Best AI research tools for evidence-led work

Editor's note: This shortlist rests on each vendor's published documentation and pricing pages, plus two independent studies, whose findings we cite from the Tow Center's published report and from the BBC's announcement of the EBU/BBC study. We have not tested these tools ourselves, and nothing here ranks them on accuracy. A tested update, based on our own citation test, will follow in Can AI research assistants substantiate their claims?

"Research" covers several jobs. Scanning the web for a market overview is one. Finding every trial on a treatment is another. Tracing who has cited a key paper is a third. Reading forty PDFs you already have is a fourth. One tool rarely does all four well, and a tool that handles one job well can be a poor fit for another.

So this shortlist is organised by job, not by brand. For each tool: what it is for, what its documentation says it searches, how it shows citations, its published price on the date we checked and the limits its makers state. Independent research comes afterwards, along with what it did and did not test.

How we chose: the selection criteria

There are no scores or performance verdicts here. A tool made the list when its official documentation showed all five of these:

1. A clear research job: built for one of the four jobs above.
2. Visible sources: citations or links back to the original material.
3. A stated search scope: the open web, a named scholarly index or your own files.
4. Published pricing: a price or free tier on an official page.
5. Stated limits: published usage caps or known weaknesses.

Inclusion means a tool passed these documentation checks, not that we have confirmed its citations are right.

Job 1: deep web research

These three tools search the open web, read many pages and write a report with links. They suit market scans and policy overviews, where evidence is spread across many websites.

ChatGPT deep research

What it is for: multi-step questions that combine many sources into a long report.

What it searches: by default, OpenAI's help centre says deep research can use "the public web" and files you upload. Depending on plan and workspace settings, it can also read connected apps such as Google Drive or SharePoint. You can limit research to sites you name, or tell it to prioritise them.[1]

How it cites: the docs say all outputs "include citations or source links". The finished report has a "sources used" section and an activity history showing how the research progressed.[1]

Price (checked 16 October 2026): the pricing page lists deep research as "Limited" on the Free and Go plans and included on Plus and Pro.[2] OpenAI's help centre gives Plus as $20 a month.[3] Prices for Go and Pro did not appear on the page when we fetched it.

Documented limits: usage varies by plan; fixed allowances reset every 30 days from first use, and availability depends on country.[1]

Gemini Deep Research

What it is for: the same kind of long, cited report, with the option to draw on your Google Workspace content.

What it searches: Google Search is the default source. You can add Gmail, Drive and Chat, upload files or add notebooks, and you can switch Google Search off to research only the sources you selected.[4][5]

How it cites: while it runs, a "Sites browsed" view lists the websites it is using. When you export to Google Docs, Google says its citations come with it in "a works cited section".[6]

Price (checked 16 October 2026, US prices): Deep Research appears in the free plan. Google AI Plus is $4.99 a month, AI Pro $19.99 and AI Ultra from $99.99. Paid plans carry higher usage limits.[7]

Documented limits: daily caps on research requests, and on how many can run at once. You must be 18 or over. Google says a report usually takes about 5 to 10 minutes.[5]

Perplexity

What it is for: fast cited answers, plus a Research mode for longer reports.

What it searches: Perplexity says it searches the web in real time, "gathering information from authoritative sources like articles, websites, and journals". It can also work with files you upload.[8]

How it cites: "Each answer includes numbered citations linking to the original sources," according to its help centre.[8]

Price (checked 16 October 2026): the free Standard plan includes 3 Pro Searches a day and 1 Research query a month. Pro and Max raise those limits. Education Pro is $10 a month with verification, and Enterprise Pro starts at $40 a seat per month.[9] Pro costs $20 a month or $200 a year, according to Perplexity's help centre; its pricing page shows Pro at $17 a month billed annually.[10][11] Our Perplexity review covers its plans, limits and data terms in more detail.

Documented limits: Perplexity says access to advanced models may be limited during weeks of heavy use.[12]

Job 2: academic literature search

These tools search indexes of research papers, not the open web: fewer press releases and blogs, but only what the index covers.

Elicit

What it is for: finding, screening and extracting data from papers, up to systematic reviews.

