Peec AI's product chief, writing in a thread about scraping Google, says he has seen ChatGPT cite Wikipedia articles and PDF files six months after they were deleted.
Peec AI's "ChatGPT built its own search index" reports the reason is a hybrid retrieval stack: Microsoft Bing used directly without scraping, third-party scraping for Google and local data, and OpenAI's own proprietary family of indexes called Labrador.
Bing-Direct, Scraped Google, and Labrador Verticals at a Glance
Grounding here means the retrieval step that supplies outside sources before ChatGPT writes an answer. Peec AI and Synscribe document the three layers this way:
| Layer | What it is | What it returns |
|---|---|---|
| Bing-direct | Microsoft Bing for general web and news, called directly without a scraping provider | General web and news results plus Microsoft Web IQ grounding output |
| Scraped Google via middleware | External scraping providers queried through OpenAI's Sonic orchestration system, the internal system that routes search calls | URL plus title plus snippet plus date metadata for web, news, and local |
| Labrador verticals | Labrador, OpenAI's internal family of specialized indexes, each tuned to one content type | Stored full page content with crawl and publication dates, including full licensed text for news and academia |
Peec AI's breakdown is the source for that table: direct search, scraped results and vertical indexes working side by side.
What Server Events Show About Who Serves What
Between May 21 and July 21, 2026, a field called result_source appeared in the server-sent events, the live stream ChatGPT servers send to a browser while generating, according to Peec AI.
That field carried four values only: Labrador, Bright, Oxylabs, and SERP, where Bright plus Oxylabs plus SERP are external scraping providers and Labrador is the own index, as Peec AI documents it.
Synscribe's "It's Not Bing or Google" analysis of the same events found 484 general web results tagged bright and five tagged labrador across three conversations, with every general web result carrying the bright tag in the raw result_source field.
Sonic is the orchestrator and Bright Data is the retrieval layer it calls, in Synscribe's account of those events. Bright Data, the web data platform formerly called Luminati Networks, returns structured metadata for standard web results. Oxylabs scrapes Google and feeds the news vertical specifically, while a third channel tagged SERP also serves news plus a mixed local variant, in Peec AI's breakdown of the provider list.
Synscribe states that Bright Data is middleware. Bright Data's own SERP product documentation lists Google, Bing, DuckDuckGo, Yandex, Baidu, Yahoo, and Naver across 195 countries, and Synscribe reports that searching the exact ChatGPT query strings manually in Google and Bing produced different ordering and different page sets, consistent with a blending or caching layer on top of upstream engines.
Which Labrador Vertical Serves Which Need
Labrador is a family of vertical indexes for general web, PDF documents, YouTube transcripts, news, arXiv papers, Wikipedia, local listings, finance including a separate finance PDF index, legal, medical, shopping, and images, as Peec AI lists them.
Each entry stores full page content with crawl date and publication date, the same structure Google spent two decades building with a general web index plus specialized verticals layered on top, in Peec AI's description.
Peec AI and Synscribe map the split this way:
- General web: Labrador general web plus Bing-direct plus scraped Google through Bright middleware.
- PDF, YouTube transcripts, arXiv, Wikipedia: Labrador verticals for documents and video text.
- News: three sources side by side. Bing News direct, scraped news through Oxylabs and the SERP channel, and Labrador news with separate tiers for the last day, the last seven days, and everything older.
- Local: scraped Google Maps through Bright plus business data from licensed partners Yelp and TripAdvisor plus internal pipes routing to Google Maps and business sites, alongside the Labrador local vertical, in Peec AI's provider breakdown.
- Shopping: OpenAI's own product index under active test against scraped results on slices of live traffic, with experiments named prefer-index-over-serp-v3 and shopping-index-q2qb running on the own index as of early September 2026, plus merchant feeds behind the shopping surface, as Peec AI reports the tests. One test affected 8 percent of chats at the time it was spotted, with slices ranging from single digits up to 20 percent.
- Finance, legal, medical, images: Labrador verticals, with finance holding its own PDF slice.
Microsoft's own Web IQ grounding platform names ChatGPT directly among the products it already powers alongside Copilot and Nasdaq, as Peec AI cites it. OpenAI's head of ChatGPT testified under oath during Google's antitrust trial that OpenAI had quality issues with non-Google search data and began building its own index in 2023, aiming to answer most queries from it over time, as Peec AI excerpts the testimony.
Why Licensed Full Text Beats a Snippet
For standard web results ChatGPT receives snippets, in Synscribe's reading of the result schema. Each Bright result carries URL plus title plus snippet of about 150 characters plus publication date plus result_source plus content type, with the content body empty.
The five Labrador results Synscribe found behaved differently. They appeared on Reuters, MarketWatch, and arXiv with content type news or academia instead of search, and their snippet fields held full multi-paragraph article text.
Peec AI describes the same advantage from the index side: Labrador incorporates licensed content with direct data agreements plus direct publisher pipelines alongside its independent crawler. A Reuters article about a company can therefore supply ChatGPT with paragraphs of product and market description, where a standard page supplies one dense snippet. That gap is why Synscribe frames sustained press coverage as a structurally richer retrieval signal.
Why Long Cache Makes Every Recency Claim Suspect
Live retrieval at ChatGPT scale is too slow to run on every query, so OpenAI relies heavily on cached page data with background updates, as Peec AI puts it. Peec AI's product chief calls these systems retrieval systems with very long cache times and treats a citation six months after deletion as proof the store qualifies as an index.
A Chinese-language synthesis shared among SEO data builders reports the consequence for AI visibility work in its sample: only about 1.6 percent of new citations came from new domains, 27 percent were new URLs on known domains, and PDF citation share rose from 0.5 percent to 1.12 percent. The reported lesson is depth on known-authority domains over new-domain expansion.
A long cache is also why a live check matters. When a ChatGPT answer cites a page, compare it with what Google ranks for the same query today. With qequals, an agent can fetch today's Google results for that query and put the two side by side, dated, so a stale citation stands out.
Builders deciding what answers can cite can run three checks:
- Stop treating Bing coverage as a proxy for ChatGPT coverage. Peec AI reports the overlap is lower than most teams assume, with some of what looks like Bing routed through Web IQ, a different product built for grounding.
- If you sell products, pull shopping queries in ChatGPT separately from web queries. Shopping is where prefer-index-over-serp-v3 and shopping-index-q2qb test the own index on live traffic slices.
- If local inquiries matter, fix Yelp and TripAdvisor listings first. Peec AI lists them as licensed partners whose data reaches ChatGPT with a direct feed, while Maps data arrives through scraping.
Whichever you start with, record the date of every check. With a cache this long, what ChatGPT cites and what the web says now can drift apart for months.
