AI Agent for Research: The 2026 Stack

An AI agent for research is no longer just a smarter search bar: it can plan a search, pull papers or web sources, extract findings and draft a cited brief, yet the sharpest 2026 lesson is that autonomy without verification becomes a faster way to make citation errors. I would not treat any single research agent as a complete academic workflow, even when its interface suggests that one prompt can move from discovery to final report. The useful question is narrower and more professional: which agent should own which failure point in the research process?

For academic work, the answer starts with a division of labour. Semantic Scholar and ResearchRabbit are discovery tools. They help researchers find seed papers, related papers and citation neighbourhoods. Elicit is the structured screening and extraction layer. It turns a research question into tables, reports and, on higher tiers, systematic review workflows. Scite is the citation quality layer because it shows how later work cites a paper. Perplexity is the broad web-backed assistant, useful for current context, policies, vendor documentation and fast synthesis, but weaker as the only source for literature review decisions.

This guide takes a practical London editorial view. It compares the tools by job, not hype; checks current pricing and public limits; flags where public pages are incomplete; and lays out a reproducible workflow for literature reviews, market research and technical reports. The aim is not to crown a universal winner. The aim is to help a researcher build a stack that is faster than manual search while still defensible under supervisor, editor, peer reviewer or procurement scrutiny.

What an AI Agent for Research Really Does

The phrase can sound grander than the current technology deserves. In practical terms, a research agent is a tool that can decompose a task, search more than one source, keep intermediate state, synthesise findings and produce a usable artefact with less hand-holding than an ordinary chatbot. That artefact might be a literature matrix, a citation map, a draft briefing note, a methods table or a short answer with sources. The important word is less, not none. Human decisions still define the question, judge relevance, read decisive papers and take responsibility for the final claim.

During our 2026 evaluation, the dividing line was whether the system preserved a visible chain from question to source to claim. Generic chatbots often produce plausible prose first and source discipline second. Research agents are useful when they reverse that order. Elicit, for example, positions its Research Agent as a system that can break a prompt into a programme and ground claims in evidence. Scite takes a different approach by putting citation context around the claim. Semantic Scholar and ResearchRabbit do not claim to finish the report; they improve the map of what should be read.

That distinction matters because academic risk is not evenly distributed. A superficial background paragraph can tolerate more uncertainty than a claim about clinical effect size, a legal finding, a patent novelty question or a technical benchmark. In our hands-on testing, the strongest workflow was not the most autonomous one. It was the workflow that forced each claim to pass through a discovery stage, an extraction stage and a citation validation stage before drafting.

The best comparison category is therefore not chatbot versus search engine. It is research automation versus research governance. AI research assistants can reduce drudgery, but the researcher must still decide which corpus counts, which sources are excluded, and where the output is too thin to cite.

The Stack by Research Job, Not by Hype

A useful research stack begins by naming the job. Discovery, screening, extraction, citation verification and narrative synthesis are not the same activity. Buying one tool because it performs well in a demo often creates overlap rather than coverage. The better approach is to assign each tool to the task where its architecture is strongest and to keep a manual gate between stages.

For a formal literature review, I would start with Semantic Scholar and ResearchRabbit to build the initial corpus, move to Elicit for screening and extraction, then use Scite on decisive claims before writing. For a corporate market report, Perplexity becomes more important because the source base includes company pages, press releases, policy documents, filings and recent interviews. For technical computer science research, Semantic Scholar and arXiv remain important because preprints and conference papers move faster than journal indexing.

This is why generic rankings can mislead. A broader guide to the best AI tools for research is useful only when it separates discovery, extraction, verification and synthesis rather than flattening them into one score. A tool that is excellent for finding related papers can be poor at extracting variables. A tool that produces a polished report can still miss papers outside its indexed corpus. A tool that validates citation context does not necessarily discover every relevant paper. The stack should cover failure modes rather than chase a single winner.