What it searches: Elicit says it covers more than 138 million papers drawn from Semantic Scholar, PubMed and OpenAlex, updated weekly, plus a ClinicalTrials.gov option. It excludes books and dissertations. Without full-text access, it works from the title and abstract.[13]

How it cites: in research reports, you click inside the report to see "supporting quotes and reasoning". You can export the papers it screened, its screening decisions and its extraction table to CSV.[14]

Price (checked 16 October 2026): a free Basic plan includes unlimited search with limited use of reports and the research agent. On the default view of the pricing page, paid plans start at $11 per user a month, billed annually. A separate industry view lists Pro at $49 per user a month, billed annually.[15]

Documented limits: Elicit is frank about this. It "can miss the nuance of a paper or misunderstand what a number refers to", and it cannot reliably tell a strong study from a weak one.[16]

"Elicit summarizes the findings of a bad study just like it summarizes the findings of a good study."

— Elicit Help Center, Elicit's limitations (2026).[16]

Consensus

What it is for: answering research questions from the scientific literature, with a synthesis across papers.

What it searches: Consensus says it searches more than 220 million peer-reviewed papers drawn from Semantic Scholar, OpenAlex and its own crawl of the scholarly web, updated weekly. It ranks results partly by recency, citation count and journal reputation.[17]

How it cites: every AI synthesis includes citations, and the docs say citations can show the supporting passage from the paper when that is available. Retracted papers are flagged and left out of AI analyses.[17][18]

Price (checked 16 October 2026): a free plan includes 10 Pro messages and 3 Deep reviews a month. Pro is $20 a month or $144 a year, and the Deep plan is $65 a month or $540 a year.[19]

Documented limits: Consensus says it does not have access to all research, and that its AI can misread a real paper.[18]

Job 3: citation discovery

Here the job is to follow the trail: who cited a key paper and what built on it.

Semantic Scholar

What it is for: a free academic search engine from Ai2, built around the citation graph.

What it searches: papers gathered by web indexing and through publisher partnerships, across all fields of science. Its main focus is published articles and preprints. Book coverage is very limited and patents are not included.[20]

How it handles citations: paper pages list citing work. It flags "Highly Influential Citations" and labels citation intent as background, method or result, both where it has the citing paper's full text. Alerts track new citations of a paper or author, and Research Feeds recommend new papers from folders you curate.[20]

Price: free. An account is optional for searching and needed for alerts and a library.[20]

Documented limits: coverage focuses mainly on English-language publications. The "Ask This Paper" feature has been tested only on English papers and works on a subset of them, and the FAQ warns that generated text "will not be free of errors".[20]

Consensus also has a citation graph, if you want discovery and synthesis in one tool.[17]

Job 4: reading and annotating your own sources

Gemini Notebook (formerly NotebookLM)

What it is for: asking questions across documents you have already chosen.

What it searches: only the sources you add. These include PDFs, Google Docs and Slides, web URLs, public YouTube videos with captions, audio files and pasted text. Its help centre says chat responses "only use data from your sources".[21][22]

How it cites: answers use direct quotes from your sources as citations. You can hover over a citation to see the quoted text, or select it to jump to that passage in context.[22]

Price (checked 16 October 2026): usable without a paid plan under standard limits. Google AI Plus gives two times the standard limits, AI Pro four times, and AI Ultra 5 or 20 times the AI Pro level.[23][7]

Documented limits: free users can add up to 50 sources per notebook, and each source can hold up to 500,000 words or 200MB. Paywalled web pages are not supported, and footnotes and comments in Google files are not imported.[21]

The catch: an answer can only be as complete as the sources you added. Our explainer on retrieval-augmented generation covers the pattern underneath.

The tools at a glance

US list prices from official pages; they can change. The table summarises vendor documentation, not measured performance.

AI Chronicle

What independent research says about AI citations

Vendor docs describe what a tool is meant to do. Two large independent studies looked at what AI search and assistant tools actually did with sources. Neither tested every tool here, or research reports.

