Research AI Agents: Web Search, Literature and Source-Grounded Work

A research agent is worth paying for when it shortens one evidence job you can check: finding current web sources, building a literature table, reading how later papers treat a claim, or answering only from a folder you already trust. A chat box that writes a confident paragraph is a different product, and it will not survive the first time a stakeholder asks, "Where did that come from?"

Use this page to pick that job first. Prices below were read from vendor pricing and help pages on 26 October 2026. They move. Open the linked page before you buy a seat.

TL;DR: Match the tool to the artifact you need to defend. Perplexity for current web briefings with citations; Elicit for extraction tables across many papers; Consensus for fast paper-grounded answers with deep reviews; Scite for checking how later papers cite a specific study; NotebookLM for answers that must stay inside sources you uploaded. No single tool wins all four jobs, and each fails in a different, visible way.

Start from the artifact, not the brand

The useful output is rarely "an answer." It is one of these:

Those jobs overlap at the edges. They fail in different ways. Perplexity can cite a blog that paraphrased a paper. Elicit can extract a cell from an abstract when the full text says something narrower. Consensus can sound settled when the field is split. Scite tells you how a paper was cited, which is not the same as whether the finding is true. NotebookLM will not go hunting for the study you forgot to upload.

A worked question, so the split is obvious

This is an illustration, not a test we ran. A support lead wants to know whether publishing a help center cuts ticket volume. The sentence they almost put in a board slide is: "Companies that launch a help center see fewer tickets."

Ask each tool a different version of that question.

• Perplexity, if the need is current practice: "What have help-desk vendors and independent write-ups said in the last year about help-center deflection, and which pages are they citing?" You are buying a sourced briefing, then you open the pages.
• Elicit or Consensus, if the need is the research record: "What do published studies report about knowledge-base or FAQ interventions and contact volume?" You are buying a paper set and, on Elicit, a column structure.
• Scite, after you have one study in hand: "Do later papers support this result, or do they only mention it?"
• NotebookLM, if legal or product has already approved five PDFs: "Using only these files, what do they actually claim about ticket volume, and where do they hedge?"

If one subscription is asked to do all four, you will get a fluent memo and a weak audit trail. Buy the tool whose output you are willing to read line by line.

Perplexity, when the evidence is still on the web

Perplexity's own pricing page describes inline citations on every answer, including the free plan, and positions Pro and Max as higher limits plus the ability to pick a model. On the page reviewed for this guide, Pro is listed at $17 a month and Max at $167 a month, both when billed annually. See Perplexity's pricing page.

For a team seat, Perplexity's enterprise help center (updated 8 October 2026) lists Enterprise Pro at $40 per seat per month or $400 per seat per year, and Enterprise Max at $325 per seat per month or $3,250 per seat per year. That article also says API use is not included, and that Enterprise Max raises file and video-generation limits, including up to 10,000 personal files and up to 15 video generations a month. Read it here: Enterprise pricing and billing.

What to test in a pilot: take a claim your company might repeat, run it, and open every citation that carries a number. A good result shows the number on the linked page, in the same units. A weak result cites a roundup that cited someone else. Perplexity is a strong fit for market, product, and policy briefings. It is a poor sole system of record for a systematic review.

Perplexity's own channel shows Labs, the multi-step workspace for reports, dashboards, and files, rather than a citation-accuracy study. Watch it after you know which job you are buying.

The Perplexity review goes further into fit, limits, and alternatives.

Elicit, when you need the same fields from many papers

Elicit's pricing page is built around literature work, not general web search. Basic is free and includes what Elicit describes as unlimited search across more than 138 million papers, plus unlimited summaries and chat with papers. Pro is $49 per user per month on month-to-month billing, or $39 per user per month when billed annually as $468. Pro includes a systematic-review workflow that can screen 5,000 papers, reports that extract from up to 135 sources, and 20 table columns at a time. Scale is $229 per user per month month-to-month, or $89 per user per month billed annually as $1,068, with higher extraction limits and live collaboration. Enterprise is custom, with a stated screening scale of 40,000 papers. The page shows different figures for monthly and annual billing, so confirm the checkout rate. Source: Elicit pricing.

