Buying guide

AI Agents for Research: Web, Documents, or Papers?

AgentsBench / Published Nov 21, 2026 / Updated Nov 21, 2026

The best AI agent for research depends on the evidence and the file you need to hand over. Start with ChatGPT deep research for an open-web brief, Gemini Notebook for questions tied to a reading pack, or Elicit for screening papers and exporting structured data. Those are starting points based on documented features and public examples. A useful first trial is to check a single claim all the way back to its source and open the export your colleague will receive.

Three tools, three useful handoffs

Your starting pointTool to tryDeliverable to inspectAccess and cost
An open question and several possible websitesChatGPT deep researchA cited report in Word, Markdown, or PDF from ChatLimited access on Free. Plus is $20/month in the US. Check the in-product task allowance.
An assigned reading packGemini NotebookAnswers linked to the passages in selected sourcesStandard access is free. Google lists up to 50 sources per notebook on Free. AI Pro is $19.99/month in the US. Usage depends on the work you run.
A literature review with eligibility rulesElicitScreening decisions with structured fields and CSV or Excel exportBasic is free with limited agent use. Research Agent table exports require Plus or higher. Systematic Review table exports start at Pro. The individual pricing page listed $132/year for Plus and $468/year for Pro.

Access figures were checked November 21, 2026. Elicit’s quoted prices are annual commitments. Open the account’s current offer before budgeting regular use. ChatGPT plans · ChatGPT Plus price · Google AI plans · Notebook source limits · Elicit pricing

ChatGPT deep research for an open-web briefing

Choose ChatGPT when finding the right sources is part of the work. A commissioning editor who asks what employers mean by a four-day week has given you a subject, not a reading list. You need to discover the different arrangements, locate primary documents, and decide which examples belong together. Deep research lets you review the proposed plan, adjust the question, and follow links from the finished report. It suits a reader who wants a coherent narrative and needs to be able to challenge individual claims. OpenAI’s deep research guide

Use the site selector when a source must carry authority. Restrict to an agency’s domain for a question about that agency’s policy. Prioritize a few domains and allow wider search when different employers’ approaches matter. OpenAI’s release page shows sample deep research answers and citations. Use the example to inspect the report format, then judge the source quality for your own task. OpenAI’s published examples · Website selection options

Try asking for each arrangement’s total hours, pay terms, and original source. Request a separate row for each example and a short passage that supports what the report says about it. If the proposed plan includes mostly opinion pieces, ask for a published study and an employer or government policy before letting it proceed. You can reuse the same table in an article or memo, where each row needs to make sense to somebody who missed the initial conversation.

A finished report in Chat can be downloaded in Markdown, Word, or PDF. Word suits a review that needs track changes. Check the displayed research allowance before promising a batch. A follow-up about page two of an existing memo can often go through a shorter question against that file. Choosing when fresh evidence is worth a new search is one practical difference between agentic and generative AI. Report formats and usage

Gemini Notebook for questions about a known document set

Gemini Notebook, previously called NotebookLM, is the first choice when the reading list is itself a requirement. A reviewer may ask you to use the approved proposal, the signed amendment, and the current price sheet. The answer needs to cite those versions even if an interesting blog post says something else. Add the documents, name each one with a date or version, and select the sources that belong to the question. Google also offers web discovery if you decide to add material later. Google’s product name announcement and guide to adding and selecting sources

For a public example, Google put the source material from I/O 2026 into a notebook and showed an Audio Overview, slides, and a Video Overview made from it. You can see the material the examples draw on and what each format gives you. For a smaller trial, put the four-day-week abstract and OPM definition in one notebook. Ask for total hours, pay terms, and the supporting passage in separate rows. “Not stated in this source” is a useful answer when the OPM page does not establish a pay condition. Google’s published I/O example

Check how the source arrived. Google’s guide says a web URL imports the page’s text and a Google file import leaves out footnotes and comments. A qualification that lives only in a footnote, a chart, or a reviewer comment may therefore be absent from the notebook even when the original contains it. Open the imported source and look for that qualification before accepting a confident answer. Where it matters, supply the relevant passage in a form you can identify, or direct the reader to the original page. Google’s source and import limitations

