Scite, Consensus, and Elicit are the three tools most researchers land on after Google Scholar stops being enough, and they are genuinely different products. As of November 2026, Scite indexes 326M+ articles and 1.7B+ Smart Citations and tells you whether later papers supported or contrasted a finding. Consensus searches what it now describes as 400M+ peer-reviewed papers and scholarly works and gives you a directional read on a yes/no question. Elicit searches 138M+ papers and extracts structured data into tables at systematic-review scale. All three return real papers with real DOIs. None of them will take the bibliography already sitting at the bottom of your draft — the one ChatGPT helped you assemble at 1 a.m. — and tell you which entries do not exist. That gap is the thing this comparison is actually about.
"Found a paper" and "have a citation you can defend" are different milestones
The phrase that matters when your supervisor is reading is citable: a reference where the paper exists, the metadata in your bibliography matches the record in CrossRef or PubMed, and the paper actually says what your sentence claims it says. Discovery tools clear the first bar by construction — they retrieve from an index, so what comes back is real. They are much weaker on the second and third.
On the third bar there is a useful evaluation from HKUST Library, which ran the same literature-review prompt through Scite Assistant, Elicit, Consensus, and Scopus AI and hand-checked whether each cited source supported the sentence it was attached to. Their summary line is the one to remember: "Generating accurate citations doesn't ensure their alignment with the arguments in the summary." They also found all four tools sometimes drew conclusions from an abstract's framing rather than its findings, and that Elicit and Consensus both quoted a brief "Note" summarizing another study as though it were the study itself. (HKUST Library, Research Bridge)
On the second bar — metadata accuracy — the numbers come from the general-purpose LLM literature, and they are worth knowing because that is where most bad references in student drafts originate. In Walters and Wilder's 2023 study in Scientific Reports, 55% of the citations produced by GPT-3.5 across 42 topics were fabricated, versus 18% for GPT-4; among the citations that were real, 43% (GPT-3.5) and 24% (GPT-4) contained substantive errors (doi:10.1038/s41598-023-41032-5). In a 2025 experimental study by Linardon and colleagues in JMIR Mental Health, GPT-4o fabricated 19.9% of citations across six simulated mental-health literature reviews, and 45.4% of the remaining apparently real citations had bibliographic errors — most often an incorrect or invalid DOI (doi:10.2196/80371). That study also found fabrication tracked topic obscurity: 6% for major depressive disorder, 28% for binge eating disorder, 29% for body dysmorphic disorder.
Scite: the strongest citation-context tool, and the closest thing to a reference checker
Scite's core product is Smart Citations — for any paper, it shows you the sentences in which later work cited it, classified as supporting, contrasting, or mentioning. When we checked on 2026-11-02, its coverage page listed 1.7B+ Smart Citations extracted from 326M+ indexed articles, 50M+ of which it holds in full text, through indexing agreements with 44 publishers (list last updated September 2026). If your question is "has anyone failed to replicate this," nothing else on this list answers it as directly.
Scite also ships Reference Check, which is the feature most relevant here: you upload a manuscript PDF and it reports how the manuscript uses each reference, whether any reference carries an editorial notice such as a retraction or correction, and how often each has been supported or contrasted. It is a real quality check and it is the only bibliography-level tool among the three.
The limitation is documented in Scite's own help article, which is to their credit. Reference detection uses machine learning and "depends upon the format of the references, whether or not the reference has a DOI, as well as the format of the PDF," and they add that "scite will nearly always miss a few references in our Reference Check" (Research Solutions help center). For retraction screening that is a minor annoyance. For fabrication screening it is the whole problem: a reference that does not exist is indistinguishable from a reference Scite simply failed to parse. Both are silently absent from the report. Absence is not a flag.
Consensus: fastest honest answer to a yes/no research question
Consensus searches a corpus it now puts at 400M+ peer-reviewed papers and scholarly works, assembled from Semantic Scholar, OpenAlex, its own crawl, and publisher partnerships with Wiley, Taylor & Francis, Sage, ACS, APA, AAAS, and De Gruyter Brill. The Consensus Meter aggregates how many retrieved studies point yes, no, or mixed on your question, which is a legitimately good way to avoid building an argument on the one paper that agrees with you.
Two details deserve credit. Retracted papers carry a ⚠️RETRACTED badge and, per the Consensus help center, are never used in its analyses or summaries. And Medical Mode narrows to roughly 8 million papers plus 50,000 clinical guidelines from the top 1,000 medical journals, which is a better default than "everything" if you are writing anything clinical.
What Consensus does not do is take text you wrote and find the source behind it, or take a reference and confirm it. It answers questions. You still have to translate each sentence in your draft into a question, run it, and then copy the metadata across by hand.
Elicit: built for screening thousands of papers, not for checking twelve
Elicit is the systematic-review tool of the three. Its Research Agent, Reports, and Systematic Review workflow screen and extract structured data into custom columns — the Pro plan screens up to 5,000 papers and extracts from up to 135 data sources; Enterprise goes to 40,000 papers and 40 columns. Its index is 138M+ papers, plus 500,000+ clinical trials on paid plans, and it exports to RIS, CSV, BIB, PDF, and DOCX from Plus upward.
If you are running a real systematic review, Elicit is the most capable product in this comparison and the pricing reflects a workflow tool rather than a lookup tool. If you have a 3,000-word essay with fourteen references and a Thursday deadline, you are paying for screening capacity you will never touch.
