For the last two years, one of the most persistent assumptions in AI search has been surprisingly simple:
If you want to rank in ChatGPT, rank in Bing.
There was a good reason for that assumption. When OpenAI launched ChatGPT Search in October 2024, it explicitly said the product used third-party search providers alongside content supplied directly by publishing partners. (OpenAI)
But that explanation is increasingly incomplete.
Recent research from a GEO tool provider, reporting by Search Engine Watch and Search Engine Land, OpenAI’s own documentation, OpenAI job postings, and somewhat unexpectedly … testimony from the Google antitrust case all point in the same direction:
OpenAI is building a substantial search index of its own.
And this isn’t merely an experimental crawler sitting somewhere in the corner of OpenAI’s infrastructure.
OpenAI now explicitly documents that some versions of ChatGPT can search “OpenAI’s indexed and cached web content.” OpenAI Help Center
That’s a pretty important sentence.
Because once ChatGPT has its own index, Generative Engine Optimization stops being simply an extension of Google or Bing SEO.
It becomes its own distribution problem.
The mysterious Labrador
Between May 21 and July 21, 2026, researchers inspecting the server-side events sent from ChatGPT to the browser repeatedly found a field called result_source.
Four values appeared:
Labrador, Bright, Oxylabs and SERP.
The GEO tool provider’s interpretation is that Bright and Oxylabs represent external search-data providers, while SERP represents another external search-results pipeline. Labrador, by contrast, appears to be OpenAI’s internal retrieval infrastructure.
And Labrador apparently isn’t just one giant bucket of URLs.
The researcher found references to separate retrieval indexes for:
- the general web
- PDFs
- news
- YouTube
- Wikipedia
- arXiv
- local information
- finance
- legal information
- medical content
- shopping
- images
That starts to look considerably less like “ChatGPT occasionally browses the web” and considerably more like the architecture of a search engine.
There is an important caveat here: OpenAI has not publicly confirmed that Labrador is the official name of its index. Search Engine Watch makes the same distinction. The existence of OpenAI-indexed content is confirmed; the reverse-engineered details of Labrador remain independent research. Search Engine Watch
But then things get more interesting.
OpenAI basically confirmed the important part itself
You don’t actually need leaked parameters or reverse engineering to establish that OpenAI is building search infrastructure.
Look at OpenAI’s current job openings.
A role for Software Engineer, Foundations Search says the team develops systems that “transform and index vast corpora” and explicitly describes work on indexing systems, retrieval pipelines and serving layers, supporting dense, sparse and hybrid retrieval. OpenAI
Another OpenAI position, Engineering Manager, Online Data Systems, says the team operates the core database and indexing services used by OpenAI’s production applications, including ChatGPT.
The job description goes further, describing:
“our in-house database and indexing technology”
and infrastructure operating at multi-region, multi-cloud, exabyte scale. OpenAI
A third search research position discusses scalable vector stores, dynamic indexing and a mixture of dense, sparse and hybrid retrieval techniques. OpenAI
That’s not the hiring plan of a company whose search strategy consists of sending a query to Bing and waiting for the answer.
But the strongest evidence actually appeared in court.
Google accidentally gave us one of the best sources on OpenAI Search
During the remedies phase of the U.S. government’s antitrust case against Google, OpenAI’s Nick Turley testified about OpenAI’s search ambitions.
The court documents say OpenAI began grounding ChatGPT using third-party search results in 2023. In 2024, OpenAI attempted to obtain access to Google’s search API, but Google declined.
The documents also state that OpenAI had experienced “significant quality issues” with non-Google search API partners.
That dependency seems to have become strategically uncomfortable.
In another section of the government’s findings, Turley’s testimony is summarized this way:
OpenAI views building its own search index as essential to ensuring the accuracy of its generative AI products. Department of Justice
And OpenAI’s ambitions were substantial.
Turley described an 80/20 strategy: ideally OpenAI could eventually cover roughly 80% of its query space itself, while the difficult long tail could take much longer.
He also acknowledged just how difficult complete independence would be. Even with access to Google’s data, Turley said it could take OpenAI at least five years simply to determine whether answering 100% of queries using its own index was achievable.
That distinction is important.
The story isn’t:
OpenAI replaced Google and Bing.
The story is:
OpenAI is building the infrastructure that allows it to depend on them less.
ChatGPT Search is becoming a search router
The emerging architecture therefore looks less like one search engine and more like a retrieval marketplace.
Depending on the query, ChatGPT can potentially draw from:
OpenAI’s own index, third-party search providers, licensed publishing content, specialized databases, merchant feeds, local-data providers and live web retrieval.
Microsoft complicated the picture even further in June 2026 when it launched Web IQ, an AI-native retrieval infrastructure built on Microsoft’s search technology.
Microsoft explicitly says Web IQ already powers grounding experiences for both Microsoft Copilot and ChatGPT. Source
Interestingly, Web IQ itself isn’t traditional search results packaged as ten blue links. Microsoft describes it as semantic, passage-level retrieval designed specifically to deliver evidence to AI models. Its architecture combines embeddings, approximate-nearest-neighbor search, ranking and live updates. Command Line
So asking “Does ChatGPT use Bing or its own index?” may already be the wrong question.
The answer increasingly appears to be:
both – plus several other systems.
Depending on what you’re asking.