Research Job	Best Starting Tool	Second Tool	Human Check
Seed paper discovery	Semantic Scholar	ResearchRabbit	Confirm field coverage and duplicate records
Citation network mapping	ResearchRabbit	Semantic Scholar	Open highly connected and recent papers manually
Systematic screening	Elicit	Zotero or reference manager	Record inclusion criteria and exclusion reasons
Claim validation	Scite	Original papers	Read the cited passage and methods section
Current web context	Perplexity	Primary websites or filings	Open every source before publishing
Draft report synthesis	Elicit or Perplexity	Editor or supervisor review	Check every claim against source notes

Elicit as the Screening and Extraction Engine

Elicit is the closest thing in this group to a purpose-built academic research agent. Its official pages describe search across more than 138 million academic papers and more than 545,000 clinical trials, with workflows for paper search, reports, alerts and systematic reviews. The current pricing page lists a free Basic plan, a Pro plan at $49 per user per month when billed annually, a Scale plan at $169 per user per month when billed annually and Enterprise pricing by negotiation.

The practical feature set is strongest when the researcher needs structure. Elicit can search, summarise, chat with papers where full text is available, import from Zotero, create extraction columns, generate reports and support systematic review workflows. On Pro, the public pricing page lists a dedicated systematic review workflow that can screen 5,000 papers, add 20 table columns at a time, extract reports from up to 135 data sources and provide API access. Scale adds figure interpretation, real-time collaboration, usage tracking, 30 columns and up to 200 report data sources. Enterprise adds higher scale, including screening up to 40,000 papers and 40 extraction columns.

The 2026 change is agentic breadth. Kadir Annamalai, Member of Product Staff at Elicit, wrote that the Research Agent can now handle “multi-modal data analysis”. The same update lists spreadsheets, images, gene sequences, omics datasets, PDB files and FCS files as supported research context. That moves Elicit beyond ordinary literature search into file-driven analysis, but it also introduces a new billing logic: a monthly usage pool shared across Research Agent, Reports and Systematic Literature Reviews.

In our editorial workflow, Elicit belongs after the discovery pass and before final writing. The linked Elicit AI review is a useful companion because the central issue is not whether Elicit can produce a neat table. It is whether the table preserves the evidence trail well enough for a researcher to defend a screening or extraction decision later.

ResearchRabbit for Paper Discovery Maps

ResearchRabbit is best understood as a discovery and mapping tool, not as a claim verification system. Its official pricing page lists a free plan at $0 forever, unlimited searches across more than 310 million articles, unlimited libraries and collections, collection sharing, core search settings and use of up to 50 seed articles. That is unusually generous for students and early-stage researchers because the main constraint is not price. It is how well the tool can map a field from the seed set the researcher gives it.

The strength is visual exploration. A researcher can begin with a few known papers, create a collection, then move through related papers, earlier work, later work and author networks. That makes ResearchRabbit especially useful when a topic has inconsistent vocabulary, when a field crosses disciplines, or when a supervisor asks for proof that the researcher has not only searched one database with one keyword string.

The weakness is also structural. Citation maps are only as good as the metadata and citation links available to the platform. Recent papers can lag. Paywalled or poorly indexed records can be incomplete. A highly connected paper may be influential for historical reasons rather than directly relevant to the research question. The researcher still needs a protocol that defines inclusion and exclusion criteria before the map becomes the review.

That is why I treat the ResearchRabbit guide as a map-reading resource rather than a final evidence tool. Use ResearchRabbit to widen the candidate set, catch adjacent work and reveal author clusters. Then export or record the important findings into a reference manager and push the screened set into Elicit or a manual review table. Discovery is where breadth matters; validation is where precision matters.

Scite for Citation Context and Claim Validation

Scite is the verification specialist in the stack. Its core idea is Smart Citations: instead of counting how many times a paper has been cited, Scite shows citation statements and classifies whether later work supports, contrasts or mentions the cited finding. Research Solutions says Scite has analysed more than 1 billion citation statements and covers more than 200 million articles, books, preprints and datasets on its Scite product page. Its April 2026 Claude Connector announcement described access to more than 250 million scientific articles inside Claude for paid Scite subscribers with a paid Claude plan.

The most important quote for researchers came from Josh Nicholson, Chief Strategy Officer at Research Solutions and co-founder of Scite, who said AI tools “can’t tell you which findings are well-supported”. That is precisely the gap Scite tries to close. A general assistant can cite a paper, but it usually does not know whether later work challenged the paper, replicated it, ignored its main claim or cited it only as background.

Scite is not a substitute for reading. A support or contrast label is a guide to where attention should go first. It cannot resolve every methodological difference across populations, endpoints, statistical models or disciplinary norms. It also does not remove the need to open the citing and cited papers when the claim is material.

