The AI Research Tool That Cites Real Sources: 10 Ranked by Grounding
Affiliate disclosure: some links here are affiliate links and we may earn a commission at no cost to you. The order below follows the stated criterion and nothing else; no vendor paid for a position or saw this page before it went up.
You asked an assistant for sources and got a citation that went nowhere: a title that does not exist, a link to a blog post dressed up as evidence, or a reference you could not find anywhere. The fix is structural, not a better prompt. A tool can only cite a real paper if it looks one up in an academic index before it writes a word. Consensus, Elicit, Scite, Semantic Scholar, ResearchRabbit, SciSpace and Connected Papers all work that way. Microsoft Copilot grounds its answers in ranked web search results and cites web pages, not papers. Gemini Notebook cites only the files you upload to it.
This page ranks ten tools on that single question, using figures read from vendor pages on 17 September 2026 and nothing else. The short version: Consensus leads because retrieval runs before generation and the vendor says fabricated sources are therefore impossible; Elicit is the runner-up on sentence-level citations across 138 million papers, viewable on the free plan; and Semantic Scholar costs nothing at all, because it is the index rather than a product sold on top of one. Where a vendor does not publish something this criterion needs — a dated corpus count, a renewal price, a free-tier limit expressed as a number — the gap is named here and it costs them position.
What this ranking measures
This ranking measures one thing: whether a citation the tool produces leads to a real, retrievable paper. That breaks into four questions a vendor either answers on its own pages or does not — does it search an academic corpus at all, does it publish what that corpus is and when it was counted, does it publish a mechanism that stops a reference from being invented, and can you see the sources behind an answer without paying. Tools grounded in an academic index rank above general assistants, and where a vendor grounds in the open web or in your own uploads instead, this page says so plainly. What it deliberately does not measure: prose quality, speed, which model is underneath, interface polish, or overall value for money. Prices appear so you can see what verification costs, not to set the order — nothing moves up for being cheap.
The short answer
It is the only tool here whose published architecture answers the criterion directly: retrieval runs before any generation, checker models verify relevance before summarising, and the vendor states that fabricated-source hallucinations are structurally impossible, with only misread-source errors possible. The free plan’s 10 Pro messages and 3 Deep reviews a month let you check that before paying $12/mo billed as $144 annually, or $20/mo month-to-month.
Sentence-level citations across 138 million academic papers and 545,000 clinical trials, and ’View sources for answers’ sits on the free Basic plan, so the grounding is inspectable at no cost. It loses the top spot because its free usage cap is published only as the words ’Limited usage’ and its corpus figure carries no date.
Free with no paid tier and no pricing page, because it is the index rather than a product sold on top of one — a citation there is a record in the corpus, so existence is settled immediately. The AI layer is narrow: Ask This Paper is under development and available on a subset of papers only.
All 10 compared
| # | Product | Best for | Price from | What decides its place | |
|---|---|---|---|---|---|
| 1 | Answers where every cited paper must be real | Free $0/mo; Pro $12/mo (shown as $144 annually) or $20/mo billed monthly; Deep $45/mo ($540 annually) or $65/mo billed monthly; Teams $30/seat/mo | Retrieval runs before generation; 10 free Pro messages a month | Visit | |
| 2 | Sentence-level citations and systematic reviews | Basic free; Academic Plus $19 per user/month, billed monthly; Academic Pro $69 per user/month, billed monthly; Industry Pro $49 per user/month, billed as $588 annually, or $75 per user/month, billed monthly | 138 million papers and 545,000 clinical trials, searchable free | Visit | |
| 3 | Seeing how later papers treated a finding | Connect free; Basic $20 / month Billed yearly; Pro $50 / month Billed yearly — no monthly-billed rate is published | 317M+ articles indexed, but only 41M held in full text | Visit | |
| 4 | Checking whether a paper exists, at no cost | Free — no paid tier and no pricing page exists | 238,070,551 papers on the homepage counter, read 2026-09-17 | Visit | |
| 5 | Mapping a field with no generated text involved | Free, ’$0, Forever!’; ResearchRabbit+ $10* per month on the annual plan ($120 USD / year) or $12.5* per month billed monthly | 310+ million articles, and no LLM in the core recommendations | Visit | |
| 6 | Questioning PDFs with section-level citations | Premium $12 /month billed annually or $20 /month billed monthly; Advanced $70 or $90 /month; Max $160 or $200 /month | Corpus sized at 200Mn+, with no share of citations that resolve | Visit | |