The column limit is the buying detail people skip. Twenty columns on Pro is enough for sample, intervention, outcome, and a few custom fields. It gets tight if every reviewer adds "country," "funding," "adverse events," and "what the authors refused to conclude." Scale's 30 columns, and Enterprise's 40, are there for that kind of extraction. If you will not maintain a column dictionary, you do not need Scale.

Elicit's own screening video is explicit that the product should be treated as an assistant and that a person should check the work. Use that as the pilot rule: spot-check extracted cells against the PDF, especially anywhere the abstract is smoother than the results section.

The Elicit review is the longer product note.

Consensus, when you want the paper trail without building the table

Consensus is aimed at questions that should be answered from papers. Its help center article on subscription plans, updated 12 September 2026, lists:

• Free: unlimited paper search, 10 Pro messages a month, 3 Deep reviews a month, 10 Study Snapshots a month, 30 API and MCP calls a month.
• Pro: $20 a month or $144 a year, with unlimited Pro messages, 15 Deep reviews a month, and 500 API and MCP calls a month.
• Deep: $65 a month or $540 a year, with 200 Deep reviews a month and 2,000 API and MCP calls a month.
• Teams: custom, described as Pro features plus 50 Deep reviews per user per month.

Source: Consensus subscription plans. The pricing page is consensus.app/pricing. These figures come from Consensus's own help center; we could not independently verify them from the material used for this review, and help-center rates can drift from the checkout page. Confirm the rate before purchase.

A separate help article describes what a Deep review does: it splits the question, runs up to 20 targeted searches, and synthesizes from a selected set. Consensus says a Deep review typically analyzes up to 50 papers and may go to 100, and that a checkmark on a citation means the full text was used rather than the abstract. The same article says the corpus it searches is 220 million academic papers. That is Consensus's description, not an independent audit. Source: How to use Deep review.

Buy Pro if 15 deep runs a month covers the questions you actually reopen. Buy Deep if literature review is the job, not an occasional lookup. Stay on free if you are still learning whether your questions even belong in the academic literature. Consensus will not price a vendor, read your contract folder, or replace a regulator-facing review protocol. That last job is closer to Elicit's systematic-review workflow, and even there a person owns the protocol.

Scite, when the question is how a paper was cited

Scite is the check you run after you already have a study. Its API docs, which point readers to the live pricing page for checkout, list:

• Free, for public paper data.
• Basic at $20 a month or $144 a year ($12 a month on the annual plan), with 250 MCP credits a month and no self-service REST API key.
• Pro at $50 a month or $480 a year ($40 a month annually), with 2,500 MCP credits and a self-service API key for evaluation.
• Team at the same Pro price per user.

The docs are explicit that a Pro key does not license unrestricted commercial use of every dataset, and that search snippets are redacted on that self-service key. Organization plans are quoted. Source: Scite pricing and plans and the checkout page. The marketing page also offers a 7-day trial that enrolls you in a paid plan unless you cancel. Read that sentence before you start the trial.

Scite's public site describes Smart Citations as a way to see whether later papers support or contrast a claim. Use it to avoid treating "cited 400 times" as agreement. Do not use it as proof the original result replicates. A supporting citation can still be wrong, narrow, or from a paper that barely tested the point.

NotebookLM, when you already chose the sources

NotebookLM answers from sources you add: PDFs, Docs, web URLs, audio, and public YouTube links, among others. Google's help pages now document the product as Gemini Notebook, but the product and its addresses are still notebooklm.google. Either way, it does not replace a literature search.

Google's help pages, reviewed 26 October 2026, describe a free Standard plan of 100 notebooks per user, 50 sources per notebook, and 500,000 words per source (or 200 MB for a local upload). Google AI Plus, Pro, and Ultra raise those caps. Pro, for example, is listed at 500 notebooks and 300 sources per notebook. The upgrade page also says limits can change. Read the current table: NotebookLM plans and source limits.

