Deutsch
Retrieval & grounding

LLM web search

LLM web search is a tool function that lets a large language model run web searches while it answers a request, read the results, and cite the pages it used. Without it, the model answers from its training data only.

How does LLM web search work?

In the API tools of the large providers, the model decides from the prompt whether to search. OpenAI describes this with tool_choice: "auto", where search is optional, and "required", which forces a search. Anthropic says Claude searches when a request depends on current information or topics outside its training data, and answers directly for stable knowledge.

OpenAI lists three modes: a quick non-reasoning lookup, an agentic search in which the model plans and repeats searches, and deep research across many sources over several minutes. With Anthropic’s tool, the API runs the searches, may repeat them within one request, and returns an answer with cited sources. Both providers return citations with URL and title, and both allow limiting results to or excluding specific domains.

Claude Code shows how little of a page may reach the model. According to Anthropic, its web search returns only the titles and URLs of results. When the WebFetch tool opens a page, it converts the HTML to Markdown, truncates long pages and, for most fetches, runs the content through a separate model call, so Claude works with that output instead of the raw page. An analysis by searchVIU (2026), based on reports about the leaked source code, concluded that content in the <head> and in <script> tags, JSON-LD included, is lost in this step.

Retrieval-augmented generation: retrieve first, then answerA question triggers a search in a search index or vector database. The best passages are added to the prompt together with the question. The language model writes the answer from them and points to the sources of the passages. The answer can only cite what was retrieved first.1 QUESTION2 RETRIEVAL3 CONTEXT4 ANSWERQuestionfrom a useror an appRetrievalsearch orvector search findsmatching passagesKnowledge sourcesearch index, web,vector databasePrompt with contextquestion + passages[1] passage · URL A[2] passage · URL B[3] passage · URL CLLMAnswertext withsource links[1][3]Citations point to retrieved passagesSchematic, based on Lewis et al. (2020) and Google’s grounding docs.
In retrieval-augmented generation, the system first retrieves matching passages and passes them to the language model as context. Google calls the link between model output and verifiable sources grounding. The answer can only cite what was retrieved first (an inference from the method, not a Google statement). Sources: Lewis et al. (2020): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv 2005.11401); Google Cloud: Grounding overview (Vertex AI); Google Cloud: Grounding with Google Search (Vertex AI).

What does it mean for SEO?

Cited pages come from search results the tool retrieves. A page that cannot be crawled or found in search is unlikely to be among them (assessment). Chat products built on these tools may behave differently from the raw API.

Example

A request such as “Which plans does tool X offer today?” triggers a search because prices change. The answer links the pages it drew on.

More on the searchVIU blog

Sources

  1. Web search - OpenAI API docs (as of 2026-10-07)developers.openai.com
  2. Web search tool - Claude API docs (Anthropic, as of 2026-10-07)platform.claude.com
  3. What the Claude Code leak reveals about AI grounding — and what it means for your GEO strategy – searchVIU blog (2026)searchviu.com
  4. Tools reference – Claude Code Docs (Anthropic)code.claude.com

Last updated: