Two distinct jobs hide inside “AI research assistant”, and buying the wrong one is how doctoral researchers end up paying a subscription they stop opening by December. Discovery tools help you find what to read, usually by walking the citation graph. Appraisal and extraction tools help you do something structured with papers you have already found. This comparison covers both, judged on what happens at 300 sources rather than at the onboarding demo. Table first.
| Tool | Job | Index / scale (publisher’s own figure) | 2026 price | Doctoral verdict |
|---|---|---|---|---|
| Semantic Scholar | Baseline search | “over 200 million academic papers” | Free (Allen Institute for AI) | The floor everyone should be standing on |
| ResearchRabbit | Citation-graph discovery | “over 310 million academic papers”; “1,000,000+ researchers” | No pricing page published | Best free discovery tool for exploratory reading |
| Litmaps | Discovery + monitoring | Not published as a paper count | Free tier (20 inputs, 2 maps, 100 articles/map); Pro $10/month, $120/year, academic email required | Buy it for the alerts, not the maps |
| Connected Papers | Single-seed graph visualisation | Not verifiable — see note | Not verifiable — see note | Useful for one job, not a subscription |
| Elicit | Extraction + screening | “more than 138 million papers” | Free tier; Plus $11/mo ($132/yr); Pro $39/mo ($468/yr); Scale $89/mo ($1,068/yr) | The one paid tool most systematic reviewers should consider |
| Scite | Citation-context appraisal | “1.6B+ citations”, “300M+ scholarly sources” | No free tier — 7-day trial; Basic $20/mo, Pro $50/mo, billed yearly | Field-dependent; buy only if contested claims are your problem |
Prices were read from each vendor’s own pricing page in August 2026 and are shown at the annual rate where one exists. Two verification notes, because a comparison that hides them is not much use. Connected Papers’ site returns “we’re sorry but Connected Papers doesn’t work properly without JavaScript enabled” to a plain fetch, and its pricing page carries no readable content — so nothing about its current tiers is quoted here rather than guessed. Consensus blocked the same check outright. Treat every price you read anywhere, including here, as needing a look at the vendor’s page before you enter a card.
The shortlist, ranked for a doctorate
1. Semantic Scholar — the free baseline
Run by the Allen Institute for AI, Semantic Scholar describes itself as providing “free, AI-driven search and discovery tools, and open resources for the global research community”, indexing over 200 million papers from publisher partnerships, data providers and web crawls. It has no paywall, no tier and no upsell.
Who it suits: everyone, as a starting point. Where it falls short: it is a search engine rather than a workflow — no screening, no extraction tables, no project structure. Use it as the layer beneath whatever else you adopt, and note that its open API is why several tools further down this list exist at all.
2. ResearchRabbit — the best free discovery tool
ResearchRabbit works the way exploratory reading actually works: start from a paper you trust, expand outward through related works, authors and citations, and let collections build as you go. Its own figures are “over 310 million academic papers” and “1,000,000+ researchers worldwide”. Notably, its site publishes no pricing page at all.
Who it suits: anyone in the first eighteen months, or anyone entering an unfamiliar sub-literature. Where it falls short at doctoral scale: discovery without discipline sprawls. A graph will happily hand you four hundred papers, and nothing in the tool tells you which forty matter. Pair it with a hard inclusion rule you wrote down first.
3. Litmaps — discovery plus the thing nobody budgets for
Litmaps builds seeded maps of a literature and — the genuinely valuable part over a three-to-four-year candidature — monitors them, pushing alerts as new work appears. The free tier allows up to 20 inputs, 2 maps and 100 articles per map with monthly alerts; Pro is $10 a month or $120 a year with unlimited maps and daily or configurable alerts, and the education rate requires an academic email address.
Who it suits: candidates past the upgrade, whose literature is defined and now needs to stay current until submission. Where it falls short: the free tier’s 100-article cap is small for a doctoral map, so this is effectively a paid tool — but at $120 a year it is the cheapest insurance against the examiner question “were you aware of the 2027 paper on this?”.
