This guide covers one narrow decision: how to answer a factual question you plan to act on, repeat, or use in another piece of work. The deciding factor is not which interface feels faster; it is whether you can identify the responsible source, open it, and confirm the exact claim. For broader task routing between retrieval, drafting, and synthesis, see AI vs Google.
The test that matters: can you click it
Plain chat - a chatbot with no live web-search or retrieval mode active - generates an answer without guaranteeing a connection to a checkable source at the moment you ask. Search, and chat with search enabled, can return links you can inspect. That makes a citation auditable, not automatically genuine: a link can be dead, outdated, low-quality, or unrelated to the sentence it appears to support.
This is a real, measured gap, not a hypothetical one. A joint study by the BBC and the European Broadcasting Union - published as the News Integrity in AI Assistants Report - tested four major AI assistants - ChatGPT, Copilot, Gemini, and Perplexity - by having professional journalists across 22 public broadcasters in 18 countries evaluate more than 3,000 responses to news questions. It found 45% of AI answers had at least one significant issue, and sourcing was the single biggest cause: 31% of responses had serious sourcing problems, including missing, misleading, or incorrect attributions (BBC/EBU media centre summary, October 2025). That study tested assistants answering questions about published news content - exactly the kind of factual question this article is about.
A confident tone from a chatbot is not evidence that a fact is grounded in a real, current source. In the BBC/EBU study above, 31% of responses had serious sourcing problems - missing, misleading, or incorrect attributions - even when the assistants were answering questions about published news.
A simple decision rule
Ask one question before you ask anything else: will I act on this, repeat it to someone, or need to defend it later?
If no - for example, brainstorming labels or exploring possible questions where no factual claim will be used - plain chat may be sufficient. If the conversation produces a fact you later plan to repeat or act on, move that claim into the verification path.
If yes - a health question, a legal or financial detail, something you will tell a colleague or forward to a group chat, anything time-sensitive - require an openable citation. That means either:
- Search directly, and read the actual page yourself; or
- Chat with search/browsing mode turned on, then open the specific citations it returns rather than trusting its summary of them.
Plain chat without search mode should not be the last step for anything in the second category, regardless of how specific or confident the answer sounds.
Why “it sounded specific” is not the same test
A model asked a factual question will often produce a specific-sounding answer - a name, a date, a percentage - whether or not it actually has reliable grounding for that specific detail. This is the same mechanism covered in why AI sometimes gives confident wrong answers: the model is built to produce a plausible continuation, and plausible and correct frequently overlap, but overlap is not a guarantee. Specificity is not a proxy for accuracy; a source you can open is.
Search mode is not automatically safe either
Turning on search or browsing mode adds clickable links, but it does not establish that the synthesis is accurate. An earlier BBC-only test that prompted assistants to use BBC News articles as sources where possible still found errors in summaries, quotations, and framing (BBC, “Groundbreaking BBC research,” 2025). The later BBC/EBU international study found the same family of sourcing failures at a larger scale. Open the citation for each material claim and confirm that the source states what the answer says, in context. Checking an AI news summary against its primary source covers that step in detail.
When the right tool is neither
Some questions need a workflow rather than one search result or chat answer. For a comparison, landscape, or briefing built from many sources, Deep Research mode is one candidate because it can produce a source-linked report for audit; it still requires coverage and citation checks. If the claim began as a viral post, use the viral-claim tracing workflow, which starts from the original post and follows the evidence chain.
A worked comparison
| Question | Plain chat okay? | What to do instead if not |
|---|---|---|
| ”Explain how compound interest works” | Yes for an introductory explanation | Verify the formula and inputs before calculating or making a financial decision |
| ”What is the current EU AI Act deadline for X” | No | Search or chat-with-search, then open the regulation text |
| ”Summarize this article I’m pasting in” | Yes - source is what you gave it | - |
| ”What did [named study] actually find” | No | Search for the paper directly, read the abstract yourself |
| ”Brainstorm gift ideas for my dad” | Yes - no fact to get wrong | - |
| ”Is this claim in my group chat true” | No | Use the viral-claim source-trace method |
What “openable” does not guarantee
Even a clickable, real link can fail you in ways worth knowing about. A source can be genuine but outdated - accurate when written, wrong by the time you read it. It can be genuine but low-quality - a real page that is itself speculative, opinionated, or poorly sourced. And it can be genuine but irrelevant to the specific claim you are checking, even if it is topically related. Opening the link is the necessary first check, not the only one; skim for the publication date, the author or outlet’s track record, and whether the page actually addresses your specific question rather than something adjacent to it.
Common pitfalls
- Assuming internet access is active because the app supports it. The mode, model, workspace, and interface can change whether search is used. Check the answer for actual source links and inspect them.
- Trusting a citation without opening it. The BBC/EBU numbers above describe assistants that were specifically asked to cite sources and still got the sourcing wrong in nearly a third of cases.
- Using plain chat for anything you will repeat. If you would be embarrassed to be wrong in front of the person you are telling, that is the test for needing an openable source.
- Assuming search mode fixes accuracy, not just sourcing. It gives you a real link. You still have to read it.
Try it today
Before your next factual question - one you plan to act on or repeat - identify the source class that should control the answer: the regulator, law, original study, manufacturer, provider, venue, or another primary authority. Search directly or use a search-enabled chat to locate it, then open the citation for every material claim before relying on the summary. Use the search vs chat decision card to apply the same rule consistently.
