How to Use AI for Research and Verify Its Sources (2026)

You are reading Chapter 20 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

To use AI for research properly you have to know one thing about every answer that comes back: whether it was fetched off the live web or written from the model's training. Only one of those can be checked, and the two arrive looking identical.

This guide shows you how to tell them apart, names the four categories of material your assistant structurally cannot reach, and gives you a four-minute check to run on any sourced report before you forward it.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Research is the first thing most senior people hand to AI. It is also where a wrong answer reaches the most people, because a sourced report gets forwarded.

The failure mode is not a tool lying loudly. It is an articulate, well-structured, fully cited document resting on sources nobody opened.

Confidence is not evidence that anything was looked up. A recalled answer and a researched one arrive in the same font and the same tone, and the assistant rarely says which one you just received.

Chapter 19 left you checking the number yourself rather than trusting the table it arrived in. If you skipped it, Chapter 19 builds that habit. This chapter applies the same habit to sources.

One boundary before we start. This chapter is about finding sources and checking whether they hold up. Interrogating a document you already have in front of you is a different job, and Chapter 18 covers how to analyze a long document with AI.

The first kind is recall. The model writes from what it absorbed during training, which stopped at a fixed date. No page was opened to produce it.

For a definition, a method, a piece of settled history, recall is fine and it is fast. For anything that moves, it is a confident description of the past. Ask about a regulation, a competitor's pricing, a person's current role, and you get the world as it was when the model was trained. It will not tell you that, because it has no way of knowing what it does not know.

The second kind is retrieval. The assistant runs actual searches, opens actual pages, and writes from what it just read. All four assistants can do this now, and each has its own logic for deciding when to bother. Claude has had web search on every plan globally since May 2025. Gemini leans on Google's index by default. Copilot reaches the web and, at work, your files and mail. ChatGPT searches when the question looks like it needs it, or when you say so.

So the practical question is not which assistant searches. It is how you tell, on any given answer, whether it did.

The tell is citations you can click. No links means it answered from memory, whatever it sounded like. Most assistants also print a line while they work naming what they searched for, and that line is worth reading rather than scrolling past.

Mixed answers are normal. One reply can be half fetched and half recalled, and the uncited half is the recalled half.

You can force the difference. "Search the web for this and cite your sources" works in all four assistants. Ask the same question twice, once plainly and once with that sentence attached, and read the two answers side by side.

The common mistake is assuming that a tool connected to the internet has the internet. It has the free, public, crawlable, already-indexed slice of it.

Your assistant does not tell you when it hits the edge of that slice. It keeps writing in the same voice, and the answer contains no marker where the evidence stopped. There are four things it cannot reach.

Journals, newspaper archives, the serious trade press in your sector. The assistant fetches the public page, which is a title, an abstract, two paragraphs and a subscribe button, then writes a paragraph that reads like a summary of the research. Sometimes it is right, because the abstract said so. Sometimes it is an inference from a headline.

Some of this is closed deliberately. Publishers have been shutting out AI crawlers for two years, and from 1 July 2025 Cloudflare began blocking them by default across the sites it fronts. The blocks are not always respected either: the Tow Center's early-2025 tests found tools correctly answering questions about articles from publishers that had explicitly disallowed them. So you cannot infer from a source appearing that it was fairly reached, and you cannot infer from a source missing that it does not exist.

Analyst research, market data terminals, licensed panel data, the paid database your team already subscribes to. None of it is on the open web, so none of it is in the answer. What you get instead is the press release that quotes one number from it, or a vendor blog citing a figure from three years ago. That is a real difference in quality and the report will not flag it.

Your CRM, your drive, your mail, last quarter's board pack. Unless you have deliberately connected them, they do not exist as far as the assistant is concerned. Connecting them is a real option and it changes the answer completely, which is what Chapter 26 covers.

Something published this morning may be hours away from being findable, and an assistant asked about it will either say nothing or fall back on recall. For anything breaking, use AI research for context, not for the newest fact.

Ask one follow-up every time: "What could you not reach, and what would change your answer?" Then read the source list for who is not on it.

All four assistants now have a mode that goes away for several minutes, reads dozens of pages and returns a long report full of citations. It is genuinely useful, and it is routinely spent on the wrong questions.

A search-backed reply runs a handful of searches inside a normal answer and comes back in seconds, using the obvious sources, stopping as soon as it has enough to answer. A deep research run plans, searches repeatedly, reads dozens of pages, then writes a structured report with a bibliography. It takes roughly 5 to 30 minutes, unattended, and it is metered.

ChatGPT calls it Deep research: roughly 5 to 30 minutes, a small monthly allowance free and a larger one on paid plans. Claude's is Research, on paid plans only, searching the web plus your Google Workspace files if you connect them. Gemini's Deep Research shows you an editable plan first, then usually works for 5 to 10 minutes. Copilot's Researcher sits under Agents, needs a paid licence, and reaches your work files, mail and meetings as well as the web.

