| Tier |
Typical use |
Source |
Evidence status |
What the evidence says |
Reference |
| 1 |
Web & search discovery |
Google Search |
Confirmed + current |
Grounding with Google Search connects Gemini to real-time web content, returning inline citations to source URLs. |
Google — Gemini API docs, Grounding with Google Search |
| 1 |
Bing Search |
Confirmed + current |
Microsoft documents Bing results being used to enhance Copilot responses. Not re-verified in this pass. |
Microsoft Bing |
| 3 |
Common Crawl |
Confirmed historical |
GPT-3 used filtered Common Crawl as roughly 60% of its sampling mixture; LLaMA 1 reported 67%. |
Common Crawl |
| 3 |
Historical web corpora (C4 etc.) |
Confirmed historical |
C4 is a cleaned derivative of Common Crawl; LLaMA reported C4 at 15% of its pretraining mixture. |
TensorFlow Datasets — C4 |
| 4 |
Web grounding services |
Strong evidence / likely |
Category inference covering third-party grounding/retrieval intermediaries. No single canonical source. |
— |
| 1 |
Products & shopping |
Google Merchant Center |
Confirmed + current |
Merchant feed data underpins Google’s shopping surfaces. Retained on user instruction; Google Shopping removed as it is a surface, not a source. |
Google Merchant Center Help |
| 2 |
Merchant / retail feeds (OpenAI) |
Confirmed + current |
Merchants share a secure, regularly refreshed CSV/JSON feed of identifiers, descriptions, pricing, inventory, media and fulfilment so ChatGPT can surface products accurately. Refreshes accepted as often as every 15 minutes. |
OpenAI Developers — Agentic Commerce, product feeds |
| 4 |
Microsoft Merchant Center |
Strong evidence / likely |
Equivalent commercial feed infrastructure; inferred parallel to Google Merchant Center rather than separately evidenced. |
— |
| 4 |
Marketplace feeds |
Strong evidence / likely |
Category inference. Shopify catalog data is already integrated into ChatGPT, which supports the pattern. |
OpenAI Help — Shopping with ChatGPT Search |
| 1 |
Local & places |
Google Maps |
Confirmed + current |
Grounding with Google Maps is a documented tool alongside Search grounding, giving models geospatial context. |
Google Cloud — Grounding API |
| 1 |
Google Business Profile |
Confirmed + current |
Business profile data feeds Google’s local surfaces. Carried from the source table; not separately re-verified. |
— |
| 1 |
Yelp |
Confirmed + current |
Yelp licenses reviews, photos and business information to OpenAI for real-time local recommendations. Beyond grounding it also drives actions: ChatGPT users can book a table or join a waitlist, and Request a Quote lets users contact providers in-chat. Yelp’s 10-Q confirms it is live. |
Axios; Yelp blog; Yelp 10-Q FY2026 |
| 4 |
OpenStreetMap |
Strong evidence / likely |
Widely used open geospatial corpus; inferred rather than confirmed for any named model. |
— |
| 4 |
Foursquare |
Strong evidence / likely |
Left in Tier 4 deliberately: the OpenAI deal is Yelp’s, and no equivalent evidence exists for Foursquare. |
— |
| 4 |
Tripadvisor |
Strong evidence / likely |
Category inference for review/travel data. No confirmed deal identified in this pass. |
— |
| 1 |
Knowledge & reference |
Wikipedia |
Confirmed + current |
Explicitly present in GPT-3’s disclosed mixture and LLaMA (June-Aug 2022 dumps, 20 languages); also widely used as a live reference/RAG corpus. |
Wikimedia dumps |
| 1 |
Wikimedia |
Confirmed + current |
Same corpus family as Wikipedia. Licensing is unusually clear: principally CC BY-SA with attribution/share-alike obligations. |
Wikimedia dumps |
| 4 |
Wikidata |
Strong evidence / likely |
Structured entity layer; strongly implied by knowledge-graph use but not separately confirmed. |
— |
| 1 |
Community / Q&A / social |
Reddit |
Confirmed + current |
The Google deal gave access to the Reddit Data API for ‘real-time, structured, unique content’, and allows Reddit content to be displayed across Google products — i.e. live grounding, not only training. |
Tom’s Guide (Google/Reddit deal) |
| 2 |
Reddit |
Confirmed + current |
Same deal, training side: Google may use Reddit posts to train its AI models and improve services such as Search; reported at roughly $60m/yr. NOTE: Reddit is reportedly weighing whether to renew — treat as unstable. |
Fortune; Neowin/WSJ on renewal doubt |
| 4 |
Social platforms |
Strong evidence / likely |
Category inference covering platform-wide social corpora. |
— |
| 4 |
Forums / communities |
Strong evidence / likely |
Category inference. Overlaps Reddit but generalised to non-Reddit forums. |
— |
| 1 |
News & publisher content |
Live publisher pages |
Confirmed + current |
Reached at inference time via search grounding rather than pretraining; retrieval selection and crawlability govern inclusion. |
Google — Grounding with Google Search |
| 2 |
Licensed publisher content |
Confirmed + current |
OpenAI has multiple explicit licensing partnerships (FT, Axel Springer, AP, News Corp). Terms differ per partner on training vs grounding vs attribution. |
OpenAI — FT content partnership |
| 2 |
Publisher partnerships |
Confirmed + current |
Axel Springer’s deal includes otherwise paywalled material in answers; AP licensed part of its text archive. |
OpenAI — Axel Springer partnership |
| 3 |
Historical news corpora |
Confirmed historical |
Archive material absorbed in pretraining; distinct from live licensed access. |
— |
| 2 |
Developer / technical |
GitHub |
Confirmed + current |
LLaMA used GitHub’s public BigQuery dataset, restricted to Apache/BSD/MIT projects; The Pile separately includes GitHub. Public visibility is not an open licence. |
GitHub |
| 2 |
Stack Overflow |
Confirmed + current |
Named in licensing-deal mapping alongside Reddit and Shutterstock as a data platform powering multiple buyers. |
LLM Pulse — AI content licensing deals mapped |
| 2 |
Technical docs |
Confirmed + current |
Vendor documentation corpora; widely used but not tied to a single disclosed agreement. |
— |
| 4 |
npm / PyPI registries |
Strong evidence / likely |
WEAKEST ENTRY IN THE TABLE. Relabelled from ‘package registries’ to name examples. No disclosed agreement or documented retrieval use found — consider cutting. |
— |
| 1 |
Travel & commerce actions |
Google Hotel Center feeds |
Confirmed + current |
When Gemini or AI Mode show hotel options with real-time prices, that data comes from the Google Hotels feed. In Aug 2026 Google added hotel booking inside AI Mode completed with Google Pay, so this is grounding plus actions. |
TechCrunch — AI Mode travel update |
| 4 |
Booking / partner feeds |
Strong evidence / likely |
Booking Holdings and IHG are reported as participants in Google’s agentic booking pilot, which supports the direction but stops short of a documented feed spec. |
InfosTourisme (IHG/Booking pilot) |
| 4 |
OTA / commerce sources |
Strong evidence / likely |
Category inference. OTAs run their own rate feeds into these surfaces. |
— |
| 4 |
Reservation / inventory APIs |
Strong evidence / likely |
Category inference covering booking/inventory endpoints exposed to agents. |
— |