ChatGPT Quietly Built Its Own Search Engine. Bing Was Never the Full Story.
It's called Labrador, and it explains why your Bing rankings stopped predicting your ChatGPT visibility.
Everyone assumed chatgpt just queried BING for search results.
The whole SEO playbook rested on that one assumption: ChatGPT asks Bing, pulls the results, aggregates them, rewrites the answer. That assumption just broke!
A technical field spotted in ChatGPT's server stream between May and July tells a different story. OpenAI has been building its own index since 2023, not a single block but a whole family of specialized indexes: web, PDF, YouTube, news, shopping, and a dozen more. So the question changes shape. It's not "am I ranking well on Bing" anymore. It's "does ChatGPT even see me."
And the answer isn't the same for everyone. Someone tested this directly across 3 assistants and found out only one of them knew their product existed. That gap is exactly what this article is about.
The Assumption Everyone in SEO Just Lost
For 2 years the mental model was simple. ChatGPT doesn't crawl. It borrows. Ask it something time-sensitive, it fires a query at Bing, grabs the top results, and stitches an answer out of them. Rank well on Bing, show up in ChatGPT. Proxy metric, done deal.
The model was clean, testable, and wrong in a way nobody bothered to check because it was convenient. It let SEO teams keep doing what they already knew how to do (optimize for Bing, watch ChatGPT visibility follow along) instead of learning a second system from scratch.
Tomek Rudzki, who runs research at Peec AI, spent months digging through ChatGPT's server events and found a field that doesn't fit that story at all.
Meet Labrador: OpenAI's Own Family of Indexes
The field is called result_source. It shows up in the server-sent events ChatGPT streams back during a search, and it can carry 1 of 4 values: Labrador, Bright, Oxylabs, SERP.
Out of the 4 values, 3 are recognizable. Bright and Oxylabs are scraping infrastructure providers, the kind of company that sells you access to the open web at scale. SERP means exactly what it sounds like, a third-party search engine result page, Bing most likely.
Labrador is the one that doesn't belong on that list, because it's not a vendor. It's OpenAI's own name for its own retrieval system, and it isn't a single index. It's a family: a general web index, a PDF index, a YouTube index, a news index tuned for freshness, an arXiv index, a Wikipedia index, and separate verticals for local, finance, legal, medical, shopping, and images.
Google spent 20 years leveling up its index, one crawl budget increase at a time. OpenAI skipped most of the grind and speedran straight toward endgame gear, starting from a company that, 3 years ago, had never run a search index at all.
The Field That Leaked the Truth
Not everyone is convinced "index" is the right word.
Malte Landwehr, who tracks this space closely, put it in one line: good news, you can be visible in ChatGPT without ranking on Google, bad news, maximum visibility now means understanding Labrador too. He treats it as real, but he's careful with the label.
Others aren't as generous. One SEO on X dismissed the whole framing as phase 2 of a badly thought-out plan by GEO agencies trying to manufacture a new thing to sell. Fair skepticism, in the sense that a server field with 4 possible values doesn't automatically mean "search engine" in the way Google is a search engine. It might just mean an aggressive cache.
I think the caching theory undersells what's in the job postings, but I'm not fully sure the word "index" carries the same weight here as it does at Google's scale.
Malte Landwehr isn't ready to call it an index. Somebody at OpenAI clearly is.
Bing was never the story. It was just the alibi.
The Paper Trail Under Oath
OpenAI has been posting roles for indexing systems and retrieval pipelines at exaoctet scale for a while now, the kind of listings you don't write for a company that just wants a smarter Bing wrapper.
The clearer signal came out of a courtroom, not a job board. Nick Turley, who runs ChatGPT, testified under oath in the Google antitrust trial that OpenAI started building its own index in 2023. The initial internal goal was to answer 80% of queries without leaning on outside search providers, and that target was set before the end of that same year.
What Turley said next is the part worth sitting with. OpenAI later concluded it would take 5 years just to evaluate whether reaching 100% internal coverage was even achievable, let alone to actually get there. That's not a company bragging about a finished product. That's a company describing a multi-year infrastructure bet made under oath, in a legal setting where overstating your capabilities carries real consequences, which is exactly why the number is worth more than a blog post claim would be. A sworn number from a competitor's antitrust deposition doesn't get the benefit of marketing polish, and it still lands at "we don't know if 100% is reachable." That's the gap between the ambition stated in 2023 and the honesty forced out of a witness stand years later, and it's a wider gap than most SEO threads discussing Labrador seem to acknowledge.
