An AI model has no single, permanent search index attached to it. The product that hosts the model picks its search tools, data sources and instructions, so the same model can read different pages in different apps.
This post is part of Zia Wasi's research on AI search. Zia Wasi is co-founder of Depra AI and a GEO researcher; GEO, or generative engine optimization, is the work of getting a brand named in AI answers. He collected everything below from public vendor documentation and checked each source on 13 October 2026. Every documented route is linked; where a vendor documents nothing, the post says so.
In short: Google AI Overviews and AI Mode use Google Search systems (Google Search Central, checked 13 Oct 2026). Microsoft documents Bing for web grounding in Copilot (Microsoft Learn, checked 13 Oct 2026). Perplexity says it built its own index (Perplexity, checked 13 Oct 2026). OpenAI and Anthropic document their own search tools and crawlers, and xAI documents its own search tools, without naming one exclusive upstream index (OpenAI, Anthropic, xAI, checked 13 Oct 2026).
| Surface | Documented discovery or grounding route | Confidence and limit | Source |
|---|---|---|---|
| Google AI Overviews and AI Mode | Google Search systems, with related searches and supporting web links | Confirmed. The two surfaces may use different techniques and show different links. | Google Search Central |
| Gemini with Google Search | The google_search tool retrieves Google results for grounded API answers | Confirmed for the API. API output is not a copy of the Gemini app or AI Mode. | Google AI for Developers |
| ChatGPT and OpenAI APIs | OpenAI search crawler and web search tool; Microsoft announced Bing integration in 2023 | Confirmed tools and a historical partnership. No complete, exclusive current upstream map. | OpenAI crawlers, OpenAI web search, Bing 2023 |
| Microsoft Copilot | Bing search for web grounding in Copilot and Copilot Chat | Confirmed for the documented products. Work data and admin settings can change the answer context. | Microsoft Learn |
| Claude | Native web search and Anthropic search crawling; Bing for image search; Brave named in a scoped government article | Confirmed tools. The Brave evidence is scoped. No verified rule that Claude uses only Brave. | Anthropic, Claude Help, Claude for Government |
| Perplexity | Its own search index, a Search API and answer services | Confirmed own index. The model a user picks does not identify the full retrieval setup. | Perplexity, Perplexity Docs |
| Grok | Separate web_search and x_search tools; X keyword, semantic and thread retrieval | Confirmed tools. The docs do not name an exclusive upstream web index. | xAI web search, xAI X search |
| Meta AI on Facebook | Facebook AI Mode draws on public discussions across Meta apps | Confirmed for that surface only. Not proof of how every Meta AI app or country works. | Meta |
Why does the product matter more than the model?
The product decides when to search, what to search and which results the model reads. So "which index does GPT or Gemini use" has no single answer; the right question is which product, in which mode, with which tools.
Grounding means giving the model outside evidence at the moment it writes an answer. It is different from training. A new page can help a searching assistant long before it is part of any future training data. A model can also name a familiar brand without reading any current page.
Search rank matters because it shapes which pages enter consideration. It is not the whole story. OpenAI separates the sources its web search consulted from the smaller set it cites inline (OpenAI web search guide, checked 13 Oct 2026). A page can be read and still not be shown.
Which index do Google AI Overviews and AI Mode use?
Google AI Overviews and AI Mode use Google Search. Google says a page needs to be indexed and eligible for a search snippet to appear as a supporting link in either one (Google Search Central, checked 13 Oct 2026). Both may run several related searches, and their links can differ (Google Search Central).
Google says no special AI file and no special schema is needed, and that traffic from these features is counted in the Web report of Search Console (Google Search Central, checked 13 Oct 2026). So Search Console Web totals are not a clean count of AI citations.
Google-Extended is a robots.txt control token, not a separate crawler. It governs some Gemini training and grounding uses, and Google says it does not control inclusion or ranking in Google Search; Googlebot controls Search crawling (Google crawlers, checked 13 Oct 2026). A brand can block one and still allow the other, so a single "AI blocked" verdict hides the detail.