| 7 | Seeing a paper’s neighbourhood as a graph | Free (5 graphs per month); Academic USD3/month* billed 36 USD annually, or USD5/month* billed 15 USD quarterly; Business USD10/month* billed 120 USD annually | Each graph analyses ~50,000 papers; free tier gives 5 a month | Visit | |
| 8 | Grounded answers over PDFs you already have | Gemini Notebook Standard free; paid only inside a Google AI plan, from 132k ₫/tháng (Google AI Plus), 489k ₫/tháng (Google AI Pro), from 2.250k ₫/tháng (Google AI Ultra) | Free: 50 sources per notebook, 50 chats a day, no paper index | Visit | |
| 9 | Drafting around sources you supply or connect | Free $0; Pro $17 / Per month with annual subscription discount ($200 billed up front), $20 if billed monthly; Max From $100 / Per month | 60+ scientific databases counted, and no paper count published | Visit | |
| 10 | Web answers inside Office, not literature work | Free version, ’subject to available capacity and limits’; Microsoft 365 Personal $9.99/month or $99.99/year; Premium $19.99/month or $199.99/year | Citations link to ranked web pages; no paper corpus claimed | Visit |
1. Consensus
Why it sits at 1 Retrieval runs before generation; 10 free Pro messages a month
Consensus is a search engine over the literature with an AI layer on top, and its architecture is why it leads here: retrieval runs before any generation, checker models verify relevance before summarising, and the product returns nothing rather than filling a gap. The vendor states that fabricated-source hallucinations are structurally impossible in its product and that only misread-source errors can occur.
You can test that for nothing — the free plan gives 10 Pro messages and up to 3 Deep reviews per month on top of basic paper search. Pro is $12/mo, displayed beside ’$144 annually’ because the billing switch loads on Annual; month-to-month is $20/mo. Students, faculty and US clinicians are offered up to 40% off, with no resulting figures published.
- The strongest published grounding claim here, and it is checkable free.
- Three corpus figures appear across its own site — 250M+, 220M+ and over 200 million — with the only dated one on a post from August 3, 2022, and its Terms of Service contain no billing or renewal section at all.
2. Elicit
Why it sits at 2 138 million papers and 545,000 clinical trials, searchable free
Elicit searches 138 million academic papers plus 545,000 clinical trials and claims every AI-generated claim is backed by sentence-level citations to the underlying sources. It is also the most candid vendor here about its own failure mode: it acknowledges hallucination, says it biases toward silence over an incorrect answer, and warns that its grounding does not extend to judging study quality — weak and strong studies are summarised alike.
Seeing sources is not paywalled. Free Basic includes ’View sources for answers’, unlimited search across the corpus, unlimited summaries and unlimited full-text chat with papers. Paid pricing follows an audience selector: under Academic, Plus is $19 per user/month billed monthly; under Industry, Pro is $49 per user/month billed as $588 annually, or $75 per user/month billed monthly.
- Citations are traced to the sentence, and the free plan can view the sources behind an answer.
- Free-tier usage is published only as the words ’Limited usage’ — no number of runs, queries or uploads anywhere — and Elicit issues no partial refunds for unused subscription time.
3. Scite
Why it sits at 3 317M+ articles indexed, but only 41M held in full text
Scite indexes more articles than anything else here, over 317 million, and its unit of work is the Smart Citation: the sentence in a citing paper where your reference is actually discussed. The vendor says every answer carries a citation traceable to the exact sentence in the source paper, verifiable in one click, asserts that answers are never generated, and positions its grounding against rivals that cannot read paywalled full text.
Two things hold it at third. Only 41 million of those articles are held in full text, while the homepage compresses the claim to ’317M+ full-text’. And the free Connect plan gives 25 MCP credits a month with no Assistant or Search access at all, so citation behaviour cannot be checked before paying $20 / month Billed yearly — the only billing basis the page shows.
- The only vendor here that dates a coverage claim, to June 2026.
- Fees are non-refundable, cancellation requires thirty days’ notice before the next billing month, and the 7-day trial auto-enrols you in the plan you selected at the start.
4. Semantic Scholar
Why it sits at 4 238,070,551 papers on the homepage counter, read 2026-09-17
Semantic Scholar is the index rather than a layer sold on top of one: free, operated by Ai2, a non-profit research institute founded in 2014, with no paid tier and no pricing entry in its navigation. The homepage counter read 238,070,551 papers on 17 September 2026. A citation here is a record in that index, and no account is needed to read papers at all.
The AI layer is where it gives ground. Ask This Paper is ’currently under development’ and ’available on a subset of papers’, answering from the single paper you are reading and returning supporting statements from it. Semantic Reader covers arXiv papers. The vendor warns that its generative features produce errors that are hard to spot and places verification on the reader.
