The right tool depends on the task, not the brand name

Start here

Choosing between assistants is usually framed as a brand choice, but the vendors' own documentation describes different jobs. Analyzing a stack of a reader's own documents and citing them, reasoning through a multi-step problem at length, and running privately on a personal machine are separate capabilities, and confusing them means picking a tool that cannot do the job at hand.

What the documents say

Google's page for Gemini Notebook, the product long marketed as NotebookLM, describes it as an 'AI research tool and thinking partner that can analyze your sources' — built around a user's own uploaded material rather than the open web. Anthropic's documentation for extended thinking describes a mode where the model thinks against a set token budget before its final answer, useful for workloads needing 'more thorough analysis' on complex problems, at the cost of added latency. Ollama's own project documentation shows a different design: its programming interface runs at a local address on the user's own computer rather than a remote server, shown in its installation and API instructions rather than claimed as a privacy feature.

Check this

This is editorial guidance built from the three documents above. Before opening any assistant, name the job in one sentence — summarizing your own files, reasoning through a multi-step problem, or working without sending anything to a remote server — and match it to the capability the vendor's page describes, not the tool's general reputation. For grounded research, check whether the assistant cites the uploaded material rather than general knowledge. For a hard multi-step problem, check whether a reasoning mode exists and whether the wait suits the task. For anything that should stay on-device, check that the tool genuinely runs locally, as Ollama's own address shows, rather than merely promising privacy in marketing copy.

What holds and what fails

This distinction holds when a reader can name the task clearly enough to match it to a capability. It fails when the task itself is vague, since a grounded tool, a reasoning tool and a local tool can each produce a plausible answer to an ill-defined question, and the choice of tool will not fix an unclear brief. It also depends on the documentation staying current: these capabilities reflect each page as retrieved on 16 September 2026, and product names, modes and installation steps change. This is an editorial framework built from vendor documentation, not a ranking of one tool over another.

• Name the task in one sentence before opening any assistant.
• Check whether a grounded tool actually cites your own uploaded material.
• Check that a local tool's data genuinely stays on your device before trusting it with anything sensitive.

The brand on the login screen says less about fit than the job each vendor's own documentation says the tool is built to do.