Does Gemini use Google Search?
The Gemini API can, through its google_search tool. The documentation describes generated search queries and answers with citation metadata (Google AI for Developers, checked 13 Oct 2026).
What is not documented: that an API answer matches what a person sees in the Gemini app or in AI Mode. A Gemini API test tells you whether that configured model names a brand when it can search. Measure and label the API and the app separately.
Does ChatGPT use Bing?
Microsoft announced a Bing search integration for ChatGPT in May 2023 (Bing blog, checked 13 Oct 2026). That is a documented historical relationship. It is not a current, complete map of where ChatGPT search gets its results, and no public source checked for this post gives one.
OpenAI documents three separate agents: OAI-SearchBot for search, GPTBot for possible training use, and ChatGPT-User for some visits a user triggers (OpenAI crawlers, checked 13 Oct 2026). Search and training settings are independent, and allowing GPTBot is not needed for search crawling (OpenAI crawlers).
OpenAI's web search API can return the sources it consulted as well as its inline citations, and settings such as domain filters change the result (OpenAI web search guide, checked 13 Oct 2026). So a Bing ranking is useful evidence, but it does not prove inclusion in a ChatGPT answer. Track ChatGPT directly.
Which index does Microsoft Copilot use?
Microsoft documents Bing search for web grounding in Copilot and Copilot Chat. It says Bing is a web grounding service for these products, that the product may generate search queries from the prompt, and that admins control web access (Microsoft Learn, checked 13 Oct 2026).
Site owners can see part of this in Bing Webmaster Tools. Its AI Performance report, announced in February 2026, shows citations and grounding queries across Microsoft experiences and selected partners (Bing Webmaster blog, checked 13 Oct 2026). A June 2026 update added intent, topic, citation share and compare views; citation share does not show rival domains and is not traffic share (Bing blog, checked 13 Oct 2026).
The old public Bing Search APIs were retired on 11 August 2025 (Microsoft Lifecycle, checked 13 Oct 2026). Microsoft now offers web search with Bing grounding for agents in Foundry (Microsoft Learn, checked 13 Oct 2026).
Does Claude use Brave Search?
There is no verified public rule that Claude uses only Brave. What Anthropic documents is a web search tool where Claude decides when to search, may search again and writes cited answers, with domain and location controls (Anthropic web search tool, checked 13 Oct 2026).
Anthropic separates ClaudeBot for possible training use, Claude-SearchBot for search quality and indexing, and Claude-User for requests a user starts (Claude Help Center, checked 13 Oct 2026). Claude's image search is documented as powered by Bing (Claude Help Center, checked 13 Oct 2026).
The Brave link comes from an earlier Claude for Government help article that named the Brave Search API for its web search connector. That article now redirects (Claude Help Center, checked 13 Oct 2026). It is scoped evidence, not a routing rule for every Claude product.
Brave says its Search API runs on its own independent web index (Brave Search API, checked 13 Oct 2026). Brave also says its crawler will not crawl a page that Googlebot cannot crawl, and offers a Submit URL route for re-fetching pages (Brave Search Help, checked 13 Oct 2026). That rule does not mean Brave uses Google's index. A fair test, proposed in the research, is to run a fixed set of buying questions through Claude, send the searches Claude ran to Brave with the same country and language, and report the URL overlap for that sample only. Overlap is evidence of shared sources, not proof that Claude uses Brave.
Does Perplexity have its own search index?
Yes. Perplexity's Search API announcement says it built an index of billions of web pages, and that the API returns ranked web results without a generated answer (Perplexity, checked 13 Oct 2026). Its documentation separates raw search results from generated answers with citations, and offers domain, language and region filters and multi-query requests (Perplexity Docs, checked 13 Oct 2026).
Perplexity describes PerplexityBot as a search crawler, not a crawler for training foundation models, and Perplexity-User as the agent for visits users start; it publishes IP ranges to verify both (Perplexity crawlers, checked 13 Oct 2026).
