---
title: "Web Search"
description: "Use current web information in meinGPT: via the + menu in the chat, the meinGPT Web model, or the Research app."
canonical_url: "https://meingpt.com/en/docs/platform/web-search"
language: en
---

# Web Search

meinGPT can pull current information from the internet into your answers. There are **three ways** to do this. All of them use the same search under the hood (our search provider Linkup) — they differ in how you start them, which model answers, and where the result ends up.

## The three ways at a glance

- **Web search in the chat** — switchable on top of any model via the "+" menu.
- **The "meinGPT Web" model** — a chat model with web search built in.
- **Research app** — its own app in the sidebar, with history and "Deep Research".

The actual mechanics are described below under **[Web search in the chat](#web-search-in-the-chat)**; the other two ways refer back to it.

## Web search in the chat

The built-in **Web Search** tool gives the model access to current information from the internet.

### Enabling it

You enable web search in three places:
- in the **"+" menu** of the chat input (Tools → Web search, see [Chat input area](/en/docs/platform/chats/eingabefeld))
- in an **[Assistant](/en/docs/platform/assistants)** that has web search set up as a tool
- **per step** in a [Workflow](/en/docs/platform/workflows-create)

### Who creates the search query

You don't provide the search terms directly — the **AI model currently running your chat** does. It reads your prompt and formulates a suitable search query from it on its own. Your prompt is not taken over verbatim but translated into search terms.

### Two modes

Depending on the task, the model picks between two approaches on its own:
- **Direct search** — a single search query for quick, single facts.
- **Research agent** — for broader questions: a dedicated research step that runs several searches in sequence and reads individual pages in full before summarizing the result. This also applies when you provide a specific page (e.g. *"Read https://example.com/whitepaper and summarize the three key claims"*).

### Forwarding to Linkup

The search query created by the model is forwarded to our search provider **Linkup**, which performs the actual web search and returns the hits (GDPR-compliant; [More information](https://docs.linkup.so/pages/security-and-privacy/data-privacy)).

### From result to answer

Linkup returns the found content (titles, excerpts, sources) to the model. The model evaluates it and writes the final answer in the chat — with a preview of the sources (title, description, favicon) and references to the pages used.

## The "meinGPT Web" model

Instead of switching web search on, you can select the model **meinGPT Web** directly in the model picker. It is based on **Claude Haiku 4.5**, is fast and EU-hosted, and has web search **permanently enabled** — you can't turn it off here. The flow and result are the same as with [web search in the chat](#web-search-in-the-chat). Use it when you want quick, source-backed answers without toggling anything.

## Research app

In the sidebar under **Apps → Research**, web search is available as its own app — with a dedicated page instead of a chat toggle. Web search is **always active** here. The search runs via Linkup, and like the [meinGPT Web model](#the-meingpt-web-model) the same fast model answers (Claude Haiku 4.5). Two things differ from web search in the chat:

- **Deep research (toggle):** Unlike in the chat, where the model itself decides whether to search briefly or research more thoroughly, you decide via a toggle here. When enabled, a more thorough, multi-step research runs (more searches, more pages read) with a more detailed, cross-checked answer and more sources. It takes longer but often delivers better research results.
- **History:** Every research is saved. Under **Past research** you'll find your previous research again — searchable via the field above.

Open **Apps → Research** and type your question into the composer.

Turn on **Deep Research** if you want a more thorough, multi-step research instead of a quick answer.

Under **Past research**, below the composer, you'll find your previous research — even while a new search is running.

The app is labeled **Research** in the sidebar but internally uses the same web search function. The web search toggle inside the chat input is still called **Web search**.

