# Elicit: reviews and analysis

> Searches 138 million papers and pulls the data into a table with the citation attached.

- Canonical: https://serchai.com/en/reviews/elicit/
- Site: Serchai (https://serchai.com) — AI tools comparator
- Language: en
- Updated: 2026-08-02

---

## Verdict

Elicit does with scientific literature what an AI notebook does with loose documents: it searches by meaning across more than 138 million papers, pulls the data into a comparable table and ties every claim to the sentence it came from. It works with your own PDFs too, though table extraction from them is paid. The serious objection is about reach rather than price: an independent measurement published in Cochrane gives its searches 39.5% average sensitivity, so it supports traditional searching rather than replacing it.

**Best for:** Anyone reviewing scientific literature who needs each paper’s data point with a checkable citation attached.

**Rating:** 3.5/5

## Pros

- Every claim carries the sentence from the paper it came from
- Extracts the same data point from dozens of papers into one comparable table
- A declared corpus of more than 138 million papers, searched by meaning

## Cons

- An independent measurement gives its searches 39.5% average sensitivity
- Table extraction from your own PDFs requires a paid plan
- The corpus is academic: no books, dissertations or company documents

## Key facts

- Price: Permanent free tier + paid from 49 USD
- Free trial: No
- Platforms: Web
- Categories: [Assistants](https://serchai.com/en/best-ai/assistants/)
- Official website: https://elicit.com

## What the internet says (agentic sweep)

The praise that repeats comes from university libraries rather than blogs: Elicit searches a declared corpus of more than 138 million papers from Semantic Scholar, PubMed and OpenAlex, pulls the data into a table and ties every claim to the sentence it came from. It does work over your own documents too, since the library takes PDFs with no storage cap and you can question several at once, though the first serious limit shows up there: table extraction from your own PDFs requires a paid plan, and simultaneous chat stops at four papers on free and eight on paid. The hardest objection is an independent measurement. A study in Cochrane Evidence Synthesis and Methods measured the sensitivity of its searches across four real reviews and got 39.5% on average, against 94.5% for traditional searching, which collides head-on with the recall the vendor advertises. Deakin's library adds two warnings a buyer should read before paying: searches are not reproducible, and uploading other people's articles may breach journal licences. And the entry price on the general catalogue is $49 a month with annual payment, a long jump from a thin free plan.

- Sweep date: 2026-08-02
- Derived score: 3.5/5

### Axes

- Results: 3.6/5
- Control: 4/5
- Real price: 2.8/5
- Integration: 3.4/5
- Support: 3.8/5

### Recurring themes in favor

- Every claim is tied to the sentence in the paper it came from, so checking it takes a click rather than an afternoon (strong theme)
- A declared corpus of more than 138 million papers from Semantic Scholar, PubMed and OpenAlex, searched by meaning rather than exact keyword (strong theme)
- It pulls data from many papers into a comparable table, the mechanical work that eats the hours of a review (strong theme)
- The library takes your own PDFs with no declared storage cap and you can question several at once (present theme)

### Recurring themes against

- An independent measurement published in Cochrane gives its searches 39.5% average sensitivity, against 94.5% for traditional searching (strong theme)
- The own-documents use case sits behind the paywall: table extraction from your PDFs is not in the free plan (strong theme)
- Searches are not reproducible, which leaves them short of what a systematic review has to be able to document (strong theme)
- The corpus is academic and nothing else: no books, no dissertations, no non-academic publications (present theme)
- A university library advises against uploading third-party articles over the risk of breaching journal licences, which is exactly what the product invites you to do (present theme)
- The price jump is abrupt: from a thin free plan to $49 a month with annual payment on the general catalogue (present theme)