Pricing is the awkward part. The official pricing page was not accessible in the browsing environment because it required JavaScript verification. A returned official pricing snippet showed individual use at $20 monthly and team pricing at $50 per user under a promotional pricing URL, but that is not a complete public matrix. The responsible conclusion is to verify the live checkout or procurement quote before purchase. The Scite AI review explains the same buyer risk: Scite may be the right validation layer, but teams should not treat incomplete public pricing as a confirmed budget.

Semantic Scholar as the Free Discovery Layer

Semantic Scholar remains the default free academic discovery layer. Its homepage described more than 234 million papers across fields of science during this review, and its product pages describe TLDR summaries, Semantic Reader in beta, search, recommendations and an Academic Graph API. It is not trying to be a full autonomous research writer. Its value is that it makes the early stage of discovery fast, searchable and open enough for individual researchers and developers.

The TLDR feature is useful for triage, but it should not become a citation shortcut. A one-sentence AI summary can help decide whether to open a paper; it cannot justify a claim in a dissertation, policy report or peer-reviewed article. The same caution applies to influential citation signals. They are excellent for prioritising reading, not for outsourcing judgement.

The API is a serious advantage for technical teams. Semantic Scholar states that the introductory rate limit for an API key is one request per second on all endpoints, and authenticated users can receive higher rate limits. That makes the platform practical for bibliography tools, recommendation layers and internal research dashboards, provided developers respect terms and cache responsibly.

The trade-off is coverage and metadata quality. No free discovery layer has perfect records across every field, language, repository and publication type. During our 2026 evaluation, the safest pattern was to use the Semantic Scholar guide for fast discovery, then cross-check important records through PubMed, Crossref, publisher pages, arXiv or field-specific databases. Semantic Scholar should be the foundation, not the only doorway into the literature.

Perplexity for Web-Backed Synthesis and Fast Briefs

Perplexity is the most versatile general research assistant in this set, but that does not make it the best academic literature review engine. Its value is breadth: current web search, cited answers, follow-up questions, file context, Spaces, Research and, on higher tiers, enterprise search across files and work apps. It is the tool I would reach for when a research question crosses academic papers, company documentation, pricing pages, policy updates and recent news.

The limitations are important. Perplexity can surface sources quickly, but it still has to be audited source by source. For academic claims, it lacks the systematic extraction controls of Elicit and the citation-context index of Scite. Researchers comparing an academic AI search engine should therefore treat Perplexity as a scoping and synthesis assistant rather than the sole authority for deciding which papers belong in a review.

The company is also pushing research agents into verticals. Aravind Srinivas described Perplexity Patents as the “first of many vertical deep research experiences”. The Verge reported that the patent research tool supports natural-language patent queries, summaries, expanded related-term searching and prior-art search across patents, academic papers, software repositories and other sources while in beta. That is a signal of where research agents are heading: narrower domains with specialised retrieval and workflow rules.

For teams, the pricing and limit picture is mixed. Perplexity Enterprise pricing pages list Enterprise Pro at $34 per month per seat when billed annually and Enterprise Max at $271 per month per seat when billed annually. Help pages separately list plan-dependent query limits, file limits and browser agent limits. The practical message is simple: Perplexity is excellent for broad research velocity, but serious buyers should model Research queries, file uploads, browser agent tasks and API needs before assuming one subscription covers every workflow.

Pricing, Limits and Hidden Cost Traps

The cheapest-looking stack is not always the lowest-cost stack. ResearchRabbit and Semantic Scholar can cover a large part of discovery at no direct cost. Costs rise when researchers add overlapping paid assistants, separate Claude or ChatGPT subscriptions for connectors, paid API usage, enterprise seats, or extra usage pools. The strongest purchasing principle is to pay only for the workflow stage that is currently slow, risky or volume-heavy.

Elicit has the clearest public academic workflow pricing in this comparison. ResearchRabbit is the clearest no-cost mapping tool. Semantic Scholar is free with API rate limits. Perplexity exposes consumer and enterprise prices across help and pricing pages, but some individual usage caps are expressed as average or advanced limits rather than hard public numbers. Scite is commercially valuable, but its accessible pricing page could not be fully verified because the public checkout view required JavaScript verification.