Two practical constraints matter more than the plan names. Copy-protected PDFs fail to import. A Drive file you lose access to becomes unreadable in the notebook, which is the correct failure if the source was never yours to keep. For a pilot, upload the actual policy or study packet, ask for the sentence a director might quote, and demand the source passage next to it. If the notebook cannot point at a passage, the sentence does not go in the memo.

This is also the right shape for internal research that must not wander onto the open web. If the corpus is your whole company's drive, tickets, and wiki, you are no longer picking a paper tool. That buying job is closer to AI knowledge management and to enterprise search products such as the one covered in the Glean review.

What to score in a one-week pilot

Run the same three prompts in every finalist. Write them down before the demo.

1. A paper or page you have already read. The tool should recover the finding you remember, including the caveat.
2. A claim where credible sources disagree. The tool should show the split, not average it into a slogan.
3. A question that can only be answered from your uploaded set. Anything pulled from outside that set is a failure for that prompt, even if the outside source is good.

Then score the citation, not the prose. Open the source. Check that the number, the population, and the date match the sentence. For extraction tools, check five random cells against the PDF. For citation-context tools, open two "supporting" citations and see whether they actually tested the claim.

Keep a person on the hook for anything that will be repeated outside the team: a board metric, a clinical or legal summary, a security claim, a price. The tools above can assemble the reading. They do not become the author of record.

How the public prices compare

Elicit's table shows different figures for monthly and annual billing. Consensus's figures are from its help center and could not be independently verified in this review. Confirm the exact amount on each purchase page.

If two tools are close, buy the cheaper one whose failure you can see. A missing citation is a better failure than a polished paragraph you cannot trace.

Where this leaves the shortlist

Choose Perplexity for current, cited web research and pay for Enterprise only when an admin needs seats, files, and a contractual plan. Choose Elicit when the deliverable is a screening table or a review protocol. Choose Consensus when you want paper-grounded answers a few times a week and you will live inside a Deep-review cap. Choose Scite when a specific paper needs citation context before you trust it. Choose NotebookLM when the sources are already in hand and the answer must stay there.

Discovery can stay free. An academic index such as Semantic Scholar is enough to find candidate papers before you pay for extraction or citation context.

If the next question is how those sources get retrieved inside a product you are building, read the RAG glossary before you add another research subscription.

Run the pilot before you subscribe

The three-prompt checklist above is the fastest way to see which tool's failure mode you can live with. When you get to the contract, including API access, seats, data retention, and cancellation terms, work through the AI agent buying checklist before you enter a card number.

Frequently asked questions

Which research AI agent is best for a systematic review? Elicit's Pro plan is built for this: it screens up to 5,000 papers in a dedicated workflow and extracts into 20 columns at a time. Consensus Deep reviews can synthesize up to 100 papers, but they do not replace the screening protocol a reviewer must own. For a regulator-facing review, you still write the protocol and check every extracted cell.

Can Perplexity replace a literature search? No. Perplexity is strong for current web briefings with inline citations, but it is not designed to build a structured paper set or a screening table. If your deliverable is a table of studies with consistent columns, Elicit or a manual PubMed screen is the better fit.

Does Scite tell me whether a finding is true? No. Scite shows how later papers cite a study: supporting, contrasting, or mentioning. A supporting citation can still come from a paper that barely tested the claim. Use Scite to check the reception of a paper, not to verify replication.

Is NotebookLM a search engine for academic papers? No. NotebookLM answers only from sources you add. It will not discover literature for you. Use it when your sources are already approved and the answer must stay inside them, such as a legal packet or an internal policy set.

Are the free plans enough to evaluate these tools? Mostly. Perplexity Free has citations on every answer. Elicit Basic has unlimited search and summaries. Consensus Free includes 10 Pro messages and 3 Deep reviews per month. Scite Free covers public paper data. NotebookLM Standard is free with 50 sources per notebook. Run the three-prompt pilot on free tiers before paying.

What is the biggest mistake buyers make? Buying one subscription for all four jobs. Each tool fails differently, and a single fluent tool will hide the failure until someone asks for the source. Match the tool to the artifact you must defend, and keep a person on the hook for anything that leaves the team.