Web search vs Deep Research

Why looking things up matters at all

A language model is trained on data up to a cutoff date, after which it knows nothing unless it looks. Ask it about a recent change (a new rule, this quarter's figures, who holds a role now) and from memory alone it may give you a confident answer that's simply out of date, with no signal that it's stale. Connecting it to live sources fixes the currency problem and, done right, gives you links you can check. This is often called grounding: tying the answer to real, retrievable sources rather than the model's memory.

ChatGPT offers two grounded modes, and they suit different jobs.

Web search: fast and current

Web search is the quick option. ChatGPT runs a few searches, reads the top results, and answers in seconds with links to its sources. It's the right tool for a factual question with a fairly direct answer: "What's the current UK VAT registration threshold?", "Has this company announced its results yet?", "What are the opening hours for X?" You get a current answer, fast, with sources you can click to confirm.

Its limits are the flip side of its speed. It reads a handful of pages, not dozens; it doesn't dig deep or cross-check exhaustively; and if the answer is contested or scattered across many sources, a quick search can land on one and present it as settled. For a simple fact, that's fine. For a nuanced question, it's thin.

You usually invoke it with a Search button or by asking a question that obviously needs current information; ChatGPT will often decide to search on its own. Either way, the tell is the citations: an answer with linked sources has looked something up; one without is coming from memory.

Deep Research: slow and thorough

Deep Research is a different beast. Instead of a few quick searches, it runs an extended investigation (searching, reading, following leads, and cross-referencing across many sources) then produces a structured, cited report. This takes minutes, not seconds, sometimes many minutes, and that's the point: you're trading time for depth. It's built for questions a quick search can't do justice: "Compare the three main options for X, with the trade-offs and current pricing of each", "What's the current regulatory position for Y in the UK right now, with sources", "Build me a briefing on this market with the key players and recent developments".

Two features make it worth understanding rather than just clicking.

The plan-review step

Before a Deep Research run charges off, it will often ask you clarifying questions and propose a plan: the angles it intends to cover, what it will look for. This is a genuine control point, not a formality. Because the run is long, a wrong assumption at the start wastes minutes and buries the answer you wanted under the one you didn't ask for. Read the plan. If it's about to research the wrong market, the wrong country, or the wrong definition of your question, correct it now: "focus on the UK only", "I mean B2B, not consumer", "skip the history, I only need the current position". Steering at the plan stage is far cheaper than re-running.

Do a deep research report on options for a mid-size UK company to run a staff feedback survey. Scope it to the UK, to tools suitable for a non-technical HR team, and to current pricing and data-protection considerations. Before you start, show me your plan and the sources you intend to lean on so I can adjust it. Output: a comparison table plus a one-paragraph recommendation.

Source restriction

You can also constrain where it looks. If you only trust official or primary sources, say so: "use official government and regulator pages only", "prioritise the companies' own documentation over blog write-ups", "ignore forums and marketing pages". Restricting sources trades some breadth for reliability, which is usually the right trade for work you'll act on. It also makes the citations easier to check, because you already trust the kind of place they come from.

For this research, use primary and official sources only: government, regulators, and the providers' own documentation. Do not rely on marketing pages, forums or opinion blogs. Where sources disagree, show both and say which is more authoritative and why.

Which to reach for

A simple test. If you could imagine getting the answer from one or two quick searches yourself, use web search; it'll be faster than reading a whole report. If you'd need to open a dozen tabs, take notes, and synthesise, that's exactly the job Deep Research does well, and the wait is worth it. And if the question is trivial or timeless, not tied to anything current, you may not need to look anything up at all.

Cost and limits matter too: Deep Research is heavier and usually capped per plan, so it's not the thing to burn on "what's the VAT threshold". Spend it on the questions that really reward depth.

Verification still matters, for both

Here's the trap. A grounded answer looks trustworthy (it has citations, it reads like a briefing) and that appearance does a lot of quiet persuading. But links are not proof. A citation can point to a page that doesn't actually say what the summary claims; the report can blend two sources into a statement neither made; a real source can itself be wrong or out of date. Deep Research reduces the odds of a bare fabrication, but it does not remove your job.

So verify, whichever mode you used. Open the key citations and confirm the source really says what the answer claims, not just that the link works. Cross-check any load-bearing figure against a second source. And treat a slick, well-cited report with the same critical eye as a quick reply: the polish is a reason to check, not a reason to relax. This is the same hallucination risk you met in Phase 0, wearing a smarter suit.

Try it now

Common mistakes

• Using Deep Research for a quick fact. It's slow and usually capped. If a single search answers it, don't spend a research run, and don't wait ten minutes for something you'd have found in ten seconds.
• Skipping the plan step. Clicking past the proposed plan is how a long run ends up thoroughly researching the wrong question. The thirty seconds you spend correcting scope saves the whole run.
• Leaving the sources wide open when it matters. For anything you'll act on, restrict it to sources you'd trust anyway. Breadth is worth less than reliability when the answer has consequences.
• Trusting citations because they exist (the over-trust trap). A footnoted, confident report is more persuasive and therefore more dangerous when it's subtly wrong: a link can point somewhere that doesn't back the claim, and a real source can be out of date. Open the important citations and confirm they say what's claimed; never let the presence of sources stand in for checking them.

Keeping current

Both features move fast: availability by plan, how many sources they read, the caps on research runs. For the current picture, see OpenAI's help articles on ChatGPT Search and Deep Research, plus the ChatGPT release notes. Accurate as of 13 July 2026.