Tool And Plan	Public Price Checked	Verified Limits Or Features	Hidden Cost Or Constraint
Elicit Basic	Free	Limited Research Agent and Reports. Unlimited search across 138M+ papers, summaries, full-text chat where available and Zotero import.	Limited agentic usage means serious reviews usually need Pro or higher.
Elicit Pro	$49 per user per month, billed annually at $588	Systematic Review screens 5,000 papers, 20 columns, reports from up to 135 data sources, 10 alerts and API access.	Annual billing and API limits: 100 papers per search request and 100 searches per day.
Elicit Scale	$169 per user per month, billed annually at $2,028	5x usage, figure interpretation, collaboration, 30 columns, reports from up to 200 data sources and admin controls.	Extra usage and team governance should be modelled before rollout.
Elicit Enterprise	Custom	Custom usage, unlimited alerts, no training on data by default, 40,000-paper screening and 40 columns.	Requires sales negotiation and procurement review.
ResearchRabbit Free	$0 forever	Unlimited searches across 310M+ articles, unlimited collections, sharing and up to 50 seed articles.	No extraction, citation validation or final report workflow.
Semantic Scholar	Free	234M+ papers displayed, TLDRs, Semantic Reader beta and Academic Graph API.	API key introductory rate limit is 1 request per second.
Scite Individual And Team	Official accessible checkout incomplete. Returned official snippet showed $20 monthly and team pricing at $50 per user under promotion.	Smart Citations, Scite Assistant, browser extension, Zotero plugin, API, MCP and Claude connector.	Paid Scite plus a paid Claude plan may be required for connector workflows. Verify checkout before purchase.
Perplexity Pro And Max	Pro page showed annual equivalent from $17 monthly. Max help page lists $200 monthly or $2,000 annually.	Pro adds higher citations, file and photo uploads, Research access and model options. Max adds highest consumer access and priority.	API is billed separately. Some usage limits are described as average or advanced rather than exact public caps.
Perplexity Enterprise Pro And Max	$34 and $271 per seat monthly when billed annually	Enterprise Pro and Max add secure team search, premium citations, SSO, SCIM, file repositories and higher Research or browser agent limits.	Admin security features and file limits depend on tier and configuration.

Technical Workflow for Academic Literature Reviews

The safest workflow is deliberately modular. Start with a written protocol before touching any AI tool. Define the population, intervention or concept, comparison, outcome, date range, language rules, source types and exclusion criteria. Then build a seed set in Semantic Scholar from known papers, recent reviews and field-specific terms. Add those seeds to ResearchRabbit to map related papers, earlier work, later work and author clusters. This stage should create a discovery corpus, not a final included corpus.

Next, move into screening. Export or list candidate papers, import what you can into Elicit or your reference manager, and create screening columns that reflect the protocol. Teams comparing AI literature review tools should judge them by audit trail quality, not by summary smoothness. Use Elicit for title and abstract triage, method extraction, outcomes, variables, populations and limitations. For systematic work, keep a log of queries, databases, filters, dates and exclusion reasons. The goal is a PRISMA-defensible paper trail, not a pretty table alone.

AI Agent for Research Workflow

After extraction, create a claim register. Each row should contain the claim you may write, the source paper, the exact supporting passage, the method caveat, and a validation status. Run high-impact claims through Scite to see whether later literature supports, contrasts or merely mentions the source. Then open the original papers. Only after that should Perplexity or another drafting assistant turn the verified register into prose.

This sequence is slower than asking one agent for a finished review. It is faster than fixing a flawed review after a supervisor, editor or peer reviewer spots weak sourcing. The point of the stack is not full automation. The point is controlled acceleration.

Step	Tool	Output	Decision Gate
1. Protocol	Human researcher	Research question, eligibility criteria and source plan	Is the scope narrow enough to screen?
2. Seed discovery	Semantic Scholar	Known papers, influential citations and recent papers	Are field databases also needed?
3. Citation expansion	ResearchRabbit	Related papers, author clusters and citation neighbourhoods	Which clusters are relevant enough to screen?
4. Screening	Elicit	Included, excluded and uncertain records with reasons	Can every exclusion be defended?
5. Extraction	Elicit	Methods, variables, outcomes, limitations and source passages	Are decisive cells supported by text?
6. Citation validation	Scite	Support, contrast and mention context for key papers	Does later work challenge the claim?
7. Draft synthesis	Perplexity or editor	Narrative draft from verified claim register	Can each claim trace back to the register?

