Consensus

What is Consensus?

Consensus is the AI-powered academic search engine, built to search and analyze peer-reviewed research rather than the open web. The problem it targets is specific: a general chatbot will answer a research question fluently and may invent the sources, while a traditional database will return five hundred results and leave the synthesis to you. Consensus sits between those failure modes — you ask a research question in plain language and get an answer where every claim traces back to a paper you can open.

The corpus is the foundation of the pitch. The company says it draws on 250M+ research papers, including licensed full text content from leading publishers, and the publisher logos on its homepage are not decorative: Wiley, Taylor & Francis, Sage Publishing, ACS, APA and AAAS appear, with several of those partnerships announced individually on the company blog. Licensed full text matters more than raw paper counts, because the difference between searching abstracts and searching full text is the difference between finding a paper and finding the finding inside it.

The company was founded in 2021 by Christian Salem and Eric Olson, two Division 1 athlete teammates from families of researchers and teachers, and it has raised $45M in funding from Union Square Ventures, GreatPoint Ventures, Nat Friedman and Daniel Gross. It has grown fast against its own baseline: a bylined VentureBeat report from April 2023 described a Boston-based company that had just secured $3 million in seed funding led by Draper Associates, with nearly 200,000 registered users at that point, against the 10 million-plus the company claims today. That report also documented the original technical approach — a customized version of GPT-4 built in partnership with OpenAI, applied to a corpus of research papers. Today the company reports 12,500+ universities represented among its users and 150+ million research questions handled to date, figures that are self-reported and, as covered further below, do not always agree across its own pages.

Core Features

• Cited answers over peer-reviewed literature: The base behavior is a search that answers rather than merely lists. Every response is tied back to specific papers, which is the mechanism that separates it from a general-purpose model answering from memory. Free accounts keep basic paper search permanently, with AI analysis metered above that.
• Deep Search for literature review automation: The flagship capability, and the one the top pricing tier is named after. Rather than running a single query, it builds a comprehensive search strategy, expanding key terms, identifying conflicting arguments, and exploring the citation graph. The "identifying conflicting arguments" part is the substantive claim — a literature review that only surfaces agreement is a biased review, and disagreement is usually the thing a researcher most needs to find.
• The Consensus Meter for evidence agreement: For questions that can be posed as yes-or-no, the Consensus Meter instantly shows how much the evidence agrees or disagrees. This is a genuinely different output format from a prose summary: it gives you the distribution of findings across the literature rather than one synthesized narrative that may quietly average away a real controversy.
• Medical mode for clinical sources: A filter that narrows results to about 50,000 clinical guidelines and 8M articles from the top 1,000 medical journals. Useful for clinical questions where the relevant evidence hierarchy is guidelines and high-impact journals rather than the broad literature — though the terms place hard limits on what this may be used for, covered below.
• Natural-language filters: You can specify timeframes, populations, designs, and more directly in your prompt, and the system applies the corresponding filters. This removes the usual friction of translating a research question into a database's filter syntax, which is where most people give up on advanced search.
• Citation graph and research tooling: Paid tiers unlock unlimited access to the full set of research tools including search, analysis and the citation graph — the mechanism for moving from one relevant paper outward to what cited it and what it cited.
• API and MCP access: Programmatic access is a metered part of the subscription rather than a separate product, which makes the corpus reachable from other tools and agent workflows.

Use Cases

1. Running a literature review for a thesis or systematic review: The classic multi-day task. The intended flow is to pose the research question, let Deep Search build and expand the search strategy, then work outward through the citation graph. The genuine advantage over a database is that conflicting findings are surfaced as a feature rather than buried in result ranking. The honest caveat is that this does not replace a registered systematic review protocol — it accelerates scoping, not methodological rigor.
2. Settling a yes-or-no empirical question quickly: "Does intermittent fasting lead to meaningful long-term weight loss?" is the shape of question the Consensus Meter is built for. Instead of one confident paragraph, you get a read on how the body of evidence actually splits, which is the correct answer format when the literature is genuinely divided.
3. Clinical lookup against guidelines and top journals: Medical mode narrows to guidelines and high-impact journals so that a question about first-line treatment returns the sources a clinician would actually cite. This is a starting point for finding evidence, explicitly not a decision-support system.
4. Coursework and student research: Students are the largest visible user group, and the value is specific — the tool produces cited sources to write from rather than prose to hand in. The vendor leans on this framing deliberately, positioning itself as compatible with academic integrity policies rather than in tension with them.
5. Institutional deployment through a library: Libraries and departments buy access on behalf of their users, either for a lab or class or across the whole institution, which is why over 170 university libraries partner with Consensus to give students and faculty access.
6. Embedding research retrieval in other tools: Through API and MCP access, the corpus becomes a retrieval source for an internal application or agent workflow — relevant for teams building research assistants who want peer-reviewed grounding rather than open-web scraping.