- Free forever with no upsell, and it is the corpus other tools license.
- Its own pages disagree on size — 238,070,551 on the homepage, 214 million on the API page, over 200 million on About — and none of the three carries a date.
5. ResearchRabbit
Why it sits at 5 310+ million articles, and no LLM in the core recommendations
ResearchRabbit removes the failure mode rather than managing it. The vendor’s explicit position is that no large language model drives its recommendations, so there is no generative step that could invent a reference. What comes back is real indexed articles reached through the citation graph — over 310 million, aggregated from three named third-party metadata providers and refreshed weekly.
That is also the ceiling: it returns papers, not cited prose, so it answers no question for you. The free plan carries unlimited searches, unlimited library and collections, Zotero import and BibTeX export, capped only at 50 seed articles per search. RR+ raises that to 300 and adds Signals integrity badges, at $10* per month on the annual plan ($120 USD / year) or $12.5* per month billed monthly.
- No generative step at all, plus unlimited free searching across the whole corpus.
- Its Terms of Service page renders with no terms text, so billing, auto-renewal and refund clauses are not publicly readable, and no refund policy exists anywhere on the site.
6. SciSpace
Why it sits at 6 Corpus sized at 200Mn+, with no share of citations that resolve
SciSpace searches a corpus it sizes at over 200 million records and says its answers carry citations linked to specific sections of the source PDF, which is the right granularity for checking a claim. It positions itself against general-purpose chatbots on exactly this point — hallucinated citations — and claims a benchmark result against Elicit and Consensus for relevant-paper retrieval.
The omissions cost it position. It states that Premium can ’Unlock Eligible Full-Text Papers Through Your Institution’ but never what share of citations resolve to a retrievable paper, and the 200Mn+ figure carries no date. The pricing page shows no free plan card, so the free Chat with PDF limits are invisible beyond ’Max. 100 MB per file’. Premium is $12 /month billed annually, or $20 /month billed monthly.
- The only money-back window on this list: 100% within 24 hours of purchase.
- A 30% promotional code stacks on a ’Save 40%’ yearly discount and the page never states what any plan costs at renewal, nor prints a free-tier credit allowance.
7. Connected Papers
Why it sits at 7 Each graph analyses ~50,000 papers; free tier gives 5 a month
Connected Papers builds a similarity graph, not an answer. It hosts no corpus of its own — it is connected to the Semantic Scholar Paper Corpus under ODC-BY — and each graph is built by analysing roughly 50,000 papers and surfacing a few dozen, selected by co-citation and bibliographic coupling. Every node is a real indexed paper because there is no generated prose anywhere in the product.
It sits here because it answers nothing: you get the neighbourhood around a paper you already have, not a cited claim. Free is 5 graphs per month with all features included. Academic is USD3/month* billed 36 USD annually, or USD5/month* billed 15 USD quarterly; Business is USD10/month* billed 120 USD annually. Subscriptions auto-renew, with a one-year initial term on annual plans.
- Every feature is in the free tier; the only free limit is 5 graphs a month.
- The Terms exclude refunds outright apart from double coverage minus a 10% processing deduction, and let the vendor revise fees at any time without notice.
8. Gemini Notebook (formerly NotebookLM)
Why it sits at 8 Free: 50 sources per notebook, 50 chats a day, no paper index
Gemini Notebook — formerly NotebookLM, whose old domains now redirect to notebook.google — answers only from the sources you upload and declines when the answer is not in them. For citation safety that is near ideal: the reference is a document already on your disk. For finding literature it is the wrong shape, because there is no searchable corpus behind it and no claim anywhere that citations resolve to published papers or DOIs.
The free Gemini Notebook Standard tier needs only a Gmail account and publishes real numbers: 100 notebooks per user, 50 sources per notebook, 50 chats a day, 3 audio and 3 video overviews a day, Deep Research 10 a month, each source capped at 500,000 words or 200MB. Paid Notebook cannot be bought standalone; it unlocks through a Google AI plan, from 132k ₫/tháng.
- Published free-tier numbers you can plan around, and answers that refuse to leave your sources.
- Citations point only to your own uploads, so the tool surfaces nothing you have not already found, and the plans page is geo-localised to Vietnam with no USD figure shown.
9. Claude
Why it sits at 9 60+ scientific databases counted, and no paper count published
Claude is a general assistant with a research posture. Anthropic quantifies research coverage as a count of scientific databases — 60+ — rather than as papers, and answers can carry citations to retrievable papers through the Wiley Scholar Gateway connector, which returns DOI links. Claude Science’s stated defence against fabricated references is an automated background reviewer, not a guarantee of accuracy.
