GPTSpy knowledge base

Knowledge base

A register: every observed object of the ChatGPT pipeline has its own page — with measurements, context and dated external sources. Method and setup · Fan-out report · Not yet described

See the whole machine run

9 steps · animated request flow · real data from 2026-10-06

From question to answer

Every step is an object in the knowledge base. As of 22 Aug 2026, GPT-5.6, Plus account, browser interface.

Session. The client loads /backend-api/me, /settings/user, /models, /system_hints and about 50 more endpoints → backend-api-settings-user, eclipse, backend-api-system-hints. The tracker's context snapshot (platform.*) comes from here.
Sentinel. Before every message sentinel/chat-requirements/prepare and finalize, Turnstile and proof-of-work in the browser → backend-api-sentinel-chat-requirements-prepare. The reason the runner drives the UI rather than the API.
Send. POST /f/conversation with 17 payload keys including model, force_parallel_switch, paragen_cot_summary_display_override → force_parallel_switch (send payload). The reply arrives as an SSE stream → UI event types in the stream (Card, Row, Pressable …).
Search decision. Does the model search or not? Formerly readable as Sonic probabilities, today only from the outcome → Search decision, Search rate.
Fan-out. A code message to web.run, line-based pipe syntax: mode, query, time window, domain → web.run syntax (pipe format), Query fan-out, Freshness window, Domain restriction, Fan-outs per question.
Retrieval. In-house index or Google scrape, recognisable by excerpt length → Retrieval pipelines (Labrador, scrape), Snippet median (pipeline signature). Result: search_result_groups, 14–22 URLs grouped by domain → grouped_webpages / search_result_groups.
Reading. Opening individual hits (open|turnXsearchY), so far only in thinking mode → open (opened results), Opened hits.
Answer. Text with references: inline links → url (inline reference), entities → entity (entity markup), widgets → Widgets (genui, dil, map, nav_list), source list → sources_footnote. Numbers → Citations per answer, Entries in the sources footnote, Widget rate.
Shopping. For buying advice, a product carousel from the purchased feed (Google Shopping IDs, provider p2): base card in the answer, merchant offers only via POST /search/product_update at render time, utm_source=chatgpt.com on every merchant URL → ChatGPT shopping, product_entity, checkoutable (instant checkout), Products per carousel, Offers per product.
Fetch. The client loads the finished object: since 21 Aug 2026 GET /backend-api/conversations/{id} as a paginated messages array → /backend-api/conversations/{id}. What is missing there (fan-outs, Sonic, offers) lives on in the stream or the follow-up endpoints.

What the client does not see and this map therefore does not show: providers behind web.run, reading cache, fusion/ranking, crawlers. For those, see the external evidence at Retrieval pipelines (Labrador, scrape) and segonzac-2026-08-17.

Ten points where ChatGPT decides on visibility

In the order they apply. Each is a filter; what fails the first never sees the others.

Search decision — Does ChatGPT search at all? 69 % of our questions, yes. Evergreen knowledge, health, long-tail problems: no. No website helps there.
Query fan-out — Is the brand already in the query? ChatGPT writes product names into the fan-outs before it searches. Named brands are cited 33× more often than retrieved ones.
Domain restriction — Is the domain on the list? For 8 of 18 searching standard questions ChatGPT searches only one to five domains. All others do not exist for that query.
Freshness window — Is the page young — or stable — enough? The window follows the fact: news 1 day, prices 7, products 30–90, stable facts (biography, specs) up to 3650 days. Outside the window, nothing is considered.
Retrieval pipelines (Labrador, scrape) — Which corpus delivers? In-house index (reads 200 characters from the H1, ignores the meta description) or Google scrape (Google's snippet). Topic decides.
grouped_webpages / search_result_groups — Is the page among the 14–22 hits? Grouped by domain; several pages of one domain compete with each other.
open (opened results) — Does the page get read? Opened pages are cited 74 % of the time, merely retrieved ones 7 %. In standard mode we see no opening.
sources_footnote — Is the page in the source list? Two to six entries. Only those are seen by the user.
url (inline reference) — Is there a link in the text? The only form that gets clicked. For local and navigational questions, partly from model knowledge without a search.
Widgets (genui, dil, map, nav_list) and entity (entity markup) — Does an object replace the website? Map, follow-up question, product carousel, entity. No link to be won there; 18 % of our runs.

Figures are from 22 Aug 2026 and are updated daily under Search rate and the other metrics.

Register

From question to answer

Ten points where ChatGPT decides on visibility

Register

Endpoints 6

Fields 6

Feature flags 1

Models 1

Retrieval pipelines 1

Reference types 5

Send payload 1

Sources 27

Stream types 1

Terms 3

web.run modes 9

Widgets 1