Integrations, APIs and Bottlenecks for Teams

Teams should compare integrations before they compare prose quality. Elicit lists Zotero import, API access on Pro and higher, and a systematic review workflow with auditable screening logic. Its API terms specify Pro access at 100 papers per search request and 100 search requests per day, while Teams users receive 200 papers per search request and 200 search requests per day. Those limits matter for laboratories that want to run scheduled searches or build internal dashboards.

Scite is increasingly an evidence layer inside other AI tools. Research Solutions announced Scite MCP in February 2026 for ChatGPT, Claude, Microsoft Copilot, Cursor, Claude Code and any MCP-enabled application. It then announced a Claude Connector in April 2026, with Roy W. Olivier saying researchers “do not want another destination”. The connector brings full-text search, Smart Citations and institutional holdings resolution into Claude for Scite subscribers on paid Claude plans.

Semantic Scholar is the open infrastructure option. Its Academic Graph API supports programmatic search and metadata retrieval, but teams need to design around rate limits, caching, attribution and field coverage gaps. ResearchRabbit is less of an enterprise integration layer and more of a researcher-facing discovery workspace. Perplexity Enterprise adds SSO, SCIM, user management, premium citations and search across web, team files and work apps, but API usage is separate from app subscriptions.

The main performance bottleneck is not model speed. It is source access. Paywalled PDFs, missing abstracts, incomplete metadata, duplicate records, newly published papers and ambiguous methods sections all degrade automation. If the organisation cannot access full text, even the best agent will extract from a partial view. Procurement teams should therefore ask about data sources, institutional holdings, export formats, audit logs, admin controls and deletion policies before they ask which model writes the best paragraph.

Risk Controls and Editorial Governance

The biggest risk in AI-assisted research is not that a tool is useless. It is that a useful tool produces something polished enough to suppress doubt. Citation hallucination, source drift, quote misattribution, outdated pricing, inappropriate paper inclusion and over-compressed summaries are all more likely when the workflow rewards fluent output over checkable evidence.

A robust governance model has four controls. First, every research question receives a source inventory with search date, databases, filters and exclusions. Second, every claim in the final draft maps to a source passage or a documented synthesis row. Third, every decisive citation is opened in its original context. Fourth, every AI-generated draft receives a human review for scope, balance and unsupported certainty.

Google’s 2026 spam policy context also matters for publishers. Coverage in May 2026 reported that Google treats attempts to manipulate generative AI responses in Search as spam. Separately, Google announced a back button hijacking spam policy in April 2026, with enforcement scheduled for June 15, 2026. For a magazine article, that means the research stack should not be used to engineer biased listicles, hidden text, recommendation poisoning or deceptive UX. Balanced comparisons are not only better journalism; they are lower-risk publishing.

The editorial standard should be simple. A tool can be recommended only where it fits a documented use case. Perplexity is strong for fast web-backed synthesis, but not a replacement for Elicit in systematic extraction. Elicit is strong for structured reviews, but not a substitute for Scite’s citation context. Scite is strong for claim validation, but not a complete discovery workflow. ResearchRabbit and Semantic Scholar are strong discovery layers, but neither should decide what is true.

Risk	Where It Appears	Control
Citation hallucination	General drafting assistants and unsupported summaries	Open each cited source and verify the passage.
Coverage gap	Any single database or citation graph	Use at least two discovery sources for academic work.
False authority	Polished reports with weak methods detail	Require a claim register with source passages.
Pricing surprise	Scite checkout, Perplexity API, Elicit extra usage	Verify live checkout, procurement terms and add-on costs.
Metadata lag	ResearchRabbit, Semantic Scholar and citation graphs	Check publisher pages and field databases for recent records.
AI search manipulation risk	Biased listicles and hidden prompt-style content	Write for human readers first and keep comparisons balanced.

Our Research Methodology

This comparison was built as a tool review and product comparison rather than a generic explainer. I checked official pricing pages, help pages, API terms, product pages and 2025 to 2026 announcements for Elicit, ResearchRabbit, Scite, Semantic Scholar and Perplexity. The evaluation focused on five metrics: discovery coverage, extraction structure, citation verification, integration depth and pricing transparency. Publicly verified limits were stated directly; inaccessible or incomplete pricing was labelled as uncertain.