Tow Center for Digital Journalism, March 2025. Researchers gave eight AI search tools excerpts from 200 news articles and asked for each article's headline, publisher, date and URL: 1,600 queries. Collectively the tools got more than 60% wrong, from 37% for Perplexity to 94% for Grok 3. Premium tiers such as Perplexity Pro answered more correctly than free versions but were also more often confidently wrong. More than half of Gemini's and Grok 3's responses cited fabricated or broken URLs. Each excerpt was queried once, and results can vary between runs.[24]

"Generative search tools fabricated links and cited syndicated and copied versions of articles."

— Klaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center, AI Search Has a Citation Problem (2025).[24]

EBU and BBC, October 2025. Journalists from 22 public service media organisations in 18 countries assessed more than 3,000 responses to news questions from ChatGPT, Copilot, Gemini and Perplexity. They found that 45% of answers had at least one significant issue. In 31% of responses, the issue was sourcing: missing, misleading or incorrect attribution. A fifth had major accuracy problems. Gemini had significant issues in 76% of its responses, largely because of sourcing.[25]

Both studies covered news questions in general assistant or search modes in 2025. They say nothing direct about Elicit, Consensus, Semantic Scholar, Gemini Notebook or any current deep research mode. They do show that a citation's presence does not prove the source supports the claim. Our AI hallucinations explainer covers why this happens.

How to check any tool yourself

1. Click every citation that carries weight. Confirm the link opens, the page exists and it is the original source, not a copy or a summary of one.
2. Find the sentence. Search the source for the specific claim. A relevant page is not the same as a supporting passage.
3. Check the date and status. Is the source current? For papers, check for retractions or corrections on the publisher's page.
4. Ask what was searched. Was it the open web, a paper index or only your files? A missing study may simply sit outside the tool's scope.
5. Run a known-answer test. Ask something you can already answer from trusted sources, and compare.

Our guide on how to verify AI-generated research sets out the full routine.

A worked example (hypothetical)

This scenario is invented to show how the jobs fit together. No tool was run, and no outputs are reported.

Priya, an analyst at a small housing charity, has to brief her board on whether "four-day weeks improve staff wellbeing without hurting output." She has a week.

Literature first. She would put the question to Elicit or Consensus and read the supporting quotes for each cited study, checking sample sizes and design herself.

Then the citation trail. For the two studies that carry most weight, she would check Semantic Scholar for later work that cites or disputes them.

Then the wider web. For recent pilots reported by governments or employers, she would run a deep web research report and treat it as a reading list, opening each linked source.

Finally, her own pile. She would add the papers and pilot reports she has verified to a notebook tool and draft the briefing from quoted passages, so every sentence traces back to a document she has read.

The test: could a board member click any claim and land on the passage that supports it? If not, the claim is cut or rechecked.

Questions readers commonly ask

Which tool is the best overall?

We do not name one. We have not tested these tools, and the right choice depends on the job: a literature tool will not find a council press release. Our citation test will compare tools on measured results.

Are the citations from academic search tools always real?

Consensus says it retrieves real papers before generating text, so the papers it cites exist, but it acknowledges its AI can still misread them.[18] Existing and supporting the claim are different things, so check the passage.

Can I use these tools with confidential documents?

Check each vendor's data policy and your organisation's rules first. Google says Gemini Notebook data is not used for training unless you give feedback.[23] ChatGPT deep research follows your regular ChatGPT data controls.[1] Our Glean profile explains why permissions matter once AI can read company files.

Should my team build its own research agent instead?

Sometimes, if you need fixed approved sources and an audit trail. Our research agent tutorial shows how.

The next step

The tool finds the sources; you still have to check them. Before you put AI research in front of anyone, work through our guide on how to verify AI-generated research.

New to AI at work? Start with how to start using AI at work.

Related reading: Can AI research assistants substantiate their claims? · AI hallucinations explained: why plausible answers fail

Sources and evidence note

Product descriptions, prices and limits come from each vendor's own help centre, pricing or product pages, and describe intended behaviour, not tested performance. Prices are US list prices as displayed on 16 October 2026. Independent evidence comes from the Tow Center's report as published by the Columbia Journalism Review, and from the BBC Media Centre's announcement of the EBU/BBC study; we did not read the full EBU/BBC report. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• AI hallucinations explained: why plausible answers fail