Ask ChatGPT about something that happened last week and you used to get a polite apology about a training cutoff. That answer has mostly disappeared. Search now runs inside the chat window, pulls live pages, and hands back a written response with numbered sources underneath. Around 800 million people open ChatGPT each week, and a decent share of them are asking questions they once typed into Google.
Searching this way is not the same skill as scanning ten blue links, though. It rewards different habits, and it punishes a certain kind of trust.
What ChatGPT search actually is
There is no separate product to install. The model decides when a question needs fresh information, fires off a query, reads what comes back, and writes an answer that folds those pages together. You can also force it: the web search button in the input bar, a small globe icon on desktop and mobile, tells it to look things up even when it believes it already knows the answer.
The output is a synthesis, not a directory. You get a paragraph that answers the question, inline markers like [1] and [2] pointing at sources, and a list of links at the end. For something like “which London boroughs have introduced overnight parking charges this year”, that is a genuine time saver. For “my local council’s website”, it is slower and more roundabout than typing four words into a search box.
How to get a useful answer out of it
- Force the search when it matters. Anything about prices, releases, deadlines or “right now” should go through the web button rather than being left to the model’s judgement about whether it needs to check.
- Ask in a full sentence. “What changed in the EU AI Act in the last six months” beats “AI Act news”. The extra words shape the query it sends out.
- Name the format. A table of three products with prices, five bullet options, a 200-word brief. It follows the instruction more reliably than you would expect.
- Ask it to show its working. “Quote the exact sentence from each source” is the single best way to catch a summary that has drifted away from what the page said.
- Stay in the same thread. Every reply can trigger a fresh lookup, so you can narrow a broad question over three or four turns instead of cramming everything into one enormous prompt.
The citation habit worth building
Citations are the best thing about searching this way and the easiest thing to over-trust. The model occasionally attaches a source to a claim that only loosely supports it, or blends two articles into a sentence that neither of them actually says.
A two-second habit fixes most of it: scan the numbered list before you act on anything. If a specific figure (a salary band, a tax threshold, a release date) has no citation beside it, treat it as a paraphrase and check it yourself. The sources are clickable, and the ones worth relying on are primary, meaning government pages, company filings, official documentation, published papers. Aggregator posts citing each other are how errors travel.
One more tell: a confident, tidy answer can hide a gap. Ask “what are you least sure about here?” and you will often get a straight list of caveats that was missing from the first response.
Where it genuinely beats a results page
Comparison shopping is the standout. “I need a mechanical keyboard under £120 with hot-swap switches, quiet enough for an open office, shipping to the UK” is a query where a page of links makes you do all the reading. ChatGPT will assemble a shortlist, flag the trade-offs, and mention which options are out of stock. Research that spans conflicting sources works similarly well: a technical bug with five contradictory forum threads, or a holiday with a dozen constraints pulling against each other.
If the answer matters, it is worth running the same question through more than one tool. Our comparison of Perplexity, ChatGPT, Google and Claude as search tools covers where each one tends to win and where they all stumble.
Where it still loses
Three categories stay stubbornly better on a conventional engine.
Breaking news. Anything minutes old can lag behind the index. For a developing story you want a live feed, not a summary written from yesterday’s coverage.
Local and navigational queries. Opening hours, a phone number, “login”, a specific shop’s address. You want the destination, not a paragraph about it.
Anything where seeing all the options matters. A search engine shows forty results and lets you spot the strange one on page three. A synthesised answer quietly drops the outliers, which is exactly the wrong behaviour when you are hunting for the odd exception.
There are engines built deliberately to avoid the advertising layer, and they behave differently again. Andi’s ad-free approach next to Perplexity, Google and ChatGPT is worth a read if that trade-off interests you.
When one lookup turns into a twenty-minute project
Some questions cannot be answered by a single search. For those there is Deep Research, which plans a sequence of lookups, reads dozens of pages, and writes a long report with a proper source list. It often takes five to thirty minutes to finish, and it works best on things like mapping a competitor’s pricing changes across a year or pulling together a literature review from scratch. We looked at what ChatGPT Deep Research actually does with those twenty minutes, including where the time goes and what it produces.
Searching your own documents instead of the web
Half the questions people aim at the web are really about their own files. Upload a supplier contract, a spreadsheet of invoices, or a 90-page PDF and the same interface will search inside it, quote the relevant clause, and compare two versions. It is slower to set up and much faster to live with. The trick is telling it what kind of document it is and what you are looking for, rather than dropping a file in and asking “summarise this”. Our guide to getting useful answers from PDFs, spreadsheets and more covers the prompts that work.
Anchor the search in time
Models drift toward whatever was true when their training data ended, and search results drift toward whatever ranks well, which is often older and more confident than the truth. Two words solve most of this. Add the month and year to the prompt, or say “as of this week”. Ask “what is the current rate” rather than “what is the rate”. It sounds fussy. It changes the answer more often than any other tweak.
A test worth running this week
Pick a question you genuinely care about, something with numbers in it and a bit of ambiguity. Run it through ChatGPT search with the web button on, then run the identical words through Google. Read both answers side by side and note two things: how long each took to get you to something actionable, and whether the ChatGPT answer named a source for every specific figure it gave you.
That last check is the whole game. When every number traces back to a page you can open, you have a research assistant that saved you twenty minutes. When a figure appears with no citation and a confident tone, you have a draft you still need to check. Knowing which one you are holding, before you send it to a colleague, is the difference between a fast workflow and an embarrassing one.