Deep Research: Which AI Researches Best

Every major AI vendor now has a feature for in-depth research: deep research in ChatGPT and Gemini, Research in Claude and Perplexity, and Researcher in Microsoft Copilot. They all plan the research, search in several rounds, read a lot of sources and hand you a report with citations. Which AI researches well for you depends less on the brand than on three questions: where do the sources you need live, how precisely can you steer them, and do you check the report before you use it? Without a subscription, ChatGPT and Gemini offer limited deep research according to their vendors. Perplexity's help center mentions limited free access, while its pricing page lists deep research only from Pro up. Claude only offers Research on paid plans. A research report is always a draft, never a verified source.

I work with AI tools every day, and my fourteen AI employees run on Claude Team. One of them, Sebastian, my AI employee for SEO and GEO, handles keyword research for my website, among other things. His rule: every figure comes from a saved tool result in his project folder. I did not run my own head-to-head test with the same question across these tools for this article. Features and plans come from the vendors' own pages, which I checked on October 28, 2026. All the basics about AI are collected on my AI Basics page.

What deep research does differently

A regular web search inside an AI chat grabs a handful of recent results and summarizes them. OpenAI puts the difference plainly: use search for quick facts, and deep research for depth and thoroughness. It takes longer, reads many sources and produces a detailed, documented report.

It's the loop you know from AI agents: break the question down, search, read, decide what to look up next, write it all up. Anthropic describes Research in Claude exactly that way, as multiple searches that build on each other. I explain what makes this way of working tick in Agentic AI Explained: What It Is and Isn't. You get an overview in minutes, but also more places where something can be wrong.

Deep research compared: what the vendors document

The table sums up what the vendors state on their help and pricing pages as of October 28, 2026. It doesn't tell you whose reports are better.

ChatGPT deep research

According to OpenAI, deep research uses the public web and your uploaded files by default, plus connected apps where your plan and workspace allow it. It only reads from those apps. Useful: you can restrict the research to specific websites, or just prioritize them. You can edit the proposed research plan before it starts and interrupt a run that's already going. Reports include citations and download as Markdown, Word or PDF. The pricing page lists deep research as limited on Free and Go, expanded on Plus and at its maximum on Pro.

Claude Research

Anthropic offers Research only on the paid Pro, Max, Team and Enterprise plans, and web search has to be on. Claude searches the web and your connected tools, such as Google Workspace. Anthropic names two catches itself: in the mobile app, Research searches only the web and skips your connected tools without telling you. And Research burns through usage limits faster than regular chats. If your documents sit in folders on your computer anyway, take a look at Claude Cowork: What It Does, Who It's For.

Gemini Deep Research

Google lists Deep Research even for users without a Google AI plan, with the caveat that it may be unavailable to them during periods of high demand. Google Search is the default source. You can add Gmail, Drive and Google Chat if the Google Workspace app is connected, upload your own files, or deselect Google Search to search only your own content. You can edit the research plan; per Google, the report usually takes 5 to 10 minutes and can be exported to Google Docs. You can only find past reports again if your Keep Activity setting is on.

Perplexity Research

Perplexity describes Research as a mode that runs dozens of searches and reads hundreds of sources. Its help center says free users get limited access and Pro subscribers extended access; the pricing page, however, only lists deep research from Pro up. Limits refill on a rolling basis, and you can't pick the model in this mode. Since the "Advanced Deep Research" update, Research asks clarifying questions on broad queries first, takes follow-up questions while it runs and works with uploaded documents.

Microsoft Copilot Researcher

According to Microsoft, Researcher is available to Microsoft 365 Premium and Pro subscribers and to business customers with a Microsoft 365 Copilot license. It searches the web and, at work, the files, emails, meetings and chats you have access to, under the same permissions as everywhere else. Microsoft has retired its earlier Deep Research feature.

Which AI researches best? Five questions to decide

Instead of a ranking, here are the questions that settle it for your situation:

1. Where do your sources live? On the open web, all five will do. If your documents sit in Gmail and Drive, Gemini is the shortest route; with Microsoft 365 and a Copilot license, it's Researcher. If you've already connected your tools to Claude or ChatGPT, use that one.
2. Do you need control over sources? If only government sites, standards or vendor pages should count, you need a way to limit the search to specific websites. ChatGPT documents this explicitly; in Gemini you can deselect Google Search entirely.
3. Can you trace every claim? Everyone promises citations. What matters is whether you can find the passage in the source quickly. Judge that on a topic you know well.
4. Should the research see your data? Once emails, files or chats become sources, your plan's privacy rules apply. OpenAI says deep research follows the same data handling as regular chats. In a company, this belongs in the approval of the tool.
5. Are the limits enough? Google rates deep research as a high-usage feature, and Anthropic says Research uses up limits faster. If you need several reports a day, you'll likely end up on a paid plan.

How to research with AI: a six-step workflow

I recommend this workflow regardless of the tool. Steps 5 and 6 are the same rules I give my AI employees.

1. Narrow the question. Add time frame, region, audience and purpose, for example: "Which public funding programs for digitalization are open to small manufacturers in Germany in 2026? Official federal and state sites only."
2. Set the sources. Say what counts and what doesn't, and restrict the search to those sites if the tool supports it.
3. Read the plan before it runs. ChatGPT and Gemini show it to you. Correct the course here; later it costs a fresh run.
4. Treat the report as a draft. Read the source list first. If sources you'd expect are missing, the report is incomplete, however good it sounds.
5. Check figures, dates and quotes against the source. For every number you plan to reuse, open the cited source and find the passage. If it isn't there, or says something different, it goes. With Sebastian, that's the rule: every figure needs a source.
6. Have it checked separately. Let someone who didn't write the report review anything important, another tool or a colleague. My rule: whoever produced something doesn't check it. Large deliveries get an independent cross-check on top, because two checks from the same source are only one check. On September 18, 2026, exactly such a cross-check found errors that Sebastian's own editing had missed.

Where deep research falls short

Deep research doesn't make hallucinations impossible; it moves them. Instead of a made-up answer without a source, the worst case is a claim with a source that says something else. How to spot those errors is covered in AI Hallucinations: Why AI Makes Things Up.

There are practical limits too. Content behind a login is only visible through connected apps. A repeat run may find different sources. And according to OpenAI, availability depends on your plan and country; Google also says availability and limits can change. If you want to search only your own documents, a RAG system is often the better fit.

Frequently asked questions

What is deep research?

A feature in AI chat tools that researches a question in several steps: plan, search the web and allowed sources, read many pages, write a report with citations. It usually takes a few minutes.

Which AI is best for research?

There's no ranking that holds for everyone. What matters is where your sources live, how precisely you can steer them and whether you check the report.

Is deep research free?

Partly. Per the vendors' pages as of October 28, 2026, ChatGPT and Gemini offer limited deep research for free. Perplexity's own pages disagree: limited free access per help center, Pro and up per pricing page. Claude and Microsoft offer it only on paid plans.

How long does a deep research request take?

Google says Gemini usually needs 5 to 10 minutes, Perplexity talks about a few minutes. In Gemini you can leave the chat and get notified when it's done.

Where to go next

Run the test I didn't run, with your own question: a topic whose answer you know, two tools, identical wording, three figures checked against the source in each report. You'll learn more about your situation than from any ranking. To condense long reports afterwards, see Summarize Texts and PDFs with AI. How others use and check AI research day to day is something we discuss in my community Claude Practitioners.