How to use Consensus

1. Open the site and run a search without an account first. Basic paper search is free and permanent, so you can test whether the corpus covers your field before creating anything.
2. Phrase the question the way you would ask a knowledgeable colleague, not the way you would query a database. The system is designed to parse natural language, including constraints — put the timeframe, population, or study design directly in the sentence rather than hunting for filter menus.
3. Choose the mode that matches the question. A quick factual lookup, a yes-or-no question suited to the Consensus Meter, and a full literature review are three different jobs, and Deep Search is the expensive one to be spent deliberately.
4. Turn on Medical mode if the question is clinical, so results narrow to guidelines and top-tier medical journals rather than the whole corpus.
5. Read the cited papers, do not stop at the synthesis. The summary is a map, not the territory; the citations exist so you can verify that each paper actually says what the summary claims it says.
6. Work outward through the citation graph from the most relevant result. This is usually more productive than reformulating the original query repeatedly, because a single well-targeted paper points to its own intellectual neighborhood.
7. Check whether your institution already provides access before paying. With university libraries partnering at scale and a student and clinician discount available, individual subscription may be unnecessary.

Tips & Best Practices

• Claim the discount you qualify for. Students and faculty with a valid school email and US clinicians with a valid NPI number can claim up to 40% off, which materially changes the calculus on the paid tiers.
• Spend Deep reviews deliberately. They are the scarcest resource on every tier — three per month on free, fifteen on Pro, two hundred on Deep. Use ordinary search for scoping and reserve Deep Search for questions where a comprehensive strategy actually matters.
• Prefer questions the evidence can answer. The Consensus Meter works on empirical yes-or-no questions. Asking it something normative or definitional produces a weaker result than asking a well-posed causal question.
• Treat disagreement as signal, not noise. When the meter shows a split, that split is often the most useful thing you will learn. Resist the temptation to reformulate the question until you get a clean consensus.
• Verify every citation you intend to use. The tool retrieves and summarizes; responsibility for accuracy stays with you, and the terms are explicit that no warranty of accuracy is given.
• Do not assume complete coverage. Users report gaps, including papers they knew existed that did not surface, so a null result is not proof of absence — especially for very recent, non-English, or paywalled work.
• Watch the billing cycle if you subscribe. Annual renewal complaints are the dominant theme in third-party reviews, so note your renewal date and cancel through the account settings well before it rather than relying on a reminder.

Who is Consensus for?

• Graduate students and PhD candidates: The heaviest realistic users, for whom literature review is a recurring, time-dominated task rather than an occasional one.
• Academic researchers and faculty: For scoping new areas, staying current, and finding the disagreement in a field quickly.
• Clinicians and medical professionals: Served by Medical mode and a dedicated NPI-based discount, within the boundary that the service is explicitly not medical advice.
• Undergraduate students: For coursework requiring real cited sources, where the alternative is a general chatbot that may fabricate references.
• University libraries and departments: The institutional buyer, purchasing for labs, classes, or an entire campus.
• Evidence-oriented professionals outside academia: Policy analysts, science journalists, and health and fitness practitioners who need to cite primary literature rather than secondary coverage.
• Teams building research-grounded AI applications: The API and MCP audience, wanting peer-reviewed retrieval as a grounding layer.

Platforms

• Web application: The primary surface, running in the browser with a thread-based interface. No installation, and basic search works before sign-in.
• API and MCP access: Programmatic access included as a metered allowance on paid tiers, letting other applications and agent frameworks query the corpus.
• Institutional access through libraries: A distribution channel more than a platform — many users reach the product through a university subscription rather than paying directly.
• No dedicated mobile or desktop applications: There is no evidence of native app builds; the responsive web application serves all devices.
• Publisher integrations upstream: Licensed full text arrangements with major publishers determine what the search can actually reach, making these partnerships part of the platform's real capability surface.

Pricing & Plans

The free tier is unusually well-defined for this category, and worth reading precisely. It keeps basic paper search permanently free but explicitly without AI analysis, and meters the intelligent features: 15 Pro messages per month and up to 3 Deep reviews per month. That structure means the free tier is genuinely usable as a search tool indefinitely, while the AI synthesis that constitutes the actual product is a sampler.

Paid tiers split by how much literature review you do. Pro runs $12 per month or $144 annually and lifts Pro messages to unlimited while allowing 15 Deep reviews per month, plus unlimited access to the full research toolset. Deep costs $45 per month or $540 annually and is aimed at researchers or clinicians conducting frequent literature reviews, raising the allowance to 200 Deep reviews per month. The distinguishing variable between the two tiers is Deep reviews, not general usage — if you run a handful of comprehensive reviews a month, Pro suffices; if literature review is your daily work, Deep is the tier built for you.