What does not follow: that Perplexity is mostly a forum search, or that picking a Claude or GPT model inside Perplexity makes the answer a Claude or ChatGPT measurement. The research reads it this way: the hosting product still sets retrieval, though Perplexity does not publish a per-request routing rule. Forums, publishers, documentation and shop pages are all source types, and which ones win depends on the question.
What do Grok and Meta AI read?
Grok has two separate search tools. web_search searches the web and browses pages (xAI docs, checked 13 Oct 2026). x_search searches X by keyword or meaning, finds users and fetches threads (xAI docs, checked 13 Oct 2026). The web documentation does not name an exclusive upstream web index (xAI docs).
Meta's June 2026 announcement describes Facebook AI Mode drawing on publicly shared material across its apps, including Groups and Reels (Meta, checked 13 Oct 2026). That is one surface. It is not proof of how WhatsApp, Instagram or the standalone Meta AI app retrieve sources.
Open models such as Llama, Qwen or DeepSeek can be paired with any search tool, private documents or no retrieval at all. The model family alone tells you nothing about the index.
What this means for a brand
A brand does not need one trick per engine. It needs to be findable in the indexes these products document, open to their search crawlers, and mentioned on the pages each engine actually reads.
- Be indexed by Google and Bing. Google AI Overviews and AI Mode need an indexed, snippet-eligible page (Google Search Central, checked 13 Oct 2026), and Copilot grounds on Bing (Microsoft Learn, checked 13 Oct 2026). Use Search Console and Bing Webmaster Tools for your own site.
- Allow the search crawlers, decide training separately. OAI-SearchBot, Claude-SearchBot and PerplexityBot are documented as search crawlers. OpenAI and Anthropic name separate training crawlers (GPTBot and ClaudeBot), and Perplexity says PerplexityBot is not used to train foundation models (OpenAI, Anthropic, Perplexity, checked 13 Oct 2026). Our AI crawlers list explains each one.
- Do not expect IndexNow to reach every engine. The IndexNow registry lists Bing, Yandex, Seznam, Naver, Yep, Internet Archive and Amazonbot; Google, Brave, OpenAI, Anthropic and Perplexity are not listed (IndexNow registry, checked 13 Oct 2026). A submission is a change notice, not a promise of indexing (IndexNow FAQ, checked 13 Oct 2026).
- Earn mentions on the sources each engine reads. Read the citations in your own category, not a general rule. In the October 2026 re-run of Depra AI's Hinglish vs English study (479 answers from ChatGPT, Gemini and Perplexity in India), amazon.in was cited in 132 answers, flipkart.com in 82 and nykaa.com in 65, so marketplaces were among the most-cited sites for those shopping questions.
How Depra AI measures it
Depra AI tracks four engines on every plan: ChatGPT with web search, Gemini, Google AI Overviews and Perplexity. ChatGPT and Gemini answers are captured from their consumer web apps, Google AI Overviews from a real Google results page in the country you sell in, and Perplexity through Sonar, its answer model with live web search, set to that country. See the methodology for the details.
ChatGPT, Gemini and Google AI Overviews are checked daily on every plan. Perplexity is checked weekly on Starter, Growth and Scale, and daily on Enterprise. Microsoft Copilot, Claude, Google AI Mode, Grok, DeepSeek, Meta AI and other models are available on Enterprise as an add-on, set up by the Depra AI team.
Each answer is stored and reported with the engine that gave it, so you see each engine on its own rather than one blended "AI" number. Depra AI also keeps the web searches ChatGPT ran behind an answer, labels every cited website by site type, marketplaces included, and checks which of 41 AI crawlers your robots.txt allows. The AI action plan turns this into steps, using a proprietary method developed by Zia Wasi. Depra AI is the worked example of that method: the team used it on its own brand and got Depra AI into ChatGPT's answers for AI-visibility questions in India, built without paid backlinks or paid marketing. The Depra AI case study shows what changed and what has not.