And OpenAI is actively testing whether its index can win
The GEO tool provider found another particularly revealing clue.
In August 2026, its researchers observed an internal ChatGPT experiment called:
prefer-index-over-serp-v3
At the time of observation, it reportedly affected around 8% of conversations.
Prefer index over SERP.
In other words: when should OpenAI trust its own retrieval infrastructure rather than obtaining search results externally?
Search Engine Watch independently reported on these findings and concluded that OpenAI appears to be actively testing how prominently its own index should be used compared with third-party results. Search Engine Watch
This matters enormously for marketers.
Because Google ranking, Bing ranking and ChatGPT visibility are no longer necessarily the same thing.
OpenAI even has its own search crawler
There’s another clue sitting openly inside millions of server logs:
OAI-SearchBot.
OpenAI tells publishers that if they want their content to be discoverable, summarized and cited in ChatGPT Search, they should allow OAI-SearchBot to crawl their websites. OpenAI Help Center
OpenAI’s documentation now also tells site owners that ChatGPT ranks search results using multiple factors related to relevance and reliability, and that inclusion is not guaranteed even if the crawler can access the page. OpenAI Help Center
This distinction is worth understanding.
OpenAI operates different crawlers for different purposes.
Allowing a model-training crawler is not the same thing as making content available for search.
For brands that previously blocked every “AI bot” with one giant robots.txt rule, this may deserve another look.
Blocking AI training while remaining discoverable in AI search are increasingly two different decisions.
The index is also a cache
Perhaps the clearest confirmation came recently from OpenAI itself.
OpenAI introduced an “offline web search” configuration for certain controlled ChatGPT environments. Instead of connecting to live external search providers when a query happens, ChatGPT can search:
OpenAI’s indexed and cached web content. OpenAI Help Center
That tells us something important about how modern AI search works.
The AI does not necessarily visit your webpage every time somebody asks about you.
Your page may already have been:
crawled → processed → indexed → cached → retrieved later.
That’s normal search-engine architecture.
And it changes how we need to think about AI visibility.
SEO isn’t dead. But “rank in Google and you’re done” definitely is.
Traditional SEO still matters.
A lot.
Google and Microsoft continue to provide important discovery and retrieval infrastructure. OpenAI itself admits third-party search providers remain part of ChatGPT Search. Microsoft confirms ChatGPT uses Web IQ. OpenAI
But there is now another layer.
A company may rank well in Google yet appear inconsistently in ChatGPT.
A competitor may be cited by ChatGPT even though it isn’t the highest-ranking organic result.
Different AI models may select completely different sources when answering essentially the same question.
And those differences are exactly why AI search monitoring has to happen at the answer level, not just at the ranking level.
This is something we see constantly with OtterlyAI.
Instead of assuming that Google rankings act as a proxy for AI visibility, OtterlyAI monitors the actual prompts customers ask across systems such as ChatGPT, Google AI experiences, Gemini and Perplexity and analyzes which brands appear, which sources get cited and how visibility changes over time.
Because increasingly, the answer itself is the search result.
And the infrastructure producing that answer is becoming more fragmented, not less.
GEO is becoming infrastructure optimization
This also changes the way I think about Generative Engine Optimization.
For a while, GEO was frequently presented as slightly modified SEO:
Create good content. Build authority. Get mentioned. Add structured data. Hope ChatGPT likes you.
That’s becoming too simplistic.
The emerging AI-search ecosystem has several layers:
Crawling → indexing → retrieval → ranking → source selection → synthesis → brand mention → citation.
You can fail at any one of them.
If OAI-SearchBot cannot crawl you, that is a problem.
If OpenAI crawls you but your important page isn’t retained in the relevant index, that’s another problem.
If you’re indexed but not retrieved for commercially important prompts, that’s another.
If you’re retrieved but a competitor becomes the source the model trusts enough to cite, that’s another.
And if you’re cited repeatedly but never mentioned as the recommended brand, you’ve discovered yet another problem.
Traditional rank tracking gives you only a small part of that picture.
Platforms like OtterlyAI therefore aren’t replacing SEO analytics. They are measuring the additional layer that appears after search engines, indexes and retrieval systems have done their work: what the AI finally tells the customer.
The next search war isn’t just about the model
Perhaps that’s the biggest takeaway from OpenAI’s index.
For several years, the AI competition looked like a model race:
GPT vs. Gemini vs. Claude.
Bigger context windows. Better benchmarks. More parameters. Faster inference.
Search infrastructure reminds us that models are only part of the equation.
To answer:
“What’s the best CRM for a 200-person SaaS company?”
the model needs information about the world right now.
Which companies exist? What do they offer? What do customers say about them? What has changed? Which source should be trusted?
That requires crawling, indexing, ranking, retrieval and freshness at enormous scale.
Google spent more than two decades building that machine.
OpenAI is trying to build its own version … while simultaneously buying and combining data from others.
And somewhere inside that infrastructure there may indeed be something called Labrador.
Which is perhaps appropriate.
Google has spiders crawling the web.
OpenAI apparently sent a dog.
The funny name shouldn’t distract from the serious implication:
ChatGPT is slowly becoming less dependent on somebody else’s map of the internet and increasingly building its own.
For anyone responsible for search, content or brand visibility, that’s the part worth paying attention to.
