How do you check an answer from an agent that browsed for you?

A research agent, such as ChatGPT's deep research or agent mode, reads dozens of pages on its own and hands you a tidy report with links. The links make the report look checked. They are not proof. They are a map showing you where to look, and the checking is still your job.

OpenAI says this plainly about its own tools: deep research can state wrong facts or draw wrong conclusions, has trouble telling an authoritative page from a rumor, and often fails to say when it is unsure. Independent testing points the same way. In a 2025 study by Columbia University's Tow Center, eight AI search tools were asked to identify the source of short news excerpts, and together they got more than 60 percent wrong, usually without any hedging. Several also linked to copies of articles on other sites, or to addresses that led nowhere.

Why a cited report can still be wrong

Four problems come up again and again, and each one calls for a different check.

There is a fifth risk that is specific to agents. A web page can contain hidden text written to steer the agent, a trick called prompt injection. OpenAI names it as a key risk for ChatGPT agent. You cannot see it in the final report, so a surprising claim, or an odd recommendation to visit a site or buy something, deserves extra suspicion.

A worked audit in five steps

Imagine you asked an agent: "Should our community center switch from oil heat to a heat pump? Summarize costs, rebates and performance in cold weather." It returns two pages with eleven links. Here is how to audit it in about twenty minutes.

1. List the claims that matter. Ignore the background and pick only the sentences your decision depends on. Ask the agent to help:

From your report, list every factual claim that affects the decision
(numbers, dates, rules, eligibility). For each one, give the exact URL
that supports it and quote the sentence from that page. If no page
supports it directly, write "no direct source".

2. Open each supporting link. Check that the page loads, that the quoted sentence is really there, and that it says the same thing as the report.

3. Check the date and the owner. A rebate page from two years ago, or a reseller's blog repeating a government program, is a warning sign. Look for the government, manufacturer or research page itself.

4. Grade each claim. Keep a short table. The results below are illustrative.

5. Send the gaps back. Paste your table into the conversation and ask the agent to fix only the weak and unsupported lines, with new sources. Then check those new links the same way.

Signs that a report needs a closer look

Some patterns are worth recognizing at a glance. Every claim comes with a neat percentage, but the links lead to general overview pages. Several sentences cite the same single page. The report never says "I could not find" or "sources disagree" even on a hard question. A link's address doesn't match the publisher named in the text. None of these proves an error, but each tells you where to start.

It also helps to ask the agent, before it starts, to separate what it found from what it inferred: "Mark any conclusion you drew yourself, rather than read on a page, with the word inferred." That one instruction makes the audit much faster.

Which links should you open yourself?

Open every link behind a claim you will act on, repeat to others, or spend money on. That usually means numbers, prices, deadlines, legal or medical statements, and anything about a named person or organization. Open the link behind any claim that surprised you, too, since surprises are where invented or stretched facts tend to hide.

You can usually skip links behind background material you already know, and general explanations that several independent sources in the report agree on. If you are short on time, rank the claims by how much harm a mistake would cause and work down the list.

When you open a link, read the relevant paragraph, not just the headline, and note the page's date. If the cited page is a copy or a summary, follow it to the original. And be careful about where you click: when a link points to an unfamiliar site, a download, or a page asking you to sign in, check the address first rather than trusting it because an agent put it there.