Knowledge baseRetrieval pipelinesRetrieval pipelines (Labrador, scrape)

Retrieval pipelines (Labrador, scrape) pipeline/labrador

Retrieval pipelines

What it is

ChatGPT fetches results from several sources: an in-house index (internally labrador), purchased Google results via scraping providers (bright, oxylabs), its own SERP source and, at times, bing. Until 21 Jul 2026 the source was stated per hit in result_source. Since then it is missing. We recognise the pipeline from the signature of the results: median excerpt length, share without excerpt, titles over 75 characters. Measured 20–22 Aug 2026: local, product, entity and comparison questions sit stably at median 202 characters (in-house index); news and price questions at 131–158 characters with titles under 60 characters (Google scrape).

External evidence

For visibility

The pipeline determines what of a document arrives. Segonzac documents for the index: excerpt ~200 characters from the H1, meta description ignored, JSON-LD stripped, pages over 4 MB discarded, reading cache 30 minutes. For the Google scrape, Google's snippet rules apply. A page relying on meta descriptions and schema is not recognised by them in the index path. Whether a question runs through the index or the scrape path depends on topic (live data goes to the scrape) and, according to Mohanadasan, on the account cohort.

External sources

mohanadasan-2026-06-24 · mohanadasan-2026-07-14 · segonzac-2026-08-17 · goodwin-2026-07-08

Change history

No event has touched this object yet.

Open

Related

Query fan-out open (opened results) Snippet median (pipeline signature)