# Deep research AI: how to choose a tool

Deep research AI is the name for assistants that take one question, search and read the web for several minutes or hours, and hand back a long report with citations. ChatGPT, Gemini, Perplexity and Claude all have a mode like this, and several companies sell standalone tools. They share a method and differ on the things that decide whether you can use the result: how much they read, whether the citations hold up, how long they take, how many you get, and what you can keep. This guide sets out those differences with published measurements, gives you a five-minute test to run on any tool, and compares the main ones from their own documentation.

## What these tools do

A 2025 survey of the field describes the same four stages in every system: [planning, question development, web exploration and report generation](https://arxiv.org/abs/2508.12752). In practice the tool turns your question into a research plan, runs many searches, opens pages, and writes a report that cites what it found. OpenAI's launch post describes its version finding and synthesizing [hundreds of online sources](https://openai.com/index/introducing-deep-research/). The same post says it "can sometimes hallucinate facts in responses or make incorrect inferences", which is why the criteria below matter more than a feature list.

## The criteria that separate them

### Sources actually used, not sources listed

The number of links at the bottom of a report is a weak signal. DeepResearch Bench, a set of 100 PhD-level tasks, counts only citations whose page supports the statement. Per task it measured [111.21 supported citations for Gemini Deep Research, 40.79 for OpenAI Deep Research, 31.26 for Perplexity Deep Research and 8.15 for Grok Deeper Search](https://arxiv.org/html/2506.11763). A separate audit, DeepTRACE, found that listing more sources did not mean better grounding: Perplexity's deep research mode listed [7.7 sources on average and 97.5% of its statements were unsupported by them](https://arxiv.org/html/2509.04499), while YouChat's listed 57.2 and still left 74.6% unsupported.

### Citations that resolve and support the claim

Two failures are worth separating. The first is a link that goes nowhere. A 2026 study of 53,090 citation URLs from DRBench and 168,021 from ExpertQA found that [3% to 13% of citation URLs were hallucinated](https://arxiv.org/abs/2604.03173), with no record in the Wayback Machine. Deep research agents produced more links per query (113.1 for Gemini's, 41.2 for OpenAI's, against 3.0 to 24.3 for search-augmented chat) and [hallucinated 10.7% of them, against 4.8%](https://arxiv.org/html/2604.03173).

The second is a real page that does not say what the report claims. DeepTRACE put [citation accuracy at 40% to 80% across the systems it tested](https://arxiv.org/abs/2509.04499). Benchmarks disagree on the exact figures, because they use different questions and graders, which is the best argument for testing a tool on your own subject.

### Time per report

Vendors publish very different times. ChatGPT's launch post gave [5 to 30 minutes](https://openai.com/index/introducing-deep-research/) (February 2025). Google's help page says Gemini [usually takes about 5 to 10 minutes](https://support.google.com/gemini/answer/15719111). Perplexity's launch post said [most tasks finish in under 3 minutes](https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research) (February 2025). Claude's help pages describe Claude Research as working [over 1 to 3 minutes](https://support.claude.com/en/articles/11095361-when-should-i-use-web-search-extended-thinking-and-research). Shorter is not worse for a quick survey of a topic. It does bound how much any tool can read and check.

### Usage limits

None of the four assistants' help pages gives a number of reports per consumer plan today. ChatGPT shows a counter in the product and [resets fixed monthly allowances every 30 days from first use](https://help.openai.com/en/articles/10500283-deep-research-faq). Gemini caps [daily research requests and concurrent ones](https://support.google.com/gemini/answer/15719111), and its plans are stated as multiples of a standard limit, with Deep Research [consuming more of your usage](https://support.google.com/gemini/answer/16275805). Perplexity Pro and Max [each include a monthly Deep Research limit](https://www.perplexity.ai/help-center/en/articles/10738684-what-is-research-mode). Claude Research is on paid plans only and [uses the same limits as ordinary chats, faster](https://support.claude.com/en/articles/11088861-use-research-on-claude). What the message means when you hit one, and when it resets, is on [deep research limit reached](/alternatives/deep-research-limit-reached); what each plan costs is on [deep research pricing](/alternatives/deep-research-pricing).

### What you can export and keep

ChatGPT reports download as [Markdown, Word or PDF](https://help.openai.com/en/articles/10500283-deep-research-faq), and deleting the chat deletes the report. Perplexity exports to [Markdown, PDF or DOCX](https://www.perplexity.ai/help-center/en/articles/10738684-what-is-research-mode). Gemini offers [Export to Docs and Copy Contents](https://support.google.com/gemini/answer/15719111). Claude's Research help page does not describe an export. In each case you get the report. The pages it read are not kept for your next question.

