How do you use Gemini 3.1 Pro for research?

Google released Gemini 3.1 Pro in preview on February 19, and it is now in the Gemini app, in NotebookLM for paying subscribers, and in Google AI Studio for developers. For research, three of its traits matter. It reasons through a problem before answering. It can take in up to about a million tokens at once, roughly several long books' worth of text. And on BrowseComp, a test of finding hard-to-locate facts on the web, Google reports a score of 85.9%. Those strengths make it a capable research assistant. They do not make it a source.

The distinction shapes everything below. A research assistant helps you find, organize and question information. A source is the document where the information actually lives. Your job is to use the first to reach the second.

Where to use it

In the Gemini app, Google AI Pro and Ultra subscribers get higher usage limits for 3.1 Pro; in NotebookLM it is available only on those plans.

A worked research session

Suppose a homeowner in a cold region wants to know whether an air-source heat pump can heat a house through a winter with frequent nights below minus 25 °C. Here is how to run that as research, not as a single question.

Step 1: frame the question and ask for a plan first.

I am researching whether an air-source heat pump can be the main heating system for a 1,600 sq ft house where winter nights often drop below -25 °C. Before answering, give me a research plan: the sub-questions that matter, the kinds of sources that would answer each (government energy agencies, manufacturer specifications, independent field studies), and what evidence would change the conclusion. Do not answer yet.

Asking for a plan exposes the model's assumptions early, when they are cheap to correct. If it forgets something you care about, such as backup heat or electricity costs, add it now.

Step 2: ask for the answer, with rules about evidence.

Now follow the plan. For every factual claim, give the source with a link and say whether it is a government body, a manufacturer, a news outlet or a study. Keep three things separate: what the sources say, where they disagree, and your own reasoning. If you cannot find a source for something, say so instead of filling the gap.

Step 3: push on the weak points. Ask what the strongest evidence against its conclusion is, which claims rest on a single source, and what a skeptical expert would question. A reasoning model handles these follow-ups well because it can re-examine its own chain of argument.

Step 4: move your best sources into NotebookLM. Once you have the five or six documents that matter, such as a government guide and two field studies, upload them to a notebook. NotebookLM answers from those files only, and its replies link to the passages they rely on, so you can ask detailed questions and see exactly where each answer comes from.

How do you check its sources?

Open every link that supports a claim you plan to rely on. This sounds obvious, and it is the step most people skip. Check three things on each page.

Does the page exist and say what was claimed? Find the actual sentence or table. Models sometimes cite a real page for a point it doesn't make, or blend two sources into one. Use your browser's find function to search for a key number or phrase.

Who published it, and when? A manufacturer's specification is useful for what a product is rated to do, less so for how it performs in a real house. A study from ten years ago may describe older equipment. Look for a date, an author and an organization.

Can you confirm it somewhere else? For any claim that drives your decision, find a second independent source. Researchers call this lateral reading: instead of judging a page by how professional it looks, leave it and see what other reliable sources say about the same point.

Keep a short log as you go: claim, source, checked yes or no. Many citations will hold up. Some may prove weaker than they sounded, and now and then one may not say what the model reported. That possibility is why the check matters. Gemini 3.1 Pro is a strong reasoner, but Google itself presents it as a preview, still being refined before general release, and no benchmark score turns a summary into evidence. The evidence is always in the source.