TL;DR
- ChatGPT increasingly answers from its own index, which researchers have traced under the internal name Labrador.
- It is a family of vertical indexes covering the web, PDFs, YouTube, news, local, shopping and more, built on cached copies of pages and hybrid keyword and vector retrieval.
- Google and Microsoft still feed ChatGPT, but ranking there is no longer enough.
- Brands need OAI-SearchBot access, content that suits each vertical, visible freshness signals and answer-level measurement.
What is ChatGPT’s Labrador index?
For two years, the working assumption in SEO was that ChatGPT borrowed its results from Bing and Google. That assumption no longer holds. However, between May and July this year, a field called result_source appeared in ChatGPT’s server-side events with four values: Labrador, Bright, Oxylabs and SERP, according to Peec AI’s research. The last three point to external scraping and search providers. Labrador appears to be OpenAI’s own retrieval index.
Although OpenAI has not yet confirmed the index itself, its help centre describes an offline web search setting that uses “OpenAI’s indexed and cached web content“. The project is also not new. In the Google antitrust case, ChatGPT lead Nick Turley said OpenAI began building its index in 2023 after quality problems with non-Google search partners, and that Google had declined to provide API access.
How ChatGPT retrieval works today
Labrador is a family of indexes. Peec AI found separate ones for the general web, PDFs, YouTube, news (split into the last day, the last seven days and older), arXiv, Wikipedia, local, finance, legal, medical, shopping and images. Each stores full page content alongside crawl and publication dates, much like a conventional search index.
OpenAI is testing how far to trust it. In mid-August 2026, an experiment named prefer-index-over-serp-v3 was running on about 8% of chats, and at least five shopping experiments were live in early September. The shopping stack visible in those events combines BM25 keyword matching with reranking and vector search, a hybrid set-up that OpenAI’s own search job postings also describe.
The index sits on a cache. A one-billion-page test site had 6 million pages crawled by OpenAI by early September, at around 35,000 requests an hour. ChatGPT is also still a router: Peec AI found evidence that it queries Google, uses Bing in Deep Research, and draws on Microsoft’s Web IQ, Yelp and TripAdvisor. Turley himself told the court it could take at least five years to learn whether OpenAI’s index could ever answer every query alone.
What Labrador changes for your AI search strategy
Treat each vertical as its own surface. A separate PDF index means whitepapers, reports and price lists compete in their own pool, so give them descriptive titles, real text layers and clear authorship. The YouTube index makes transcripts and video titles worth optimising, and the tiered news index rewards content published inside its one-day and seven-day windows. For local intent, Peec AI recommends fixing Yelp and TripAdvisor listings first.
Exact wording matters again. In hybrid retrieval, keyword matching narrows the candidate pool before semantic ranking decides what is used. Pages that use the precise terms buyers type, and cover the topic in depth, give both stages something to work with. This is where expert-led content earns its place.
Your cached page is the page ChatGPT knows. OpenAI says there is no refresh guarantee for any specific URL. Its reasons for missing or stale pages include robots.txt rules, CDN bot protection, login walls and script-heavy rendering. A pricing change can stay invisible until OpenAI recrawls the page, which makes a technical AI search audit a priority. Visible publication and update dates help too, since the index stores both.
Google rankings are a weaker proxy. Because ChatGPT blends its own index with several outside sources, a page can rank well in Google and still appear inconsistently in ChatGPT. Measure at the answer level, then pair it with first-party data: OpenAI’s new Business Hub reports mentions, clicks and the topics where your site is cited.
A practical Labrador checklist
- Allow OAI-SearchBot in robots.txt and at your CDN, and decide separately whether to allow GPTBot for model training.
- Check server logs for OAI-SearchBot visits to your priority pages.
- Test what ChatGPT has cached. Peec AI suggests using Lockdown Mode, which serves cached results instead of browsing live pages.
- Map priority prompts to Labrador’s verticals and publish the matching formats, from PDFs and video to local listings.
- Keep key pages server-rendered, clearly dated and on stable URLs.
Conclusion
Traditional SEO still matters, and Google and Microsoft remain part of how ChatGPT grounds its answers. Labrador adds a second map of the web that OpenAI controls and is still tuning week by week, so some details here may change. Brands that are easy to crawl, cache and retrieve will be better placed whichever source ChatGPT routes a question to. If you want a review of how your site performs across these surfaces, talk to our AI search team.