The research workbench is published as a beta product rather than a generally available research tier, and Anthropic does not publish which subscription plans include the Scholar Gateway connector or whether Wiley access costs extra — exactly the fact this criterion needs. Free is $0. Pro is $17 / Per month with annual subscription discount ($200 billed up front), or $20 if billed monthly.
- Citations can resolve to DOIs through a publisher connector, which most general assistants do not offer.
- No paper count, no dated coverage claim, no statement of which plans include the connector, and no guarantee anywhere that generated citations resolve.
10. Microsoft Copilot
Why it sits at 10 Citations link to ranked web pages; no paper corpus claimed
Copilot cites the web. Microsoft describes its retrieval as grounding in ranked web search results, its citations are hyperlinks to web pages rather than to a curated paper corpus, and nothing on its Copilot pricing pages claims that answers cite retrievable peer-reviewed papers or link to DOIs. Its own transparency note says Copilot can fabricate plausible-sounding but false content and that its safety mitigations are probabilistic and can fail.
Last place is about the criterion, not usefulness — inside Word, Excel and Outlook it is doing a different job. A limited free version exists ’subject to available capacity and limits’ with no numeric cap published, and paid limits appear as bands such as ’Higher than free’ and ’Extensive’. Microsoft 365 Personal is $9.99/month, or $99.99/year.
- A free version exists and the consumer entry price is the cheapest subscription here.
- No corpus figure of any kind is published, and the only refund window Microsoft states is a 7-day prorated one on the business page, not on the consumer page.
Which one fits your situation
Semantic Scholar. It is free, has no paid tier, and no account is needed to read papers. If a title is absent from a 238-million-record index, treat it as fabricated. Do not buy anything yet.
Elicit. Academic Pro is $69 per user/month billed monthly and includes a Systematic Review Workflow that screens 5,000 papers, 20 table columns, and reports extracting from up to 135 data sources. Start on free Basic, which already searches all 138 million papers.
Gemini Notebook Standard, free. 100 notebooks, 50 sources per notebook and 50 chats a day covers most reading loads, and paid Notebook cannot be bought standalone anyway — it only unlocks inside a Google AI plan. The free tier is enough.
Scite, and you will have to pay for it. Smart Citations are the only feature here that shows citation context, and the free Connect plan has no Assistant or Search access at all. Basic is $20 / month Billed yearly, with no month-to-month rate published.
ResearchRabbit free for unlimited searches across 310+ million articles at up to 50 seed articles per search, or Connected Papers free for 5 graphs a month with every feature included. Neither writes prose, so neither can invent a reference.
It does not cover literature work. Copilot grounds in ranked web search results and cites web pages. Anthropic counts 60+ scientific databases rather than papers and does not publish which plans include the Wiley Scholar Gateway connector. Pair what you have with a free corpus tier instead of buying a second subscription.
What this market gets people for
- Reading a headline per-month price as what you will actually be charged each month. Consensus’s pricing page loads with the Annual switch already selected, so the only figure shown without interaction is $12/mo — which is $144 taken up front. Month-to-month is $20/mo. SciSpace splits the same way at $12 /month billed annually versus $20 /month billed monthly, and Elicit’s Industry Pro is $49 per user/month billed as $588 annually versus $75 billed monthly.
- Treating a vague free tier as a budget you can plan around. Elicit publishes its free cap only as the words ’Limited usage’, with no run, query or upload number anywhere. SciSpace’s pricing page shows no free plan card at all. Microsoft publishes bands such as ’Higher than free’ and ’Extensive’. Consensus (10 Pro messages a month) and Gemini Notebook (50 sources per notebook, 50 chats a day) are the ones that print real numbers.
- Buying on the corpus headline. Consensus shows three different figures across its own site — 250M+, 220M+ and over 200 million — and the only dated one sits on a post from August 3, 2022. Scite’s homepage compresses its claim to ’317M+ full-text’ while its coverage page says 41 million are held in full text. Semantic Scholar’s homepage, API page and About page disagree. Only Scite dates a coverage claim, to June 2026.
- Committing to a year where no refund exists. Scite states fees are non-refundable and requires cancellation at least thirty days before the following subscription month, while its 7-day trial auto-enrols you in the plan you picked. Elicit refuses partial refunds for unused time. Connected Papers allows no returns and no refunds except double coverage, minus 10% processing fees. SciSpace’s 24-hour money-back window is the only guarantee on this list.
Questions buyers ask
Can any of these actually invent a citation?