Where You Can Watch It Happen Live: Shopping
If Labrador is real, shopping is where you can catch it acting like one.
In August, an A/B test internally labeled prefer-index-over-serp-v3 showed up affecting 8% of shopping conversations. Yes, v3. Somewhere a v1 and v2 got quietly retired, the same way every game patch fixes what the last one broke and never mentions it in the notes. The name tells you what it's testing: whether pulling results straight from the internal index beats going out to a third-party search page for that specific vertical.
Behind that test sits an actual retrieval stack, not a black box. Lexical preselection through BM25 pulls candidates from 10 different sources. Vector search runs on embeddings at 12,288 and 4,096 dimensions depending on the index. The reranker then narrows a field of around 400 candidates down to whatever ChatGPT shows you. None of that is exotic. It's the boring version of Skynet, no killer robots, just BM25 and a reranker doing paperwork at industrial scale. But running it in-house on shopping queries specifically means OpenAI trusts its own product data enough to skip the middleman there first.
The A/B test only tells you it's live in shopping. It doesn't tell you where it stops.
The Billion-Page Experiment That Explains the Caching
Metehan Yesilyurt, a GEO researcher at Peec AI, built something closer to a trap than a website: a test site with a billion pages, created specifically to watch how ChatGPT's crawler behaves when nothing is holding it back.
By early September, ChatGPT had explored around 6 million of those pages, at a sustained pace of roughly 35,000 requests per hour. That's not a chatbot occasionally checking a fact. That's a crawler running a systematic sweep.
I've watched a scraper job blow past its expected stopping point once, the request counter climbing long after I figured it should have quit, and there's a specific kind of dread in realizing the thing you built is still going and you don't know why yet. Multiply that by OpenAI's infrastructure budget and you get a plausible explanation for why Labrador can respond fast: if you're precomputing embeddings for millions of pages ahead of time, the actual query at inference time isn't doing much heavy lifting, it's matching against work that's already done.
Google and Bing Still Matter, Just Not Alone
None of this makes Google or Bing irrelevant. ChatGPT still leans on Google for some queries, on Microsoft's Web IQ product for others, and on Yelp or TripAdvisor specifically for local and travel verticals where those platforms have data OpenAI doesn't.
The honest picture is hybrid, not a clean replacement. Google didn't get one-shot killed like some final boss in a patch note, it just stopped being the only boss in the dungeon. Labrador covers the verticals where OpenAI decided the payoff justified building and maintaining its own retrieval pipeline. Everything else still routes outside. Treating this as "Google is dead" would be exactly as wrong as the assumption it's replacing, just pointed the other direction.
What Actually Changes for You
Stop treating your Bing ranking as a stand-in for ChatGPT visibility. It was never a perfect proxy, and now it's an incomplete one on top of that.
Check ChatGPT's lockdown mode against your own domain if you want a rough signal of whether your pages are already cached somewhere in Labrador. It won't tell you which vertical or how deep, but it beats guessing.
If your business runs through shopping queries at all, watch that vertical closer than the others. It's the one place where the shift from external results to internal index is already measurable, not theoretical.
If you want to go past checking and actually work the surface, I laid out the playbook for actual ChatGPT visibility separately, worth reading once you've confirmed you have a gap to close.
There's also a tool built specifically around this problem: ChatSEO tracks cross-assistant visibility instead of relying on a single proxy ranking, which is the exact blind spot this whole story is about. Worth a look if Bing rank has been your only signal until now.
I went deeper on the "every AI surface has its own retrieval logic" thesis back in 4 AI SEO Tactics Google Says Are Dead. This is that same argument, just playing out on ChatGPT specifically, with a name and a server field attached to it now.
The System Is Still Being Built in Public
The report that surfaced all this says the system changes every week. There's no way to know today, with certainty, whether a given site sits inside Labrador, at what depth, on which vertical.
What's confirmed: Bing stopped being the only proxy that counts. What isn't: where the line falls between what OpenAI already pulled in-house and what it still reaches out for. Even OpenAI doesn't seem to have a firm answer, if Turley's own testimony is anything to go by.
The map is still being drawn. It isn't finished.
Sources
This post may contain affiliate links. If you click them, I might earn a small commission — costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.
ChatGPT built its own search index family (Labrador) and stopped asking Bing. If you're shipping AI products, the demo-vs-product checklist in the kit shows you exactly which production layer you're missing when your AI can't find what matters.