Google, schmoogle: When to ditch web search for deep research

Mint

2 days ago

Searching for the perfect electric car could have taken hours. Instead, I opened ChatGPT, clicked the deep research button and walked away from my computer.
By the time I'd made coffee, ChatGPT delivered an impressive 6,800-word report.
This year, ChatGPT and other popular AI chatbots introduced advanced research modes. When activated, the AI goes beyond basic chat, taking more time, examining more sources and composing a more thorough response. In short: It's just more.
Now free users can access this feature, with limits. Recent upgrades, such as OpenAI's latest GPT-5 model, have made research even more powerful.
For the past few months, I've experimented with deep research for complicated questions involving big purchases and international trip planning. Could a robot-generated report help me make tough decisions? Or would I end up with 6,000-plus words of AI nonsense?
The bots answered questions I didn't think to ask. Though they occasionally led me astray, I realized my days of long Google quests were likely over. This is what I learned about what to deep research, which bots work best and how to avoid common pitfalls.
Deep research is best for queries with multiple factors to weigh. (If you're just getting started, hop to my AI beginner's guide first, then come back.)
For my EV journey, I first sought advice from my colleagues Joanna and Dan. But I needed to dig deeper for my specific criteria including a roomy back row for a car seat, a length shorter than my current SUV and a battery range that covers a round trip to visit my parents.
I fed my many needs into several chatbots. When I hit enter, the AI showed me their 'thinking." First, they made a plan. Then, they launched searches. Lots of searches. In deep research mode, AI repeats this cycle—search then synthesize—multiple times until satisfied. Occasionally, though, the bot can get stuck in its own rabbit hole and you need to start over.
Results varied. Perplexity delivered the quickest results, but hallucinated an all-wheel drive model that doesn't exist. Copilot and Gemini provided helpful tables. ChatGPT took more time because it asked clarifying questions first—a clever way to narrow the scope and personalize the report. Claude analyzed the most sources: 386.
Deep research can take 30 minutes to complete. Turn on notifications so the app can let you know when your research is ready.
My go-to bot is typically Claude for its strong privacy defaults. But for research, comparing results across multiple services proved most useful. Models that appeared on every list became our top contenders. Now I'm about to test drive a Kia Niro, and potentially spend tens of thousands based on a robot's recommendation. Basic chat missed the mark, proposing two models that are too big for parallel parking on city streets.
Other successful deep research queries included a family-friendly San Francisco trip itinerary, a comparison of popular 529 savings plans, a detailed summary of scientific consensus on intermittent fasting and a guide to improving my distance swimming. On ChatGPT and Claude, you can add your Google Calendar and other accounts as sources, and ask the AI to, for example, plan activities around your schedule.
Deep research isn't always a final answer, but it can help you get there.
Ready for AI to do your research? Switch on the 'deep research" or 'research" toggle next to the AI chat box.
ChatGPT offers five deep research queries a month to free users, while Perplexity's free plan includes five daily. Copilot, Gemini and Grok limit free access, but don't share specifics. Paid plans increase limits and offer access to more advanced models. Claude's research mode requires a subscription.
Here are tips for the best results:
Be specific. Give the AI context (your situation and your goal), requirements (must-haves) and your desired output (a report, bullets or a timeline). Chatbots can't read your mind…yet.
Enable notifications. Deep research takes time. Turn on notifications so the app can ping you when your response is ready.
Verify citations. AI can still make mistakes, so don't copy its work. Before making big decisions, click on citations to check source credibility and attribution.
Summarize the output. Reports can be long. Ask for a scannable summary or table, then dive into the full text for details.
Understand limitations. The information is only as good as its sources. These chatbots largely use publicly available web content. They can't access paywalled stuff, so think of it as a launchpad for further investigation.
Whatever the imperfections of deep research, it easily beats hours and days stuck in a Google-search black hole. I have a new research partner, and it never needs a coffee break.
News Corp, owner of Dow Jones Newswires and The Wall Street Journal, has a content-licensing partnership with OpenAI. Last year, the Journal's parent company, Dow Jones, sued Perplexity for copyright infringement.