Gemini shows you that plan before it starts, and people click past it. Edit it instead. Twenty seconds spent narrowing the plan is worth more than any prompt wording.

The runs are metered, so you are spending one of a small number of them. Spend it on ground you do not already know: a market you have never sold into, a regulation that just landed in your inbox, a vendor category with eleven players you cannot name. In those cases you do not yet know the right search terms, and what the report gives you is not its conclusion. It is the names, the factions, the terminology, and the four sources you should now read yourself.

Where it is oversold is your own field. The assistant has no sense of your profession's hierarchy of evidence. It weighs a well-optimised content marketing page and a regulator's own notice at roughly the same value, because on the open web they look similar. Where you know the field, you already know which sources count and it does not, even though it reads more pages than you will.

Because these modes read more, people assume their citations are safer. Work published in April 2026 across more than 221,000 cited links found the opposite: deep research agents produce more citations per question than ordinary search-backed chat, and hallucinate them at a higher rate. More footnotes means more citations to open, not fewer.

If you would rather run research in a tool that searches on every question instead of deciding for itself, that is a different setup, and our 10 prompts for turning Perplexity into a personal assistant covers it.

The output is fast, articulate, well-organised and carries no consequence for being wrong. You already know how to read work like that. The only difference is that the citations have to be opened rather than trusted, because this junior cannot be asked why and will not be embarrassed next quarter.

Four steps, in this order. It takes about 4 minutes.

Ignore the rest. A long report usually turns on two numbers and one assertion, and those are the only lines worth your time.

Read what it actually says. This is the step everybody skips and the one that catches the most. A citation can fail in four ways: the link is dead, the link was invented and never existed, the link is real but about a different thing, or the link is real and relevant and simply does not say what the sentence above it claims. The last one is the most common, and the only way to catch it is to open the link. The first two have been measured: across more than 221,000 cited links studied in April 2026, 3 to 13% were fabricated outright and 5 to 18% did not resolve at all.

Count origins, not links. A confident synthesis with five citations can rest entirely on one weak source, because the other four repeat it or address something adjacent. Syndicated copies of one wire story are one source. The Tow Center's early-2025 test of 1,600 queries across eight AI search tools found them citing syndicated copies rather than originals as a matter of course, and wrong more than 60% of the time overall. The same study found the tools almost never admit doubt: one of them flagged low confidence 15 times in 200 answers and never once declined to answer.

Ask what it did not find, what is behind a paywall or a subscription, and which of these claims rests on a single source. A good assistant answers that honestly, and the answer is frequently more useful than the report, because it tells you where the real work still is.

Some research should not be handed over at all. Anything where the answer depends on knowing which sources are respected and which are marketing. Anything regulated, contractual or legally consequential, where a fabricated citation is a professional problem rather than an inconvenience. Anything where the cost of being wrong lands on someone other than you. The assistant is still useful there for the reading list and the vocabulary. The judgement stays with a person who knows the field.

The chapter closes with 3 exercises that take about 20 minutes. The short version: run the four steps with a timer on the next research output that reaches you, from any source, and you will find something in the first one.

Before: a sourced report arrives, it reads well, the citations are there, and you forward it. You have no way of knowing whether any page was opened.

After: you can tell a researched answer from a recalled one, you can name the four categories your assistant structurally cannot see, and you can check a sourced report in about 4 minutes.

Two things carry into Chapter 21: the research output you checked, and the list of what your assistant could not reach. Chapter 21 turns the assistant against your own conclusions rather than letting it agree with them.

If you want better wording for any of this, the 26 principles of prompt engineering is the reference, our free AI cheat sheets are the one-page version, and the rest of the syllabus lives in the AI Central Library.

Look for citations you can click. No links means it answered from training, whatever it sounded like. Most assistants also show a line naming what they searched for while they work. To be certain, add "search the web for this and cite your sources", which works in all four assistants.

No. It fetches the public page, which is usually a title, an abstract, two paragraphs and a subscribe button, then writes as though it read the whole thing. Journals, newspaper archives and the serious trade press are the material most professionals would actually cite, and most of it sits behind that wall.

Not automatically. Work published in April 2026 across more than 221,000 cited links found that deep research agents produce more citations per question than ordinary search-backed chat, and hallucinate them at a higher rate. The long mode is a survey of ground you do not know, not a more accurate search.

Find the 2 or 3 claims the conclusion rests on, open the citation under each one and read what it says, count independent origins rather than links, then ask what the assistant could not reach. It takes about 4 minutes.

Anything where the answer depends on knowing which sources are respected and which are marketing, anything regulated, contractual or legally consequential, and anything where the cost of being wrong lands on someone other than you. It is still useful in those cases for a reading list and the vocabulary.