Do LLMs Build Their Own Search Index or Rely on External Search Engines?

Are LLMs search engines?

Not the models, but the products are partly. A language model on its own answers from what it learned in training, with no index and no ranking. The assistants people use, such as ChatGPT, Claude, Gemini and Perplexity, add a search layer. When a question needs current or specific information, the assistant writes one or more search queries, sends them to an index, reads some of the results, and writes an answer with citations. That process of splitting a question into several searches is known as query fan-out, covered in what query fan-out is and how LLMs use it.

So each answer can draw on four sources:

1. Training data: what the model learned before its cutoff.
2. An external search engine: Google, Bing or Brave.
3. The company's own index: built by its own crawler.
4. Live fetches: the assistant opening a specific page during the conversation.

Which search providers do LLMs use: Google, Bing, or Brave?

Based on company statements and published analysis:

Two cautions. First, companies rarely publish the full mix, and it changes. OpenAI describes ChatGPT search as using third-party search providers plus its own crawler, without naming them all. Second, overlap with an engine is not proof of exclusive use. It does, however, tell you which index to be strong in.

Which search index do ChatGPT, Claude, Gemini and Perplexity use for recommendations?

When someone asks for a recommendation, such as the best project management tool or a dentist nearby, each assistant blends searched results with what its model already knows:

• ChatGPT searches through its providers and its own index, and opens pages live through ChatGPT-User. Studies disagree on how closely its citations match Bing: one 2025 study found 87% within Bing's top 20, a larger one found about 8% overlap with Bing's top 10 and the same with Google's. For shopping, research on its product carousels found that most products matched Google Shopping data, so product visibility depends on more than one source.
• Claude searches through Brave and Anthropic's own index, and fetches pages through Claude-User. A brand strong in Google but weak in Brave can be missing from Claude's recommendations. An analysis of roughly 35,000 citations from Claude found that 79% came from pages in Brave's top 10 results, compared with 34% from Google's top 10.
• Gemini draws on Google's index, and on Google's other data, such as Maps listings for local questions.
• Perplexity searches its own index and fetches pages live through Perplexity-User.

Without a search, all four fall back on training data, which is why established brands with a lot of coverage are recommended more often in general questions.

Why do answers differ between assistants?

Because almost every step differs. Each assistant writes its own search queries from the same question, sends them to a different index, and reads a different handful of results. The indexes themselves differ in which pages they hold, how fresh they are and how they rank them. Some assistants open pages live and others rely on stored snippets. And when no search happens, each model falls back on training data gathered at different times from different sources.

The practical result is that a brand can be recommended by one assistant and ignored by another for the same question. Measuring only one assistant, or checking only Google, gives a partial picture. Testing the same buyer questions across the assistants your customers use, and comparing their cited sources, shows where the gaps actually are. Which assistants to prioritise is covered in which LLMs brands should monitor.

What does this mean for being found?

Treat each index as a separate audience:

• Google: Search Console for indexing and coverage; Google's AI features depend on it.
• Bing: Bing Webmaster Tools and IndexNow. Bing feeds Copilot and may be one of ChatGPT's providers.
• Brave: search your key pages on Brave Search to confirm they are indexed. Brave offers a URL submission form that re-fetches pages it already knows.
• Own crawlers: allow OAI-SearchBot, Claude-SearchBot and PerplexityBot in robots.txt, and make sure your CDN or firewall lets them through. Blocking the search crawlers keeps your page text out of those assistants' own indexes and makes you much less likely to appear in their answers, although a blocked page can still show up as a bare link or headline.
• Live fetchers: ChatGPT-User, Claude-User and Perplexity-User open pages during conversations. Anthropic notes that blocking Claude-User may reduce your visibility for user-directed web search.

Indexing and canonical problems that stop citations are covered in how indexing, canonical and schema issues prevent LLM citations.

How can you tell which index an assistant used?

A simple test, the same one used to confirm Claude's use of Brave:

1. Ask the assistant a question that makes it search, and note the sources it cites.
2. Run the likely search query in Google, Bing and Brave.
3. Compare the cited pages with each engine's top results. The engine whose results overlap most is probably the one behind that answer.
4. Repeat for five to ten questions in your category, since one query can mislead.

If an assistant keeps citing pages that rank in Brave but not Google, improving your Brave presence will do more for that assistant than more Google work.

What might this look like in practice?

Consider an illustrative HR software company, Clearwell. It ranked well in Google and appeared in AI Overviews, but rarely in Claude's answers. Running its buyer questions in Brave showed that its comparison pages were missing from Brave's index, and that two review sites dominated Brave's results. It made sure Googlebot could crawl those pages, since Brave's crawler skips anything Googlebot cannot, earned listings on both review sites, and confirmed Claude-SearchBot was allowed through its CDN. Over the next two months, Claude began citing Clearwell's pages for several comparison questions.

What do people ask about LLMs and search indexes?

Does ChatGPT use Google or Bing?

OpenAI says ChatGPT search uses third-party search providers and its own crawler. Evidence on Bing is mixed, so check indexing in both Bing and Google.

Which search engine does Claude use?

Brave Search, which Anthropic listed as a subprocessor in 2025, plus Anthropic's own Claude-SearchBot index.

Does Perplexity have its own index?

Yes. Perplexity says its index covers hundreds of billions of webpages.

Do I need to optimise for each search engine separately?

Mostly the same work helps everywhere, but check indexing in Google, Bing and Brave, and allow each assistant's search crawler.