Consensus

Consensus is an AI search and evidence-synthesis product built specifically around scholarly literature. Its current search modes range from paper browsing to Pro synthesis across up to 20 papers and Deep reviews across up to 50 selected papers, with Research Agent planning multi-step searches and applying filters. The searchable corpus is now described as more than 220 million research papers, while Medical Mode narrows discovery to a curated clinical subset and guidelines. ¹

The product emphasizes traceability rather than answer-only chat: generated claims link back to papers, Study Snapshots expose structured study details, and the Consensus Meter summarizes directional findings when enough relevant papers are available. For programmatic workflows, a public API supports academic search with filters for date, study type, sample size, citations, journal, field and country, while MCP brings Consensus search into compatible clients such as Claude and ChatGPT. ²³⁴

That breadth does not make the system a reproducible systematic-search engine by itself. Deep review selects a bounded paper set for synthesis, and current comparative research finds substantial performance differences among AI literature-search systems depending on domain and task. Formal evidence reviews should therefore preserve database-specific searches, inclusion criteria and human verification alongside AI-assisted discovery. ⁵

The free tier is substantial: $0 with unlimited Paper searches, limited Pro messages, three Deep reviews per month, ten Study Snapshots and 30 shared API/MCP calls. Pro costs $20 monthly or $144 yearly and adds unlimited Pro messages, 15 Deep reviews, unlimited Study Snapshots and 250 API/MCP calls. Deep costs $65 monthly or $540 yearly and raises Deep reviews to 200 per month plus 1,000 API/MCP calls. Teams and Enterprise use custom pricing. ⁶

Paid users can enable API/MCP overage at $0.10 per call within plan-specific maximums; Pro can reach 1,000 calls per month and Deep 10,000. The free plan is strong for ordinary discovery, but frequent Deep-review or programmatic use moves quickly into paid tiers. ³

Independent institutional adoption is meaningful. Yale Library began a year-long Consensus trial in March 2026 for scientific and medical discovery, and Yale classifies the product for low-risk data use. The University of Geneva library likewise describes it as a rapid literature-analysis and synthesis tool for students and researchers. These are adoption and workflow signals, not controlled proof of answer accuracy. ⁷⁸

A July 2026 European Heart Journal - Digital Health study compared Consensus with ChatGPT-5, Elicit, Scite and PubMed across four cardiology topics using blinded judgments from three electrophysiologists. Performance varied materially by tool; ChatGPT-5 led the tested setting, and the authors concluded that AI retrieval should complement rather than replace transparent database searches. The study is domain-specific and does not establish a universal ranking, but it provides current independent evidence that literature-search quality remains heterogeneous. ⁵

A 2025 peer-reviewed review of Consensus found that the product can make academic discovery and organization easier, while highlighting risks of oversimplifying research findings, limited reporting of its use in published work and broader concerns about accuracy, transparency and reproducibility. That evidence predates several 2026 product upgrades, so it is best read as a continuing verification caution rather than a measurement of the present release. ⁹

Current privacy documentation is more explicit than older summaries. Consensus says user content is not used to train its or third-party models, personal identifiers are not attached to content sent to model providers, data is encrypted in transit and at rest, and processing providers include OpenAI, Anthropic, Google and Baseten under service agreements. Its servers are located in the United States. The help center separately says query logging can be anonymized or disabled under custom organizational security arrangements. ¹⁰¹¹

The fixed panel adds limited user-sentiment evidence. G2 currently has an exact product page but zero reviews, so its empty-state display supplies no numeric rating evidence. Trustpilot has six reviews and a 2.4/5 aggregate, with complaints concentrated around subscriptions, billing and support; the sample is too small and self-selected for product-quality inference and Trustpilot is qualitative-only under the rating rules. ¹²¹³

Consensus fits students, academics, clinicians, analysts and research teams that want cited answers and rapid orientation across scientific literature without starting from a conventional keyword-only database interface. Pro and Deep are particularly useful for scoping questions, reading lists, evidence maps and multi-paper synthesis, while API/MCP access makes the product relevant to research applications and agent workflows.

It is less suitable as the sole retrieval method for publication-grade systematic reviews, for questions that depend heavily on non-scholarly web sources, or for organizations that require high-risk-data approval without additional contractual controls.