenAI’s index relies heavily on the first 200 characters of body content and routinely ingests H1 tags while bypassing meta descriptions, traditional on-page SEO playbooks must evolve. Site architects must carefully audit their HTML templates to ensure that boilerplate metadata, category kickers, and promotional banners do not cannibalize the precious 200-character window that the AI uses to parse and summarize page context. Front-loading core semantic value immediately after the H1 tag may soon become a cornerstone of Generative Engine Optimization (GEO).
3. The Methodology Shift: Multi-Account Auditing
Perhaps the most immediate practical takeaway for the SEO community is methodological. Mohanadasan’s initial error—mistaking a single account’s traffic sample for a universal platform rule—serves as a cautionary tale for researchers analyzing opaque AI systems. As algorithms become increasingly dynamic, personalized, and geographically fragmented, relying on a single data point or account profile is dangerously misleading. Rigorous future research demands multi-account, cross-geographical, and cross-tier testing methodologies, as successfully demonstrated by Resoneo.
Conclusion
OpenAI’s search infrastructure remains a rapidly shifting landscape. While the corporate push toward exclusive media partnerships continues to dominate industry headlines, technical reality proves that the underlying technology is far more democratic than previously assumed. As AI search continues to siphon traffic away from traditional search engines, understanding the granular mechanics of indices like "labrador" will be essential for any publisher striving to remain visible in the age of conversational AI.