Consensus

What Consensus is and what it returns

Consensus is a scientific research service from Consensus NLP, Inc. It lets users start with a question, find publications and obtain answers linked to the sources used. Its working material is academic literature: useful output combines references, explanations and comparable study characteristics, alongside conversation with the assistant.

For a professional entering an unfamiliar subject, it helps turn a broad concern into an initial reading list. For a specialist, it can support exploration of disagreements or a research trajectory. In either case, three products should remain distinct: the retrieved documents, the generated interpretation and the conclusion the researcher ultimately writes. An authentic reference addresses only the first of these levels.

Reference documentation: Consensus FAQs · Privacy Policy — November 1st, 2026.

Origins and product development

Christian Salem and Eric Olson founded Consensus in 2021; the company now identifies San Francisco as its home. Its search engine launched in 2022. In July 2024 it announced an $11.5 million Series A led by Union Square Ventures. These are business milestones, not measures of scientific reliability.

The Meter appeared in February 2023. Developments during 2025 included Meter 2.0, Threads for continuing conversations, Deep Search and analysis of full text when available. In October that year, OpenAI described Scholar Agent as a multi-agent system using GPT-5 and the Responses API. That case documents an architecture at a particular time, not a permanent model roster.

In May 2026 the documentation presented Research Agent as the successor to Scholar Agent. October’s Skills added prepared research workflows. This trajectory explains the shift from finding and summarizing papers towards maintaining reading projects and performing connected tasks.

Reference documentation: About Consensus · Announcing Our $11.5M Series A Fundraise · Consensus Meter: Guardrails and Limitations · Consensus Product Changelog · Consensus uses GPT-5 and the Responses API · Research Agent · How to use Skills.

How it retrieves and ranks papers

The provider’s technical explanation describes hybrid search combining semantic similarity with BM25 keyword matching. It then reranks results using relevance and bibliographic signals. The corpus combines aggregators such as Semantic Scholar and OpenAlex, its own crawling and publisher agreements.

This design creates a practical choice: a very broad question may retrieve conceptually related work with little applicability to the problem. An overly narrow question may miss alternative terminology. A useful strategy is to retain an initial formulation, explore synonyms and document expansions. Asking for “the best studies” is insufficient: specify which outcome, population and comparison matter.

Coverage figures published on different provider pages do not always agree. The advertised size of the corpus is therefore not used here as a guarantee of completeness or access to every full text.

Reference documentation: How it Works & Consensus FAQ's.

Filters and relevance criteria

Filters include date, design, human or animal population, sample size, duration, country, journal and open access; preprints can also be excluded. Bibliographic labels help narrow results, but an influential journal or numerous citations do not independently validate a finding.

In a workplace search, filtering should follow the protocol rather than the desired answer. Excluding small studies may remove noise, but also the only available work on a very specific activity. Restricting results to randomized trials may be reasonable for an intervention and unhelpful for exposures that cannot be experimentally assigned.

Before combining restrictions, review which documents disappear and why. This makes the selection explainable to another person and helps reveal whether apparent agreement actually results from narrowing the query too far.

Reference documentation: Advanced Search Filters.

Interpreting the Consensus Meter

The Meter classifies between five and twenty relevant results into positions such as yes, no, possibly or mixed. Its detailed view shows age, method, journal and citation indicators by position. It represents the retrieved set, not a vote by the entire scientific community.

A high percentage is not the probability that a measure will work in a company. It does not show how much an outcome changes, for how long or with what uncertainty. Two favorable studies may measure different things or draw on the same cohort. Useful reading means opening each position and asking which population, comparison and outcome underpin it.

If the indicator does not appear, that should not be interpreted as negative evidence. The question may not be binary or there may be too few classifiable results. Rephrasing merely to obtain an attractive bar has no value if it changes the question being answered.

Reference documentation: The Consensus Meter.

Study Snapshot and reading an individual paper

Study Snapshot extracts details such as population, sample, duration, location, methods, variables and findings from the abstract. Fields may remain empty when the abstract supplies insufficient information. Full-text chat supports work on selected papers or uploaded documents in the enabled modes.

The snapshot serves as an initial reading record. A missing field should remain unknown until the original is checked; it should not be filled by intuition or analogy with another study. In particular, establish whether a sample refers to participants recruited, analyzed or followed to completion, and whether duration covers the intervention or the entire follow-up.