## A five-minute test with one question of your own

Pick a question in your own field where you already know the answer, or can check it, and where the answer turns on one number. Then:

1. Run it in the tools you are comparing and note the time each takes.
2. Count the distinct sources each report cites, not the total number of citation marks.
3. Pick three citations at random. Open each link. If one fails, look it up in the Wayback Machine: an archived page is a dead link, no record at all suggests an invented one.
4. On each page that loads, search for the number or phrase the report attributes to it. Mark it supported, partly supported or not there.
5. Find the key number in each report and trace it to the page it came from.
6. Export the report and check that the citation links survive the export.

Three citations will not give you an error rate, but one invented URL or one figure that is not on its page tells you how much checking the tool leaves to you. Our guide to [verifying citations in AI output](/guides/verify-citations-in-ai-output) covers step 4 in more depth, and [the data from our own check of ChatGPT and Claude citations](/benchmarks/citation-check) shows what this looks like across many answers. If you want step 4 done for you on text you already have, that is what an [AI citation checker](/ai-citation-checker) does.

## A worked example of a cited claim checked against its source

Hyperresearch [Answers](/answers) runs that check automatically on a single question: it finds sources, quotes them, and checks that each quote is on the page. We asked it about the audit above. The output below is trimmed to its fields and not edited.

> **Question:** What range of citation accuracy did the DeepTRACE audit (arXiv 2509.04499) report for deep research agents?
>
> **Answer:** 40–80%
>
> **Check:** supported
>
> **Evidence:** [arxiv.org/html/2509.04499](https://arxiv.org/html/2509.04499), quote marked verified: "unsupported statement rates remain high for several DR engines (YouChat(DR) 74.6%; PPLX(DR) 97.5%) and citation accuracy is well below perfect (40–80%)"

## The main tools side by side

Each row comes from the vendor's own pages, re-read on 19 November 2026.

| | Time per report, vendor's figure | Usage limit | Export | Per-claim check result shown to you |
|---|---|---|---|---|
| [ChatGPT deep research](/alternatives/chatgpt-deep-research) | 5 to 30 minutes (2025 launch post) | Counter in product, resets every 30 days | Markdown, Word, PDF | None described |
| [Gemini Deep Research](/alternatives/gemini-deep-research) | About 5 to 10 minutes | Daily requests, higher on paid plans | Google Docs, copy | None described |
| [Perplexity Deep Research](/alternatives/perplexity-deep-research) | Under 3 minutes for most tasks (2025 launch post) | Monthly limit on Pro and Max | Markdown, PDF, DOCX | None described |
| [Claude Research](/alternatives/claude-research) | 1 to 3 minutes | Shares chat limits, paid plans | Not described | None described |
| Hyperresearch | About 30 to 40 minutes (Light), 3 to 5 hours (Deep), targets | Pay per run | Report, evidence table, vault as Markdown | Cited claims checked on both tiers; a verdict per citation on Deep |

An assistant's built-in mode is the better choice when you need an overview in minutes, when you already pay for the plan, or when the research has to read your own mail and files: Gemini [can search your Gmail, Drive and Chat if you choose](https://gemini.google/overview/deep-research/), and ChatGPT can use [connected apps such as Google Drive or SharePoint](https://help.openai.com/en/articles/10500283-deep-research-faq). It is also the better choice when you will check the sources yourself anyway. [Perplexity vs ChatGPT for research](/alternatives/perplexity-vs-chatgpt) sets two of these assistants side by side, and [NotebookLM vs ChatGPT](/alternatives/notebooklm-vs-chatgpt) covers a tool that answers from the sources you load.

Hyperresearch is built for the other case, where the report goes to someone who will ask where a figure came from. A run plans the question, reads tens to hundreds of sources, writes the report, then reads each cited claim against the source it cites and corrects what does not hold. A Deep report adds a [verification receipt](/verification-receipt) with a verdict for every citation, and every source stays in your workspace for the next run. It reads public pages only, logged out, and it does not read Crunchbase or PitchBook. On ten tasks from the [DRACO](https://huggingface.co/datasets/perplexity-ai/draco) benchmark, graded by one judge, Deep scored [77.7 against 73.0 for Claude Research](/benchmarks/draco); that is our own measurement on 10 of 100 tasks, and the page lists its limits. A finished report you can read in full is [who is winning AI coding agents](/samples/ai-coding-agents). If you are building research into your own software instead, see the [deep research API](/deep-research-api).

## Starting

Your first Light report is free when you sign up at the [console](https://console.hyperresearch.ai/login), with no card. After that each run has a flat price, listed on [pricing](/pricing); [research subscription costs](/alternatives/research-subscription-costs) sets that against the plans you may already pay for.