For handoff, keep the selected filenames, the answer date, and a few decisive passages. Check that the reader can open the originals. When a source changes, revisit the Studio outputs you plan to use. Google says they can still reflect the previous source. The provider also describes a compute-based usage allowance, so extractions of very different lengths may cost different amounts of that allowance. Source updates · Usage limits

Elicit for screening papers against the same criteria

Elicit is built for reviews where each paper has to satisfy the same rule. Define the population, comparison, and result you need to extract. For a review of reduced working time, one possible rule is that scheduled hours fall and pay stays the same. A paper about compressing unchanged hours would fall outside the question. The product’s public walkthrough shows how a reviewer can record inclusion criteria and add custom extraction fields. Elicit’s systematic review walkthrough

Elicit supports title and abstract screening, optional full-text screening after that stage, and extraction into columns you define. Use columns a colleague can verify, such as participants, country, length of follow-up, pay condition, and the passage behind the classification. Treat an unavailable full text as a reason to mark a value uncertain. Check the passage behind borderline decisions and override a recommendation when the paper supports a different call. Retain a short reason so another reviewer can reproduce your choice. Screening, overrides and extraction

The difference shows up in the handoff. A prose report is fine for a meeting, but an analyst who needs to sort by follow-up period needs the underlying table. Elicit documents Word and PDF exports for research reports. At the November 2026 check, exporting tables from a Research Agent session required Plus or higher. Exporting systematic review tables, including screening and extraction, required Pro, Scale or Enterprise. Confirm the tier for the exact table you intend to deliver before a team spends a day filling it. Elicit’s export formats and plan requirements

If your question is about yesterday’s company announcement or a provision in a contract, Elicit’s paper screening workflow is a poor fit. Even in an academic project, a single interesting study is only a starting point. Save the search terms and date, the exclusion reasons for close calls, and a note on any full texts still missing. That is what lets a colleague update the review when another study appears.

Decide what one row represents before exporting the table. An original trial, a later paper about the same participants, and a commentary discussing the trial may all appear in a search. Counting three papers as three independent groups of workers would give the study more weight than it deserves. Keep a column for the underlying study or cohort alongside the publication identifier. Where the relationship is unclear, flag it for the person reviewing the full texts. The same discipline helps with web research when a press release has been picked up by several outlets.

If a colleague has already found a paper that plainly belongs, use it as a check on the search strategy. See whether the search retrieves it, then inspect the phrase or criterion that caused it to disappear if it is missing. Finding one known study never proves a search is exhaustive. Missing it, though, gives you a specific problem to fix before you spend time reviewing a long list.

An acceptance test for the answer you receive

Pick the sentence your reader is most likely to quote. Open its source and check the population, date, and definition. For the four-day-week brief, our reference check records two different statements. The research paper describes a reduction in working time with pay preserved. The OPM page describes fitting a full 80-hour pay period into fewer than ten workdays. These are checks against the original sources. Use them when judging whether a report can keep a reduced-hours study distinct from a compressed schedule. Original study abstract · OPM compressed schedule definition

One final way to decide is to send the smallest useful sample of the deliverable to its recipient. Give an editor a paragraph with two working citations, an internal reviewer one answer with a passage they recognise, or an analyst two completed rows they can sort. Ask whether they could use it in tomorrow’s work. A request for a different file or a more precise source at this stage costs little. Hearing the same request after a full report or a hundred screened papers is considerably more expensive.

Frequently asked questions

Which AI research tool should a student try first?

For a course reading pack, try Gemini Notebook with the assigned material. For a topic requiring a search across public sources, try ChatGPT deep research. Follow the course’s rules on tool use and citations.

Can Gemini Notebook find new sources as well as use my files?

Yes. It offers web and Drive discovery. Select the sources you decide to bring into a notebook and check the imported content before relying on it.

Can I restrict ChatGPT deep research to particular websites?

The site selector allows restriction to specified sites or gives preferred sites priority while allowing a broader search. Choose the option that matches your brief.

Can I take Elicit’s screening table into Excel?

Elicit supports CSV or Excel table exports. Systematic review table export requires an eligible plan, and the Research Agent table has a different plan threshold.

What should I do if a citation points to a source I cannot inspect?

Mark the claim as unverified, narrow it to what the accessible material establishes, or replace the source. Preserve the distinction between an abstract and a full text if it affects your conclusion.