4. Connected Papers — one job, done well
Give it a seed paper and it produces a similarity graph of the surrounding literature — a fast orientation to an unfamiliar area, and genuinely useful the week you take on a new chapter. It is not a workflow, does not manage a project, and its current commercial terms could not be verified for this comparison. Use it as an occasional instrument rather than something you subscribe to and forget.
5. Elicit — the extraction engine
Elicit is the tool that does something other than find papers. Its free Basic tier already offers unlimited search across “more than 138 million papers”, unlimited summaries, chat with full-text papers and Zotero import — which is more than most candidates will exhaust. Paid tiers buy structured work: Plus at $11 a month ($132 annually) adds exports to RIS, CSV, BIB, PDF and DOCX plus five extraction columns at a time; Pro at $39 a month ($468 annually) adds a dedicated systematic review workflow that can screen 5,000 papers, twenty columns, and extraction across up to 135 data sources; Scale runs to $89 a month and Enterprise adds screening at 40,000 papers.
Who it suits: anyone building a structured evidence table — a systematic or scoping review, or a methods-comparison chapter where you need the same eight fields from ninety papers. Where it falls short at doctoral scale: extraction accuracy is not verification. Every extracted cell you intend to cite must be checked against the paper, and the checking is not optional overhead — it is the review. Note also that the vendor reserves language about “PRISMA-grade” accuracy for its enterprise tier, which tells you something about how to treat the cheaper ones.
6. Scite — citation context, if that is your problem
Scite’s distinctive claim is that it classifies how a paper has been cited — supporting, contrasting, or merely mentioning — across “1.6B+ citations” drawn from “300M+ scholarly sources”, which lets you see whether a finding you are about to build on has actually been corroborated or quietly contradicted. Basic is $20 a month billed yearly with unlimited assistant use and collections to 1,000 papers; Pro is $50 a month with API access, larger collections and patent, clinical-trial and grant datasets. There is no free tier — only a 7-day trial — and student discounts are handled by referring your institution rather than by presenting a student card.
Who it suits: fields with replication problems, contested effects or a retraction history worth checking. Where it falls short: classification is automated and imperfect, coverage skews to well-indexed STEM literature, and $240 a year is a real fraction of a stipend. If your field’s problem is finding literature rather than adjudicating it, this is the wrong purchase.
What none of them do
Four boundaries worth naming before you attribute powers to any of these tools.
They do not read for you. A summary is a lossy compression of an argument, and the parts it loses — the caveat in the methods, the sample that is not what the abstract implies — are precisely the parts a doctorate is examined on. Anything you cite, you have opened.
They do not confer method. A tool that screens 5,000 records does not make your review systematic. A protocol, an inclusion rule fixed in advance, dual screening where your method requires it and transparent reporting of numbers do that; the software only makes the labour survivable.
They are not complete, and their incompleteness is uneven. Every index here is assembled from partnerships, feeds and crawls, and coverage is markedly better in indexed STEM literature than in humanities monographs, non-English scholarship, grey literature and policy documents. If your field lives in books, these tools are a supplement to your library’s catalogue, not a replacement for it.
They are not your library, your coding software, or your writing environment. References belong in a reference manager; qualitative coding belongs in NVivo, ATLAS.ti or Taguette; the conceptual notes these tools generate belong in your notes system; and none of them is where 80,000 words gets drafted.
One more thing to check before you subscribe
Two questions that cost nothing and change the answer. First, what does your university already hold? Institutional subscriptions to discovery and appraisal platforms are common and badly advertised — your subject librarian knows, and ten minutes with them is the highest-return conversation in this whole comparison. Second, what does your ethics approval and data-management plan permit? Uploading a published paper is unproblematic; uploading your own unpublished chapter, participant data or a collaborator’s manuscript to a third-party service is a different decision, governed by commitments you have already signed.
