Knowledge baseSourcesOlivier de Segonzac (RESONEO)

Inside ChatGPT's retrieval stack: index, cache and live pages

Author Olivier de Segonzac (RESONEO) Date 2026-08-17 Model state GPT-5.x, free and paid accounts, France, extension v5.1 Open source →

Key findings

Segonzac describes the retrieval layer behind web.run as three stages: an own index (Labrador) with query-independent excerpts of about 200 characters that begin at the H1 and ignore the meta description; a reading cache that holds pages as markdown with 30 minutes of freshness and, in doing so, ignores Cache-Control, noindex and JSON-LD; and live openings. Opened pages are cited 74%, only-found ones 7%. Free and paid accounts run through different pipelines in thinking mode.

Why this source matters for the project

It is the most detailed description of what arrives at ChatGPT from the document, and thus the basis for every recommendation to site operators. The parts the client does not see (cache, size limit) we cannot check; the parts it does see (excerpt length, title) we check daily.

Checked against our data

The 200-character signature of the index is stably attested in our data and the reason search.result_signature can recognise the pipeline even though result_source is missing. The cohort claim about paid thinking does not fit our Plus account, where thinking runs mostly show the index signature. That is either a difference between France and Germany, between cohorts, or a change after 17 Aug; without a second account not decidable.

ClaimStatusEvidence
Three layers: Labrador index, reading cache (markdown, 30 min), live openunverifiableNot visible from the client; we see only the hit formatting and open calls.
Index snippets ~200 characters, query-independent, at the H1, meta description ignoredconfirmedSnippet median 202 characters for local, product, entity and comparison questions on three measurement days (search.result_signature).
Opened pages cited 74%, only-found ones 7%unverifiableNo open observed in standard mode; thinking sample too small.
Free thinking 74.7% Labrador, paid thinking 75.3% Google scrapeoutdatedPlus account thinking for us: K1-01, K2-01, K6-01 thinking median 202 (index), only K5-01 thinking on 21 Aug 158 (scrape). Either cohort or a change after 17 Aug.
Pages over 4 MB are discarded, JSON-LD stripped, noindex ignoredunverifiableServer-side, not measurable.

Related

segonzac-2026-05-14 segonzac-2026-02-01