Programmatic use is metered separately: Pro includes 250 API and MCP uses per month and Deep includes 1,000, metered at $0.10 per use once the included allowance runs out, with monthly caps above that. Anyone planning to build on the API should model this cost explicitly rather than treating it as included.

There is a substantial identity discount: students and faculty with a valid school email and US clinicians with a valid NPI number can claim up to 40% off. For institutions, labs, classes, or teams of under 200 users can buy directly, while a sitewide license requires contacting the company — so the first question for any university-affiliated user is whether access already exists through the library. (Pricing changes; verify current figures on the vendor's pricing page before subscribing.)

Alternatives

• Semantic Scholar: The closest comparison and a genuinely different proposition — a free, non-profit corpus from the Allen Institute for AI, with no subscription and no paywall on its own tooling. Less polished as a synthesis product, but the reference point for anyone unwilling to pay.
• Elicit: The most direct competitor on literature review automation, strong on extracting structured data across many papers into a table. Often preferred for systematic-review-style extraction workflows.
• Scite: Focused on citation context rather than answers — it tells you whether subsequent papers supported or contradicted a given claim, which is a complementary question to the one Consensus answers.
• Google Scholar: Free, unmatched in breadth, and the default for most researchers. No AI synthesis and a dated interface, but nothing beats it for simply finding whether something exists.
• General assistants like ChatGPT or Perplexity: Faster and more flexible for open-ended questions, but without a guaranteed peer-reviewed corpus underneath. The failure mode is a plausible citation that does not exist, which is precisely what a purpose-built academic engine is designed to prevent.
• Traditional databases such as PubMed, Scopus, or Web of Science: The institutional standard for systematic reviews, with rigorous controlled vocabulary and reproducible query syntax. Harder to use, but a registered systematic review still needs them.

Limitations & Considerations

The most consequential limitation is what the legal terms say about accuracy, and it is worth quoting rather than paraphrasing. The service is provided on an as is, where is, as available, and with all faults basis and without any warranty, and the company expressly disclaims warranties of completeness, accuracy and reliability. For a product whose entire value proposition is trustworthy evidence, that gap between marketing and contract is the single most important thing a prospective buyer should register.

The medical boundary is stricter still, and sits in direct tension with the existence of a Medical mode. The terms state the company does not provide medical advice, and information accessed through the site is not intended for the diagnosis, evaluation, or treatment of any individual. Medical mode should therefore be read as a source filter for literature discovery, never as clinical decision support. Separately, children under 13 are not allowed to use the Services.

Coverage is not complete and should not be assumed to be. One reviewer reported that they do not have access to paywalled articles and that papers cited in their own dissertation did not surface in search. Whether or not that reflects typical behavior, the operational lesson holds for any retrieval tool: absence of a result is not evidence of absence, and a literature review that must be exhaustive still requires a traditional database.

Third-party sentiment is poor, though the sample is too small to treat as a verdict. The listing carries a TrustScore of 2.4 across just 6 reviews in the last twelve months, with 83% at one star, and Trustpilot itself notes the company has never invited customers to review, so the sample may not be representative. Under twenty reviews is directional only, so the useful signal is not the score but the recurring themes, and those are consistent: Trustpilot's own top mentions for the listing are Cancellation and Subscription. Multiple reviewers describe being charged after signing up for a free trial without receiving any clear notification that the trial was ending, followed by unanswered refund requests. For context on the score itself, the same comparison widget shows Perplexity sits at 1.5 across 753 reviews and Claude at 1.5 across roughly 2,000 — low ratings are endemic to subscription AI products and this score should not be read in isolation. G2 could not be verified first-hand as both access channels were blocked, so no G2 figure is quoted here.

One billing complaint deserves separate attention because it is externally corroborated rather than merely asserted. One reviewer reports monthly MCP lookups dropping from 1,000 to 250 partway through an annual contract, requiring an upgrade to a substantially more expensive tier to restore the original allowance. The current pricing page is consistent with that account: 1,000 API and MCP uses now sit on the $45 Deep tier while 250 sit on the $12 Pro tier. Anyone building a dependency on the API allowance should assume the quota is subject to revision within a contract term and plan accordingly.

Vendor-reported figures are internally inconsistent, which argues for discounting them rather than repeating them. On corpus size, the persistent sidebar advertises 200M+ peer reviewed research papers while the homepage body claims 250M+. On users, the pricing page says over 5 million researchers, students, and clinicians while the homepage says 10 million and the about page says 10+ million. University counts appear variously as 5,000, 10,000 and 12,500 — the first two on the same page. All were captured the same day.