Any tool that generates prose can misread a source. ResearchRabbit and Connected Papers have no generative step at all, so there is nothing to invent with. Consensus states that fabricated-source hallucinations are structurally impossible because retrieval runs before generation, and that only misread-source errors can occur. Scite claims answers are never generated. Elicit acknowledges hallucination directly and says it biases toward silence over an incorrect answer. Microsoft states in its own transparency note that Copilot can fabricate plausible-sounding but false content.
Is a free tier ever enough for real literature work?
Often, yes. Semantic Scholar has no paid tier at all. ResearchRabbit frames its Free Forever plan as everything needed to complete a literature review, with unlimited searches across 310+ million articles and up to 50 seed articles per search. Consensus free gives 10 Pro messages and up to 3 Deep reviews per month. Elicit’s free Basic searches all 138 million papers, summarises without limit and includes ’View sources for answers’.
Does NotebookLM search academic papers?
No. It is now called Gemini Notebook, and notebooklm.google redirects to notebook.google. It answers only from the sources you upload and declines when the answer is not in them, and the vendor makes no claim anywhere that citations resolve to published papers or DOIs. The free tier allows 50 sources per notebook, each capped at 500,000 words or 200MB.
Will my price rise at renewal?
No vendor on this list publishes a renewal figure — that absence is consistent across all ten. Scite’s Terms confirm auto-renewal and reserve the right to change fees on at least thirty days’ notice. Connected Papers’ Terms allow it to revise fees at any time without notice. Anthropic’s pricing page says prices and plans are subject to change at its discretion. SciSpace shows a 30% promotional code stacked on a 40% yearly discount and states no renewal rate.
Which one actually gets me full text rather than an abstract?
None of them promises all of it. Scite holds 41 million of its 317 million-plus indexed articles in full text and says it reads paywalled full text where rivals guess. Semantic Scholar says some articles are only available on the publisher’s site behind paywalls and routes through GetFTR and LibKey. SciSpace Premium offers to ’Unlock Eligible Full-Text Papers Through Your Institution’ but never states what share of citations resolve.
Is there an academic or student discount?
Consensus says students and faculty with a valid school email, and US clinicians with an NPI number, can get up to 40% off, but publishes no resulting figures. Elicit shows an Academic view where Plus is $19 per user/month billed monthly, though it publishes no eligibility rule or verification requirement. Scite offers a discount by email request with no percentage published. ResearchRabbit runs country-parity discount codes whose per-country prices are not published.
What none of them publish
These are the questions a buyer would want settled that no vendor on this page answers in public. If one of them decides your purchase, ask the vendor and hold them to the answer.
- A dated corpus figure from almost anyone. Scite dates its coverage to June 2026 and nobody else attaches an as-of date: Consensus, Elicit, SciSpace, ResearchRabbit’s marketing pages and Semantic Scholar all publish counts with no measurement date, and three of them contradict themselves across their own pages.
- What share of citations actually resolve to a retrievable, full-text paper. No vendor publishes this. SciSpace says citations are section-linked and that Premium can unlock institutional full text, but never the proportion; Scite says one-click verification without a rate.
- Any accuracy, hallucination-rate or citation-precision benchmark. Scite claims answers are never hallucinated and publishes no measured error rate; Consensus guarantees real sources without a figure; Exa declines to guarantee accuracy at all. SciSpace claims a retrieval benchmark against rivals but no accuracy rate of its own.
- Renewal prices. Not one of the ten publishes a second-term figure, and several reserve the right to change fees — Connected Papers at any time without notice, Scite on thirty days’ notice, Anthropic at its discretion.
- Free-tier limits expressed as numbers. Elicit publishes only ’Limited usage’; SciSpace shows no free plan card; Anthropic states there is no fixed message count, only a rolling five-hour window; Microsoft publishes usage bands.
- How many sources an answer draws on. No vendor on this list publishes a per-answer source count for any tier.
- Institutional and enterprise pricing. Elicit is ’Custom’, Consensus ’CUSTOM’, Scite directs buyers to sales@scite.ai, Connected Papers is ’Request a Quote’, ResearchRabbit is ’Contact Us’ and SciSpace is application-only.
- Whether Claude’s Wiley Scholar Gateway connector is included in a given plan or costs extra, and a USD price for Gemini Notebook — its plans page geo-localises to Vietnam and displays VND only.
- What discounted prices actually are. Consensus publishes ’up to 40% off’ with no figures, ResearchRabbit publishes country-parity codes without per-country prices, and neither says whether the discount survives renewal.
Reading this from the other direction? NotebookLM Alternatives, Ranked by Source Grounding covers the same market against a different rule.