Every tool on this page, by where its references come from
On 11 August 2026 we went to each vendor's own website and looked for one thing: does it say where a citation it hands you actually comes from. Not how many papers it claims, not how good the writing is. Just the provenance.
One result stands out and it is the reason this table exists. Of every tool here that claims a paper corpus, exactly one names the databases behind it. Elicit says, in its own help centre, that it "searches across over 138 million academic papers from Semantic Scholar, PubMed, and OpenAlex." Jenni, Paperpal, Consensus and Sourcely all publish a headline number in the same range and name nothing at all. A number with no named source is not a check you can run; it is a number.
| Tool | Where a reference comes from | Corpus the vendor publishes | Databases named |
|---|---|---|---|
| Elicit | A literature search across named indexes, presented as extraction tables rather than prose | "over 138 million academic papers" | Yes. Semantic Scholar, PubMed, OpenAlex |
| CiteOwl | Papers the agent retrieved and read before the claim was written, with the supporting quote stored on the citation | Not a corpus of its own | Yes. OpenAlex for papers, Crossref and Unpaywall for DOI resolution, Exa for the web |
| Consensus | A search over peer-reviewed papers, answering what a field agrees on | "over 200 million peer-reviewed academic research papers" | No |
| Jenni AI | Its own index, plus anything you import from Zotero, Mendeley or a PDF | "200M+ papers", described as "our full academic index" | No |
| Paperpal | Its own article collection, offered inside an academic editor | "250M+ research articles" | No |
| Sourcely | A search run against text you already wrote, to find backing after the fact | "over 200 million papers" | No |
| Yomu AI | Sourcely. Its citation feature is "powered by Sourcely" rather than an index of its own | None of its own | It names its supplier, which is more than most |
| Scite | Citation statements pulled out of full text, so you see how a paper was cited, supporting or contrasting | "More Than 1 billion citation statements extracted and analyzed", published by its parent company | No |
| Perplexity | Live web search, with inline citations to the pages it read | The open web | Not applicable. You can open every link it gives you |
| ChatGPT, Claude, Gemini | The model itself, unless web search or a deep-research mode is running. With search on, the pages it fetched | None | Not applicable, and this is the point: with search off there is no source to name |
| Grammarly | Citation Finder "pulls in trusted sources" and formats them | None stated | No. "Trusted sources" is the whole disclosure |
| Writefull | Nothing. Writefull Cite "flags sentences that have no citation but need one" and leaves finding it to you | Not applicable | Not applicable |
| Trinka | Nothing. It checks and formats citations you already have | Not applicable | Not applicable |
| Scribbr | Human editors, plus a free generator that looks a source up "by its title, URL, ISBN, or DOI" | Not applicable | Not applicable |
Quotes are from each vendor's own pages on 11 August 2026 and vendors rewrite these pages: Elicit, Consensus, Jenni, Paperpal, Sourcely, Yomu, Scite via Research Solutions, Grammarly, Writefull, Trinka and Scribbr.
Two gaps we are not going to paper over. SciSpace's own site returned nothing readable to repeated requests that day, so it has no row rather than a guessed one, and the Scite figure comes from its parent company's page because scite.ai would not serve one either. Prices are absent on purpose: they change monthly, they arrived in three different currencies depending on where the request came from, and several were sale prices dressed as list. The individual comparisons below carry them with the date attached.
What the four bands actually mean for your paper
- Read the source, then write the claim. The reference cannot be invented, because the sentence was written from a paper that was already open. This is the only arrangement where a fabricated citation is structurally hard rather than merely unlikely.
- Retrieve and report, but do not draft. Elicit, Consensus, Scite and Perplexity give you real, openable things and leave the writing to you. For a literature review or a methods chapter this is often the better half of the job to automate.
- Draft, then attach references from an unnamed pile. Fast and fluent, and the reference is real often enough that spot-checking feels unnecessary. Check it anyway. You cannot audit a corpus nobody names.
- Fix the words, supply nothing. Grammarly, Writefull and Trinka are good at what they do and none of them will find you a source. Writefull is the honest edge case: it tells you a sentence needs a citation and does not pretend to have one.
The test is the same whichever band you buy from, and it takes five minutes. Ask the tool for three references on a topic you know. Copy each DOI into a checker or paste it after doi.org/. Then open the one that resolved and find the sentence it is supposed to support. A tool that survives that is worth paying for. A tool that fails it is worth exactly nothing to you at 2am the night before a deadline, whatever it costs.
Which comparison to read
- Deciding at all. Start with the ranked overview, then can AI write my research paper, which carries the essay case, or the literature review generator guide depending on the job.
- Coming from a chatbot. how we compare to what you already pay for, ChatGPT vs Claude, Gemini vs ChatGPT and Claude vs Gemini, one pair at a time.
- Coming from a reading tool. the SciSpace alternative and the Scite alternative. If the question is accuracy rather than price, does Perplexity make up citations collects what has been measured, and does Gemini make up citations covers the assistant most students meet without choosing it.
- Coming from an editor. the Grammarly alternative, and Paperpal, Writefull and Trinka on the comparison page, one section each.
- Coming from a drafting tool. Jenni, and Yomu and Sourcely on the comparison page.
- Already have a draft. Sample five references before you fix anything, then repair or replace the list; the integrity hub covers the record you keep while you do it.