Research anything, and check the answers.

AI research tools are excellent at the first pass: finding sources, mapping a topic and drafting a summary with citations in minutes. The final pass is yours. Open the sources, confirm each one says what the AI claims, and go to the original for anything that matters.

That split is not caution for its own sake. In a 2025 study by 22 public broadcasters, 45% of AI assistant answers about the news had at least one significant issue. This guide shows how to brief a deep research tool, how to check what comes back, and the mistakes to expect.

01What AI research tools do now

The three big assistants all have one, with different reach into your own data:

Limits apply. Gemini, for example, caps research requests per day and gives Google AI Pro and Ultra subscribers more. If your question is about your own documents rather than the web, a tool that answers only from files you provide is often a better fit. Our guide to how RAG works explains the difference.

02How to brief an AI research tool

A deep research run is only as good as its brief. A one-line question gets a generic survey of whatever ranks well in search. A real brief gets you closer to what a good research assistant would bring back.

• The decision. Say what the research is for. “Should we enter the German market next year?” steers better than “German market info”.

• The scope. Time range, countries, what is out of scope.

• Source rules. Prefer primary sources, name the ones you trust, and ask it to skip content farms and unsourced listicles.

• The output. A short summary, then a table of claims with the source, date and exact supporting quote for each.

• Honesty. Ask it to flag conflicting sources and anything it could not confirm.

Research question: [your question].
Why I need it: [the decision this informs].
Scope: [time range, countries, what to leave out].
Sources: prefer primary sources, such as official statistics, company pages, papers and court or government documents. Avoid unsourced blog posts and aggregators.
Output: a 200-word summary, then a table with one row per key claim: the claim, the source, its publication date and the exact sentence that supports it.
Finish with a list of points where sources disagree or where you could not find strong evidence.

If the tool shows a plan before it starts, as Gemini does, read it. Removing one wrong angle at the plan stage saves you checking a whole wrong section later.

03How to check AI citations

A citation proves nothing until you open it. In a Columbia Journalism Review test of eight AI search tools, more than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs. Check the sources your conclusion rests on:

1. 01

   Open it

   Make sure the link loads and the page exists. A dead link or a homepage in place of an article is a warning sign.

2. 02

   Find the exact passage

   Search the page for the number or phrase the report cites. Check that the source says it with the same scope: the same year, country, sample and definition.

3. 03

   Check the date

   Research tools happily present a 2021 number as current. Look for a newer edition of the same survey or dataset.

4. 04

   Check who is speaking

   Is this the original, or a copy? The same study found AI tools often cited syndicated or republished versions instead of the original publisher.

5. 05

   Separate claim from inference

   The source may report a fact, while the AI draws a conclusion from it. Keep the fact, and judge the conclusion yourself.

A second model can speed this up, but it cannot replace opening the page, because it can misread sources too.

Below is a research report with citations.
For each factual claim, make a table: the claim, the cited source, the exact sentence from that source that supports it, and whether that sentence fully supports the claim, partly supports it or does not support it.
If you cannot access a source, say so rather than guessing. List the claims I should check by hand first.

04Read laterally before you trust a source

Checking that a page says something is half the job. The other half is whether the page deserves your trust. This habit is called lateral reading: instead of reading one site top to bottom, you open new tabs to see what others say about the source and the claim.

Mike Caulfield’s SIFT method turns it into four moves:

• Stop. Before you use a source, ask whether you know what it is.

• Investigate the source. Find out who is behind it and what their expertise or agenda might be.

• Find better coverage. Look for trusted reporting on the claim itself rather than relying on one source.

• Trace claims to the original. Follow quotes, numbers and images back to where they first appeared, and check they were presented accurately.

The last move matters most with AI, because each summarizing step can drift further from the original.

05When to go to primary sources

For a quick orientation, a good AI summary is enough. For anything you will publish, decide on or pay for, go to the source the claim came from.

06The usual ways AI research goes wrong

The failures are predictable, which makes them easier to catch:

• Invented or broken citations. A plausible title with a dead link, or no such paper at all.

• Real source, wrong claim. The page exists but says something narrower, older or different.

• Stale facts. Last year’s price, an old version or a superseded law, presented as current.

• Blended numbers. Figures from different years or definitions merged into one neat total.

• False consensus. A confident summary where the sources actually disagree.

• Planted instructions. OpenAI’s own docs warn that a web page can hide text with instructions that a research model may follow.

These errors come from how language models work, which our guide to why AI makes things up explains. If you publish on the web, the same tools now decide which pages get quoted; see how to get cited by AI search.

07FAQ

Read next: why AI makes things up, and how to catch it, or how RAG answers from your own documents.