### Sweep sources

- [Official pricing] https://elicit.com/pricing — Página oficial de precios, leída dos veces de forma independiente sobre el HTML servido. No tiene JSON-LD ni declara ningún código de divisa: solo imprime el glifo del dólar. Tiene dos conmutadores, uno de facturación y otro que cambia el catálogo entero entre «Industry» y «Academic», y el estado por defecto en escritorio es anual con catálogo Industry. Ahí el plan de pago de entrada es «Pro» a «$49» con la etiqueta literal «Per user/month, billed as $588 annually», y suelto «$75» con «Per user/month, billed monthly». El catálogo académico enseña otro escalón, «Pro» a «$39» «billed as $480 annually» y un plan «Plus» a «$11». El gratuito, «Basic», se describe como «Limited usage for Research Agent and Research Reports» y la cadena «free trial» no aparece en toda la página
- [Docs] https://support.elicit.com/en/articles/14758040-elicit-s-source-for-papers — Centro de ayuda oficial, artículo sobre el origen del corpus. Dice literalmente «Elicit searches across over 138 million academic papers from Semantic Scholar, PubMed, and OpenAlex» y acota el alcance con la frase que más importa para decidir si encaja con tu trabajo: «Elicit focuses on peer-reviewed journal articles, conference proceedings, preprints, and working papers. It doesn't include books, dissertations, or non-academic publications». Conviene saber que la cifra del corpus no es estable entre páginas del propio fabricante, que en otros sitios sigue diciendo 125 millones
- [Press] https://pmc.ncbi.nlm.nih.gov/articles/PMC12483133/ — Lau O. y Golder S., «Comparison of Elicit AI and Traditional Literature Searching in Evidence Syntheses Using Four Case Studies», en Cochrane Evidence Synthesis and Methods, 27 de septiembre de 2025. Es la única medición independiente y revisada por pares de este expediente, y va justo contra la afirmación comercial: «the sensitivity of Elicit was poor, averaging 39.5% (25.5-69.2%)» frente al 94,5 % de la búsqueda tradicional, con una precisión media del 41,8 % frente al 7,55 %. Su conclusión es que sirve de apoyo y no de sustituto, «not sensitive enough to replace traditional searching»
- [Review sites] https://deakin.libguides.com/AI-Evaluations/Elicit — Evaluación institucional de la biblioteca de la Universidad de Deakin, hecha para decidir si la recomiendan a su comunidad investigadora. Aporta tres avisos que no salen en ninguna reseña comercial: «Developers claim Elicit outputs are about 80-90% accurate, urging users to critically evaluate results», «Searches are not directly reproducible and currently lack the transparency required for systematic review reporting» y «There is an increased likelihood of copyright violation by users of Elicit in the upload functionality, due to nuanced rights and permissions for different articles/journals». Detalla además los topes del plan gratuito, «screen up to 50 papers and extract data from 8 papers»
- [Review sites] https://libguides.library.arizona.edu/ai-researchers/elicit — Guía de la biblioteca de la Universidad de Arizona, una segunda institución independiente de la anterior. Describe el trabajo que hace en palabras de quien lo enseña a investigadores, «find relevant papers without perfect keyword match, summarize takeaways from the paper specific to your question, and extract key information from the papers», y remite expresamente a las páginas de límites, fiabilidad y uso responsable de la herramienta antes de recomendarla. Que dos bibliotecas universitarias mantengan guía propia es la mejor señal de adopción real que tiene esta herramienta


> **TLDR:** Elicit is a research assistant that searches more than 138 million papers by meaning, pulls the same data point from dozens of them into one table, and ties every claim to the exact sentence it came from. It also works with PDFs you upload, though table extraction from your own library is paid. The serious objection is not the price: an independent measurement published in Cochrane put the sensitivity of its searches at 39.5% on average, so it supports traditional searching rather than replacing it.

## What Elicit is and how it works

Elicit started inside Ought, a research lab, and spun out as a company in 2023. That explains its shape, which is not a chatbot with a text box. The workflow is literature review: you ask a research question, the tool searches its paper index by meaning, screens out what does not fit, then extracts from each paper the fields you asked for, one per column, into a table you can sort and export.

The corpus is declared by its own help centre: more than 138 million papers from Semantic Scholar, PubMed and OpenAlex, with clinical trial records as an option. It also declares what is excluded, which matters just as much when deciding whether it suits you: no books, no dissertations, no non-academic publications. If your working material is internal reports, contracts or minutes, this is not the place.