On data handling the news is genuinely good, and it is the strongest differentiator for anyone handling unpublished research questions. The privacy policy states that your queries, content, and personal information are never used to train AI models, neither its own nor those of any third party, alongside a commitment not to sell personal information. Two caveats belong beside that: the policy collects any questions, prompts, or content you submit, as well as the output generated for you, and it states that it is not subject to regulation under the California Consumer Privacy Act — an unusual assertion that limits the statutory rights a California resident might otherwise expect. The policy also notes that in the event of a merger, acquisition, or sale of assets, your information may be transferred as part of that transaction.

FAQ

Q1. How is this different from asking ChatGPT a research question?

The corpus and the citations. A general assistant answers from model weights and may produce a fluent citation to a paper that does not exist, whereas this is an academic search engine that retrieves from peer-reviewed literature and ties responses back to real papers you can open. The trade-off is flexibility: a general assistant handles any question, while this is built for questions the research literature can actually answer.

Q2. Is the free tier actually usable, or just a demo?

Both, depending on what you need. Basic paper search is free permanently, but the vendor is explicit that the free tier is search without any AI analysis, with the intelligent features metered at 15 Pro messages per month and up to 3 Deep reviews per month. As a finding tool it works indefinitely; as a synthesis tool it is a sampler.

Q3. What exactly is a Deep review and why is it rationed?

It is the literature review automation feature, and it is the scarce resource on every tier because it is computationally expensive. Rather than one query it builds a comprehensive search strategy, expanding key terms, identifying conflicting arguments, and exploring the citation graph. Allowances run three per month on free, fifteen on Pro, and two hundred on Deep, which is the main axis distinguishing the paid tiers.

Q4. Can I rely on it for clinical decisions?

No, and the terms are unambiguous on this. The company does not provide medical advice, and information accessed through the site is not intended for the diagnosis, evaluation, or treatment of any individual. Medical mode narrows sources to about 50,000 clinical guidelines and 8M articles from the top 1,000 medical journals, which makes it a better literature discovery filter for clinical questions — but discovery is not decision support.

Q5. Are my search queries used to train AI models?

The policy says no, explicitly. Your queries, content, and personal information are never used to train AI models, neither its own nor those of any third party, and the company states it does not sell personal information. Note that data is still collected and stored — the policy covers any questions, prompts, or content you submit, as well as the output generated for you — so the commitment is about training use, not about non-collection.

Q6. What should I know before starting a subscription?

Watch the renewal terms. Third-party reviews cluster heavily on billing: reviewers describe being charged after signing up for a free trial without receiving any clear notification that the trial was ending, and Trustpilot's own top mentions for the listing are Cancellation and Subscription. Before subscribing, check whether your institution already provides access, claim the student or clinician discount if you qualify, and diarize the renewal date.

Q7. Does it cover every paper in my field?

No retrieval tool does, and you should not assume it. The vendor claims a corpus of 250M+ research papers including licensed full text content from leading publishers, but at least one user reports that they do not have access to paywalled articles and that known papers failed to surface. Treat a null result as inconclusive rather than as evidence that nothing exists.

Q8. Is it credible as a company, or an unproven startup?

It has substantive backing and independent coverage. It was founded in 2021 by Christian Salem and Eric Olson, two Division 1 athlete teammates from families of researchers and teachers, and has raised $45M in funding from Union Square Ventures, GreatPoint Ventures, Nat Friedman and Daniel Gross. Independent reporting corroborates the trajectory: a bylined VentureBeat report from April 2023 described a Boston-based company that had just secured $3 million in seed funding led by Draper Associates. Publisher partnerships with Wiley, Taylor & Francis, Sage Publishing, ACS, APA and AAAS are a further external signal, since those agreements require counterparty diligence.

Q9. Can I use it through my university instead of paying?

Very possibly, and it is worth checking first. Over 170 university libraries partner with Consensus to give students and faculty access, and institutionally labs, classes, or teams of under 200 users can buy directly, while a sitewide license requires contacting the company. If your library has not subscribed, the identity discount is the next best route.

Q10. Should I trust the user and paper counts on the site?

Treat them as vendor-reported and approximate, because they do not agree with each other. The persistent sidebar advertises 200M+ peer reviewed research papers while the homepage body claims 250M+; the pricing page says over 5 million researchers, students, and clinicians while the homepage says 10 million. University counts range across 5,000, 10,000 and 12,500. The directionally reliable facts are the funding, the founding date, and the publisher partnerships, all of which have external corroboration.