## The three ways compared

All three use the same Linkup search — here are the differences at a glance:

| Criterion | Web search in the chat | "meinGPT Web" model | Research app |
|---|---|---|---|
| **Access** | "+" menu in the chat | Model picker | Apps → Research |
| **Web search** | switchable (on/off) | always on | always on |
| **Answering model** | freely selectable | Claude Haiku 4.5 | Claude Haiku 4.5 |
| **Research depth** | model decides | model decides | you – "Deep Research" toggle |
| **History** | in the chat | in the chat | own, searchable history |
| **Ideal for** | research right in the conversation, any model | quick, source-backed answers | standalone research with history |

## Good to know

Applies to all three ways:

**10 sources per search query**, not the whole internet — comparable to the first Google results page. For broader research, use the Research app with **Deep Research** or build a [Workflow](/en/docs/platform/workflows-create) with several research steps.

- Paywalled and login-protected pages are usually not accessible.
- Very recent events (last few hours) may be missing depending on indexing.
- JavaScript-heavy pages without server rendering sometimes return only skeleton content.

### Restricting sources to specific domains

For the web search on this page, there is **no platform setting** that restricts sources to specific domains. If you want to allow only certain sites, instruct the model directly in the prompt:

> *"Research GDPR compliance, use only heise.de and golem.de as sources."*

This works reasonably well in practice but is not a hard security boundary. For critical research, verify the result manually against the named sources.

If a soft recommendation in the prompt isn't enough, **Advanced Web Search** is a separate tool that enforces a domain whitelist technically — see [Advanced Web Search with a technical domain whitelist](#advanced-web-search-with-a-technical-domain-whitelist) below.

## Advanced Web Search with a technical domain whitelist

For cases where a domain restriction needs to be more than a recommendation — where it needs to be **technically enforced** — there's a separate, additionally activatable tool: **Advanced Web Search** (internally: `linkup-web-search`). Unlike the standard web search on this page, it passes the source selection directly to the search API instead of only recommending it via the prompt.

### Difference from standard web search

| | Standard web search | Advanced Web Search |
|---|---|---|
| Activation | "+" menu in chat, base tool in assistants/workflows | Additional tool, **Databases** tab in the tool settings of assistants/workflows |
| Domain restriction | Prompt only (soft recommendation to the model) | Via configuration fields, passed directly to the search API (technical boundary) |
| Other options | – | Image results on/off, maximum result count (1–50) |

### Adding it as a tool

1. Open the assistant editor or a workflow step.
2. In the **Tools** section, switch to the **Databases** tab.
3. Activate **Advanced Web Search**.
4. Open its configuration via the gear icon.

The configuration fields (including the domain whitelist) can be edited by anyone with edit or owner rights on that specific assistant or workflow — this isn't necessarily tied to the organization Admin role. Whether the tool itself is available organization-wide is a separate decision an Admin makes via tool management in settings.

### Configuration fields

| Field | Description |
|---|---|
| `Include Domains` | Comma-separated domains (e.g. `microsoft.com, learn.microsoft.com`). When set, search results come **exclusively** from these domains. |
| `Exclude Domains` | Comma-separated domains to exclude from search results. |
| `Include Images` | Image results on/off (default: on). |
| `Max Results` | Maximum result count, 1–50 (default: 10). |

If you leave `Include Domains` empty, the tool searches the whole web without restriction. For a hard allowlist, you need to fill in the field explicitly — a domain restriction in the prompt alone (see above) is **not** enough for this tool.

## Prompting tips for better web search results

| Principle | Description |
|---|---|
| **Search-engine perspective** | Formulate keywords you would also type into a Google search. Prefer "GDPR data processing agreement obligations 2026" over "Tell me about GDPR". |
| **Few examples in the prompt** | Few-shot examples can mislead the web search. Better to phrase the task directly. |
| **State scope and timeframe** | "Q3 2025 quarterly reports from DAX companies" yields better hits than "quarterly reports". |
| **Define a fallback** | "If no reliable sources are found, say so explicitly and do not speculate." |
| **Verify claims** | For important statements, add a verification step: "Check the statements above against the found sources and flag anything unsupported." |