The other piece is the citation. Every claim in an answer is anchored to the sentence in the paper that supports it, so checking takes one click. That is the real difference against asking a general assistant the same thing, which can hand you a well-written answer and a reference that does not say what it promises.

## What using it is like day to day

The first session is surprising for how little it resembles a chat. You type the question, the tool takes its time, and what comes back is not a paragraph but a set of papers and a table under construction. You work it like a spreadsheet: you add columns (sample size, method, main outcome, country) and it fills them by reading the full text.

It does work over your own documents, with caveats worth knowing before you pay. The library accepts PDFs with no declared storage cap and they can be imported in batches. Questioning several at once is on every plan, but the number you can cross in a single conversation is small: four on free and eight on paid. Table extraction from your own PDFs, which is what actually saves the labour, is not in the free plan.

And here comes the objection a buyer should read in full. In September 2025 a study in Cochrane Evidence Synthesis and Methods compared its searches with traditional searching across four real reviews. The result was 39.5% average sensitivity against 94.5% for the classic method, with considerably better precision, 41.8% against 7.55%. Translated: it finds little of what is relevant, but almost everything it brings is worth having. The authors conclude it is in no position to replace standard searching.

On top of that sit two institutional warnings from Deakin University's library, which evaluated it for its research community. One, that searches are not reproducible and therefore fall short of what a systematic review has to document. Two, that uploading third-party articles may breach journal licensing, which is exactly what the product invites you to do.

## Price and plans

There is a free plan, Basic, described by the page itself as limited usage of the research agent and reports. It is enough to see the flow and little else: the caps published in Deakin's evaluation are 50 papers screened and data extracted from 8.

The entry paid plan on the general catalogue is Pro, from $49 per user per month billed annually, which the page itself renders as $588 a year. Month to month, the same plan runs $75. The page declares no currency code anywhere and offers no selector, so what is published is the dollar and nothing more.

One detail sits behind a dropdown and changes the arithmetic a lot: the academic catalogue. Switch to it and a Plus plan appears from $11, with Pro dropping to $39 a month billed annually. If you research from a university, that is the door to come in through, and it is worth checking before signing up at the general rate.

## Who it is for (and who it is not for)

Elicit is for people working with scientific literature who need the same data point from many papers laid out in a column, with the citation attached so they can defend it. Scoping reviews, grant proposals, states of the art, and everything currently done by reading a hundred abstracts by hand.

It is not for anyone wanting to feed it documents that are not papers. Minutes, contracts, company reports or study material are better handled in [NotebookLM](https://serchai.com/en/reviews/notebooklm/), which takes any source and does not charge to question fifty at once. Nor is it for anyone who needs the search to be exhaustive and documentable, because there the independent measurement sides with the old method.

And it is not for a quick answer to a general question: for that there is a search engine with judgement like [Kagi](https://serchai.com/en/reviews/kagi/) or an assistant like [Claude](https://serchai.com/en/reviews/claude/). The full map of the category is in the [best AI assistants](https://serchai.com/en/best-ai/assistants/).

## Frequently Asked Questions

### Can I upload my own PDFs?

Yes, and the library declares no storage cap. The limit is how many you can cross in a single conversation, four on the free plan and eight on paid, and the fact that table extraction from your PDFs requires a paid plan.

### Does it find everything published on a topic?

No, and this has been measured by third parties. A Cochrane study calculated 39.5% average sensitivity across four real reviews. It works to start and to narrow, not to call a systematic search complete.

### Is it good enough for a formal systematic review?

As support yes, as the method no. Its searches are not reproducible, and reproducibility is a reporting requirement in that kind of work.

### Is there an academic discount?

There is a separate academic catalogue, with a Plus plan from $11 and Pro from $39 a month billed annually, against $49 on the general catalogue. It sits behind a dropdown on the pricing page.

### Can it be used for free?

It can be tried for free. The Basic plan screens a few dozen papers and extracts data from a handful, and the work that justifies the tool starts just above that ceiling.

## Alternatives

- [NotebookLM](https://serchai.com/en/reviews/notebooklm/) — The research notebook that answers only from your own documents, with the citation attached. (4.1/5)