Consensus

Positioning

Consensus occupies a different slot from Elicit or Scite in the research-tool landscape, and getting the distinction clear is most of the buying decision. It is a search engine first: you type a research question in plain English and it returns relevant papers from a corpus of over 200 million peer-reviewed documents, with AI-generated summaries and — the feature the product is named after — a Consensus Meter that shows how strongly the retrieved literature supports or contradicts a claim. It answers "what does the science say?" in minutes. It does not build you an extraction table or run a PRISMA-style review.

The company leans on transparency as its differentiator: answers are grounded in cited papers you can open, and the interface shows the supporting evidence rather than a generated paragraph pretending to be authority. The pricing page claims more than 5 million researchers, students and clinicians use it — a vendor figure I cannot verify independently, though the product is widely assigned in university settings and has institutional plans that suggest real penetration.

I have used Consensus on and off for fact-checking work and returned to it over several weeks for this review, running it against questions I already knew the literature on (so I could judge whether it found the right studies) plus a few open questions where I wanted the meter to teach me something.

What it actually does

You ask a question — "does creatine supplementation improve memory in healthy adults?", "are remote teams less productive?" — and Consensus runs a semantic search over its corpus, which it describes as 200+ million scientific documents, primarily peer-reviewed journal articles plus some preprints and conference proceedings. Results come back as individual studies with a one-line AI summary of each paper's finding, and the Consensus Meter aggregates the retrieved papers into a simple visual showing the balance of Yes / Possibly / No answers to your question. A new Medical Mode, introduced in late November 2025, tunes the experience for clinical questions.

On top of search, the product now has an agentic layer with distinct consumption limits: Pro messages (deeper per-result analysis), Deep reviews (multi-step research reviews that synthesise across studies), and Study Snapshots (structured one-page summaries of a single study). There is also an API and an MCP server, meaning you can wire Consensus into an AI assistant as a retrieval source — I did exactly that with an MCP-compatible client and it worked as advertised.

Pricing at time of writing, straight from the vendor's help-centre plans article: Free is $0 and includes unlimited Papers searches, 10 Pro messages per month, 3 Deep reviews per month, 10 Study Snapshots per month and 30 API/MCP calls per month. Pro is $20 per month or $144 per year (equivalent to $12 per month) — unlimited Pro messages, 15 Deep reviews, unlimited Study Snapshots, 500 API/MCP calls per month. Deep is $65 per month or $540 per year (equivalent to $45 per month) — everything in Pro plus 200 Deep reviews and 2,000 API/MCP calls per month. Teams is custom-priced (Pro features plus 50 Deep reviews per month per user, volume discounts up to 200 seats, centralised billing), and Enterprise is custom for organisations above roughly 200 users, with library integrations, training and a development-partner programme.

My testing notes

My first test was recall against known literature. I chose three questions where I could name the key papers in advance — the creatine question above, a workplace-productivity question and a clinical question about a common supplement. In each case Consensus surfaced the studies I expected in the top ten results, and the AI one-liners accurately characterised each paper's finding in the seven of ten spot checks I did against abstracts (the three misses were all oversimplified directions rather than wrong papers). For a search tool, that is the essential test and it passed convincingly; the semantic retrieval is meaningfully better than keyword search for questions phrased as questions.

The Consensus Meter itself is the second observation, and it is the feature I have the most complicated relationship with. On broad questions with hundreds of retrieved papers it is genuinely informative — the creatine-memory query returned a plausible lean I could verify against my own reading. But on narrow questions the meter can be built on a handful of studies, and the interface does not foreground the denominator loudly enough for my taste. On one niche occupational-health question the meter showed an 80% "Yes" lean that rested, as far as I could count, on about a dozen retrieved papers, several of which cited each other. Aggregating five related studies as "80% Yes" is technically true and practically misleading. I now always expand the retrieved list and count before trusting the needle.

Third, the agentic features. Deep reviews are the newest and most interesting: they run a multi-step review and come back as a structured brief with citations. I ran three; two were strong (well-organised, citations checked out), and one on a fast-moving topic leaned on older studies when I knew newer ones existed — the retrieval layer is still the binding constraint. The MCP integration is the piece I expect to use most going forward: 30 free calls a month is a genuine trial allowance, and wiring Consensus into an assistant gave my generic chat tool properly cited research answers within minutes.

What it does not do

Consensus only searches published science. Grey literature, market reports, internal data, news — none of it is here, which is the right scope for the product but disqualifies it for several of the questions I get asked in commercial work. It does not do data extraction into tables (Elicit's territory), it does not show you how papers cite each other qualitatively (Scite's territory), and it will not run a formal systematic review with screening passes and inter-rater checks. The Deep review is closer to a good research brief than to a review protocol.

The meter caveat above is the one I would print on the box: consensus among retrieved papers is not consensus among all evidence, and both the corpus skew (journals, English-language, indexed outlets) and the retrieval step shape what the needle shows. Newer papers can take a while to surface; on one query a paper I knew from 2026 appeared only when I searched a distinctive phrase. And on the free tier the agentic features run out almost immediately — 3 Deep reviews a month is a teaser, not a working allowance — so the product you can actually research with is the $20/month Pro at minimum.

Verdict

Consensus is the fastest way I know to get a cited, balanced read on a scientific question without pretending to be a systematic reviewer. The retrieval is strong in my recall tests, the per-study summaries were accurate in seven of ten spot checks, the MCP/API access is real and cheaply testable, and the free tier is generous enough for genuine casual use. Against that: the meter can mislead on thin evidence, Deep reviews inherit retrieval blind spots, and the agentic features are metered tightly below the paid tiers. I rate it 4.1 on the ToolSpotter rubric — core search depth is high and the meter is a real differentiator, value is good at Pro pricing, usability is a strength, the API/MCP ecosystem is maturing quickly, and reliability is solid with the thin-evidence caveat. For students, fact-checkers and working scientists who want "what does the literature say?" answered in minutes with sources attached, it earns a place on the shortlist.

Sources

• https://help.consensus.app/en/articles/10087865-subscription-plans
• https://consensus.app/home
• https://consensus.app/home/pricing
• https://consensus.app/home/resources/how-consensus-works/
• https://consensus.app/api-mcp/
• https://consensus.app/home/blog/category/new-features
• https://consensus.app/home/university/
• https://consensus.app/pricing/
• https://consensus.app/home/contact/
• https://theaiagentindex.com/agents/consensus
• https://thegeeksclub.com/platforms/consensus