Researching with AI without being fooled

In short

For specific questions, use an assistant with web search (Perplexity, ChatGPT, Claude, Gemini): answer in seconds, with links.
For a complete picture, use Deep Research: 5-30 minutes, dozens of sources, a cited report.
To study your own sources, use Gemini Notebook or Projects: the AI answers only from what you upload.
A citation isn’t verification: always open the sources for data you’ll use.

Three ways to search, three tools

What you need	Tool	Time	Example
A precise, up-to-date answer	Web search (Perplexity, ChatGPT, Claude, Gemini)	Seconds	“What’s the revenue cap for Italy’s flat-rate regime in 2026?”
A broad picture of a topic	Deep Research	5-30 minutes	“State of the art of AI in SME accounting in Europe”
Working on your own documents	Gemini Notebook, Projects	Minutes	“What do these 12 PDFs say about termination clauses?”

Web search: fast and verifiable

When the assistant searches the web, it runs several queries, opens pages, reads passages and writes an answer with links. 2026 models decide on their own when to search, but for current-events questions it pays to ask explicitly: “search the web”.

Perplexity was built for this and remains the fastest for specific questions. ChatGPT, Claude and Gemini do the same inside the chat, with the advantage that you can continue the work (write, calculate, compare) in the same conversation.

Deep Research: a researcher working for you for half an hour

With Deep Research the assistant plans the research, consults dozens or hundreds of pages, compares sources and produces a structured report. Useful for:

market and competitor analysis;
state of the art on a technical or regulatory topic;
preparing an article, thesis or presentation.

Limits: it’s slow, sometimes verbose, and tends to weigh authoritative sources and blogs equally. It works much better if you tell it which sources to prefer and what to ignore.

Gemini Notebook and Projects: research on your sources

When you already have the material (PDFs, articles, videos, notes) the best move is loading it into a tool that answers only from those sources. Gemini Notebook (the new name of NotebookLM) cites the exact passage for every claim and can create summaries, mind maps, quizzes and podcast-style “audio overviews”. ChatGPT and Claude Projects also let you work on a set of documents.

A real example

To prepare this site I loaded the videos from my AI playlist (almost 300 sources) into a notebook and asked: “Extract the 20 most useful ideas, grouped by theme, saying who says them and in which video”. In two minutes I had a map to start from. But then I checked the attributions: a couple were wrong (right ideas attributed to the wrong person). Verification is still your job.

How to write a research request

A good research request has five parts: precise question, scope, sources, format, level of certainty.

Specific research

Search the web: [precise question]. Scope: [country], rules in force as of [month year]. Use official sources first (tax authority, official journal, government sites), then professional ones. Answer in max 150 words, with a link for every claim. If sources disagree, tell me.

Deep Research

Do in-depth research on [topic] for [who will read it and why]. Questions to answer: 1) … 2) … 3) … Sources: prefer [institutional sources, peer-reviewed studies, industry reports]; ignore promotional content and undated articles. Period: from [year] to today. Format: 10-line executive summary, then one section per question, a table with key data and the year of each, and a list of what could not be verified.

Checking sources in five minutes

You don’t need to check everything. Check what you’ll use: numbers you’ll quote, rules you’ll apply, claims you’ll base a decision on.

Open the link. Does the page exist? Does it really say what the AI attributes to it? Sometimes the link is right but the data comes from elsewhere on the page, or isn’t there at all.
Check the date. A 2023 figure on AI prices is archaeology.
Who says it? Primary source (the law, the official announcement, the paper) or secondary (a blog summarising)? For important things, go back to the primary.
Announcement or rumour? In AI many “news items” are features spotted in testing, like those TestingCatalog finds in apps. Valuable, but treat them as such until the company announces them.
A second opinion. For important decisions, ask a second assistant or find a second independent source.

Rules and deadlines

For taxes, deadlines and legal obligations, AI is a great starting point to understand, but the final answer must come from the official text or a professional. Rules change often and old versions stay online for years.

Recommended sources on AI

To follow the AI world specifically, these are the sources I check daily (and that Lumen’s chatbot consults in real time):

Artificial Analysis for model benchmarks, prices and speed;
Arena for rankings based on people’s preferences;
There’s An AI For That for finding tools for a specific task;
TestingCatalog for news and upcoming features.

They’re described on the Sources & method page.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.