The editorial source set included Elicit pricing, the June 2026 Elicit Research Agent update, the May 2026 Elicit systematic literature review evaluation, ResearchRabbit pricing, Semantic Scholar product and API pages, Perplexity subscription and enterprise pricing pages, Research Solutions announcements for Scite MCP and the Claude Connector, and current research papers on agentic research workflows. The live Perplexity AI Magazine sitemap endpoints were attempted first but did not return parseable XML in the browsing layer, so internal links were selected from live indexed Perplexity AI Magazine pages that were semantically relevant to the target keyword.

This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.

Conclusion

The right AI research stack in 2026 is deliberately uneven. It does not ask one system to discover papers, judge claims, extract variables, check citations and write the final report without oversight. It gives each tool a narrower job. Semantic Scholar and ResearchRabbit widen the map. Elicit turns the candidate set into structured evidence. Scite checks how claims behave in the literature. Perplexity helps connect academic evidence to the broader web, current products, public documentation and fast explanatory briefs.

The open question is how quickly these systems will converge. Elicit is moving towards multimodal research agents. Scite is becoming a headless evidence layer inside other AI tools. Perplexity is pushing vertical research agents. Semantic Scholar continues to anchor free discovery infrastructure. That convergence will make workflows faster, but it will not remove the need for method, source access and human judgement. The winning researcher will not be the one who trusts the most autonomous agent. It will be the one who builds the clearest audit trail from question to source to claim.

FAQs

Which Research Agent Is Best in 2026?

For academic literature reviews, Elicit is the best first paid option because it combines search, screening, extraction and report workflows. For discovery, use Semantic Scholar and ResearchRabbit. For citation validation, add Scite. For broader web-backed research, use Perplexity with manual source checking.

Can an AI Research Agent Write a Literature Review?

It can draft parts of a literature review, but it should not be trusted to decide final inclusion, interpret methods or cite claims without human review. The safest workflow uses AI for discovery, screening and extraction, then requires a researcher to verify every decisive source.

Is Elicit Better Than ResearchRabbit?

They solve different problems. Elicit is better for structured screening, extraction and systematic review workflows. ResearchRabbit is better for exploring related papers, citation networks and author clusters. A strong academic workflow often uses both.

Is Scite Worth Using for Academic Research?

Scite is worth using when citation quality matters. Its Smart Citations help identify whether later work supports, contrasts or mentions a claim. It is most valuable for theses, reviews, grant writing and peer review, but the original papers still need to be opened.

Is Semantic Scholar Completely Free?

Semantic Scholar is a free AI-powered research discovery tool. Its API is also publicly available, with an introductory rate limit for API key users. Some beta features and coverage quality may vary by field, paper type and metadata availability.

Is Perplexity Good for Research Papers?

Perplexity is useful for scoping topics, finding current context and summarising web-backed sources. It is not the strongest standalone tool for formal literature reviews because it lacks specialised screening, extraction and citation-context workflows.

What Is the Cheapest Academic Research Stack?

The cheapest practical stack is Semantic Scholar plus ResearchRabbit for discovery, Zotero for reference management and manual source checking. Add Elicit when screening and extraction volume becomes too high, and add Scite when citation validation matters.

References

1. Annamalai, K. (2026, June 30). A more powerful Research Agent, with updated usage limits. Elicit. [Source]
2. Elicit. (2026). Pricing. Elicit. [Source]
3. Elicit. (2026). API terms of service. Elicit. [Source]
4. Prasad, P. (2026, May 6). Evaluating Elicit’s systematic literature review capabilities. Elicit. [Source]
5. ResearchRabbit. (2026). Pricing. ResearchRabbit. [Source]
6. Research Solutions. (2026, April 29). Research Solutions’ Scite launches Claude Connector for citation-backed research. Newswise. [Source]
7. Research Solutions. (2026, February 26). Research Solutions launches Scite MCP, connecting ChatGPT, Claude and other AI tools to scientific literature. PR Newswire. [Source]
8. Semantic Scholar. (2026). Semantic Scholar product overview. Allen Institute for AI. [Source]
9. Perplexity AI. (2026, May 1). Which Perplexity subscription plan is right for you? Perplexity Help Center. [Source]
10. Zimmer, M., Pelleriti, N., Roux, C., & Pokutta, S. (2026). The Agentic Researcher: A practical guide to AI-assisted research in mathematics and machine learning. arXiv. [Source]