The recommendation
Start with a free stack and make it prove insufficient before you spend anything: Semantic Scholar for search, ResearchRabbit for citation-graph discovery, Elicit’s free tier for summarising and chatting with papers you have found. That covers the majority of doctoral literature work at zero cost.
Then buy at most one paid tool, chosen by the specific job you cannot otherwise do. Structured extraction across dozens of papers for a systematic or scoping review — Elicit Plus or Pro. Keeping a defined literature current across a four-year candidature — Litmaps Pro at $120 a year. Adjudicating contested findings in a field with a replication problem — Scite. Buying two of these is usually a sign that the underlying problem is an unwritten inclusion rule, which no subscription fixes.
Where the bottleneck actually moves to
Solve discovery and the constraint relocates, quickly and predictably, to writing. Candidates who have found and read three hundred sources do not stall for want of a three-hundred-and-first; they stall converting the material into linear chapters — the failure mapped in our notes-to-chapter workflow, and the reason the introduction chapter is so often the last thing written and the worst thing written.
Tesify is built for that stage: chapter structure, section-by-section drafting and an automatic bibliography that formats your references as you cite them. The reading, the judgement and every sentence remain yours — 100% written by you — which is the only arrangement that survives a viva. It is free to start.
Frequently asked questions
What is the best AI research assistant for a PhD in 2026?
There is no single best, because discovery and extraction are different jobs. For most doctoral researchers the strongest starting configuration is free: Semantic Scholar for search, ResearchRabbit for citation-graph discovery, and Elicit’s free tier for working with papers you have already found.
Is Elicit free?
It has a substantial free tier — unlimited search across more than 138 million papers, unlimited summaries, chat with full-text papers and Zotero import. Paid plans start at $11 a month billed annually and exist chiefly to buy structured extraction and systematic-review screening.
Does Scite have a free plan?
No — only a 7-day trial. Individual plans are $20 and $50 a month billed yearly, and academic discounts run through recommending the tool to your institution rather than through a student rate at checkout.
Connected Papers or ResearchRabbit?
ResearchRabbit for ongoing discovery and collection-building across a candidature; Connected Papers for a fast one-off orientation to an unfamiliar area from a single seed paper. They answer different questions, and the free way to find out is to run both on a paper you know well and see which graph tells you something you did not already know.
Can these tools do my systematic review?
They can carry the labour — search, deduplication support, screening at volume, extraction into tables. They cannot supply the protocol, the inclusion criteria fixed in advance, the second screener your method may require, or the transparent reporting of how many records were excluded and why. The method is yours; the tool is a lever.
Will my examiners object to AI-assisted literature searching?
Searching and screening with software is ordinary research practice and has been since databases replaced card catalogues. What examiners test is whether you know the literature you cite — so the operative rule is that anything in your bibliography is something you have opened and read, regardless of how you found it. Follow your institution’s disclosure requirements for any generative component.
Do these tools hallucinate references?
The retrieval-based ones surface real indexed records rather than inventing them, which is the main argument for using them over a general chatbot for literature work. That is not a guarantee: summaries can still misstate what a paper found, and a citation you have not opened is a risk whatever produced it.
Are they any use in the humanities?
Less, and honestly so. Coverage is built on indexed journal literature, so monograph-based fields, non-English scholarship and archival work are poorly served. Your library catalogue, subject bibliographies and a good subject librarian remain the stronger route.
Should I pay for one of these on a stipend?
Only against a named job the free stack cannot do, and only after checking what your institution already licences. The comparison worth making is not the subscription against zero, but the subscription against the hours it genuinely removes from a specific piece of work.
What about general chatbots for literature review?
Useful for explaining an unfamiliar method or critiquing your own writing; unsuitable for finding sources, because a general model without retrieval will produce plausible references that do not exist. Fabricated citations are found in seconds at doctoral level and are catastrophic when they are.