Scite vs Consensus vs Elicit compared, as of November 2026
Prices and limits below were checked on the vendors' own pricing pages on 2026-11-02. Prices may change; we re-check monthly.
| Scite | Consensus | Elicit | Citely | |
|---|---|---|---|---|
| Core job | Citation context: who supported or contrasted this | Directional evidence on a yes/no question | Structured extraction across many papers | Claim → source, and citation → verification |
| Index size (per vendor) | 326M+ articles, 1.7B+ Smart Citations | 400M+ peer-reviewed papers and scholarly works | 138M+ papers | Searches Crossref, OpenAlex, Semantic Scholar, Google Scholar, PubMed, arXiv, Wanfang and CNKI |
| Sources searched | 44 publisher agreements + open access | Semantic Scholar, OpenAlex, PubMed, own crawl, publisher partners | Semantic Scholar–derived index + clinical trials | Crossref, OpenAlex, Semantic Scholar, Google Scholar, PubMed, arXiv, Wanfang and CNKI |
| Checks a reference list you already have | Partly — PDF upload, retraction and editorial notices; "nearly always miss a few references" | No | No | Yes — per-entry Verified / Mismatch / Not Found |
| Flags retractions | Yes | Yes (badge; excluded from analyses) | No dedicated feature | No dedicated feature |
| Free tier | No free plan listed on the pricing page; 7-day free trial of Basic or Pro | Free: basic paper search, 10 Pro messages, up to 3 Deep reviews/mo | Basic: free, limited Research Agent and Reports usage | 3 free credits per new account, no credit card |
| Entry paid tier | Basic $12/mo billed yearly | Pro $12/mo billed annually ($144/yr) | Pro $49/user/mo billed $588/yr | Trial $9 one-time, 15 credits |
| Next tier up | Pro $40/mo billed yearly | Deep $45/mo billed annually ($540/yr) | Scale $229/user/mo billed $2,748/yr | Monthly $19/mo (150 credits); Yearly $168/yr (1,500 credits) |
Two caveats on that table. Elicit's pricing page showed a single price list when we checked, with Pro at $49 per user per month billed annually and no separate academic Plus tier, so older academic prices quoted elsewhere may be out of date. And Consensus's public page showed annual prices only and did not load a month-to-month price for us, so we do not quote one.
How Citely fits
Citely is built for the two moments the three tools above leave open, and it is deliberately narrower than any of them.
The first is claim → source. You have a sentence and no citation. Paste it into the source finder and it searches Crossref, OpenAlex, Semantic Scholar, Google Scholar, PubMed and arXiv, plus Wanfang and CNKI for Chinese papers for papers that actually support it, returning ready-to-cite metadata in APA, MLA, Chicago, Harvard, IEEE, Vancouver, GB/T 7714 or BibTeX. A search that finds nothing is free, and the credits a search will use are shown before it starts. The input cap is 300 characters on every plan, which is an honest constraint worth knowing up front: this is one-sentence-at-a-time sourcing, not "paste the whole essay." For essay work where you are sourcing paragraph by paragraph, the academic source finder page walks through the same flow.
The second is citation → verification, which is where the comparison table has a column the others cannot fill. Paste a reference list into the citation checker and each entry comes back as one of three states. Verified means a published record matches the reference. Not Found means no database holds a matching record — most often a fabricated paper, sometimes grey literature or something brand new, and Citely shows the closest records so you can decide. Mismatch is the interesting one: the paper is real, but a detail such as the year, an author or the DOI does not line up, and Citely names the wrong field and shows the correct record. That third state is exactly the failure mode the JMIR study found in 45.4% of GPT-4o's non-fabricated citations, and it is the one that survives a casual eyeball check, because the reference looks fine and the paper does exist.
Pricing is by credits rather than per seat, and the credits a check will use are shown before it starts. Every new account gets 3 free credits, with no credit card required, so you can try it on real references before deciding anything. Paid plans are Trial at $9 one-time for 15 credits (valid one month), Monthly at $19 a month for 150 credits, Yearly at $168 a year (about $14 a month) for 1,500 credits, and Believer at $347 one-time for 5,000 credits over three years (prices checked 2026-11-02).
When you should stay with Scite, Consensus, or Elicit
Plenty of the time. These are better tools than Citely at the jobs they were built for, and switching would be a downgrade.
Stay with Scite if your question is about how the field received a paper. Nothing else tells you that 14 papers contrasted a finding you were about to cite as settled. If you are an editor or reviewer screening manuscripts for retracted references, Reference Check is purpose-built for that and Citely has no equivalent feature.
Stay with Consensus if you mostly need fast directional reads on empirical questions, or if you work clinically — Medical Mode plus automatic exclusion of retracted papers from summaries is a meaningful safety property, and the free tier is generous enough that many students never need to pay.
Stay with Elicit if you are screening hundreds or thousands of papers with inclusion criteria and extracting the same fields from each. Nothing in this comparison replaces that workflow, and no amount of per-reference verification substitutes for it.
And skip all of this if your bibliography is six references you personally pulled from a database and have open in tabs. You already know they exist. Verification tools earn their keep on references you did not personally retrieve — anything an LLM drafted, anything inherited from a collaborator's file, anything copied from another paper's reference list without opening the paper.
Key Takeaways
- Scite, Consensus, and Elicit all return real papers, because they retrieve from indexes. The risk in your draft is not what they returned — it is the references that arrived some other way.
- Scite's Reference Check is the only bibliography-level tool of the three, but its own docs say it "will nearly always miss a few references," so a missing entry cannot be read as a fabricated one.
- Real citation, wrong metadata is the common failure, not outright invention: 45.4% of GPT-4o's non-fabricated citations carried bibliographic errors in the 2025 JMIR study, usually a bad DOI.
- Fabrication risk rises on obscure topics — 6% for major depressive disorder versus 29% for body dysmorphic disorder in that same study. The niche chapter of your thesis is the risky one.
- Pick by job, not by brand: field reception → Scite; yes/no evidence → Consensus; large-scale screening → Elicit; per-entry existence and metadata check → a citation checker.
