Turley says OpenAI planned its own index from day one
Sworn testimony and hiring point to a proprietary retrieval stack, not a Bing wrapper.
What happened
Tomek Rudzki published a LinkedIn post on 4 September 2026 describing ChatGPT's retrieval system, internally called Labrador. He reports it is a family of indexes covering general web, PDF, YouTube, news, arXiv, Wikipedia, local, finance, legal, medical, shopping and images. Rudzki cites Nick Turley, Head of ChatGPT, testifying under oath in a 2025 US antitrust hearing that an in-house index was the plan from day one. Rudzki adds that recent OpenAI job posts mention exabyte-scale database systems.
What we think is going on
The working assumption across our field has been that ChatGPT rides on Bing. Turley's sworn testimony, as Rudzki reports it, says the opposite about intent. OpenAI wanted Google's data for quality reasons and to improve its own index.
That reframes the licensing deals as a bridge, not a destination. Practitioners have been planning against a vendor relationship OpenAI always intended to exit. The vertical split Rudzki lists, with separate indexes for medical, legal, shopping and the rest, points to retrieval tuned per domain rather than a single web crawl reranked.
We think the testimony and the job posts are directional. They speak to intent and capacity. They are not a measurement of how much of today's answer set comes from a proprietary index versus a licensed one.
What to do about it
Measure the gap on your own properties this week. Pull the queries where you rank in Bing's top ten for your category. Check which of those surface you as a citation in ChatGPT, and which do not.
The delta is your exposure to the shift Rudzki is describing. If ChatGPT already cites you where Bing ranks you, a Bing-to-Labrador transition is a smaller risk. If it does not, Bing rank is not carrying you today either.
From there, work on being the source a careful editor in your vertical would pick. A specialist reading your medical pages should want to cite them. Retrieval-side tuning by OpenAI will not rescue pages that a human expert would pass over.
Treat schema, freshness and site structure as table stakes that support that judgement, not substitutes for it. Rudzki's post is one practitioner's read of testimony and hiring signals, so watch and verify before rebuilding anything.
What we're watching
Two things would firm this up. Public disclosure or benchmarking that shows what share of ChatGPT citations already come from OpenAI's own crawl versus Bing. And any change in the Microsoft-OpenAI licensing terms, which would tell us how close the exit Rudzki implies actually is.