When requesting a figure, also ask for its context: group, unit, measurement time and supporting passage. This helps prevent a correct estimate being transferred to a table but attributed to the wrong group.

Reference documentation: Using the Study Snapshot · How to Chat With Full Text.

Research Agent, Pro and Deep

Research Agent can chain searches, trace authors and citations, retrieve by DOI and compare studies. Deep breaks questions down and typically analyzes up to fifty papers, potentially reaching one hundred when useful. These are documented capabilities, not a commitment to find all relevant evidence.

Pro suits a bounded exploration; Deep allows more retrieval and synthesis work on a complex question. Before choosing, define the output: a candidate list, a comparison of methods or a map of claims. Requesting a long report without selection criteria can produce a convincing explanation of an incomplete set.

Chained tasks also require checking intermediate steps. If an initial search confuses muscle fatigue with sleepiness, a more elaborate subsequent synthesis may carry that error forward. Final quality depends on framing and checking the problem correctly.

Reference documentation: Research Agent · How to Use Deep review.

Skills and checking claims

Skills provide workflows for reviewing a topic, following references, finding evidence for a claim and checking citations in a draft. They run with Deep. Verify Citations connects a claim with a passage and assesses whether the source supports, contradicts or does not address it.

A reasoned editorial application is to check a sentence before publication: separate its population, intervention and outcome, locate the passage and confirm that its scope has not been expanded. For example, a claim about knowledge retention should not become a claim about fewer accidents.

Reverse reference searching can also introduce bias if it only seeks support for a conclusion already chosen. Request favorable and contradictory evidence, together with conditions under which the claim would cease to hold. The purpose is to correct the argument, not decorate it with citations.

Reference documentation: How to use Skills.

Collections, exports and integrations

My Library organizes papers, searches and documents in collections, with reference imports and viewing or editing permissions. Exports allow references to be reused through CSV, RIS or bibliographies. MCP connects Consensus to compatible assistants, subject to authorization and, in some environments, administrator involvement.

A team should distinguish its candidate library from its reviewed-document library. Saving a paper does not mean approving it. A simple convention can record who checked the reference, which question it addresses and where the reading note is stored. Reviewing duplicates, DOIs and versions on export avoids mixing a preprint with its later publication.

When consulting from another assistant, retain the link to the original scientific search. A second reformulation may lose nuances from the first answer. Traceability should reach the paper and passage, rather than stopping at the connected system’s conversation.

Reference documentation: My Library – How to Save Papers and Searches · How to Connect Consensus to Other AI Tools · How to Export Consensus Results to Reference Managers.

Access, languages and choosing a plan

Consensus offers free access and paid plans, including institutional options. Deep quotas depend on the plan; “unlimited Pro” should not be interpreted as unlimited use of every feature. Settings allow users to select the language of answers and summaries.

Selection should be based on a sample of real questions and expected monthly volume, including repetitions and verification. An institution can provide access to the tool without granting universal access to all journals. Using the search engine, reading a publication and reusing its content are separate permissions.

When working in Spanish or Catalan, retain original scientific terms when checking references. Fluent translation can conceal differences between related constructs. This profile is available in three languages; that is not evidence of equivalent product performance across those languages.

Reference documentation: How to Use Deep review · Oklahoma State University: Consensus AI-powered Academic Search Engine · Consensus FAQs.

Reception and documented limitations

Oklahoma State University Libraries document institutional access to Consensus and position it as an aid to exploration. They warn about possible misinterpretation, incomplete coverage and full-text restrictions. This demonstrates institutional use and library guidance, not a trial measuring overall accuracy.

Three evaluation questions should remain distinct: whether it finds relevant studies, interprets them correctly and preserves disagreements in the synthesis. A tool may perform well on the second task with supplied articles while struggling to discover them in the first. Comparing answers without controlling the document set obscures where differences arise.

To assess local usefulness, a team could prepare a small set of questions with known bibliographies, record omitted documents and check important claims. This is an evaluation proposal from Sabentis; it was not performed for this profile and cannot establish an expected accuracy percentage.

Reference documentation: Oklahoma State University: Consensus AI-powered Academic Search Engine.

Data, documents and privacy

The policy dated November 1, 2026 states that queries, conversations and documents may be processed by model providers, while excluding their use for training. This does not mean content never leaves Consensus. The policy addresses account data, retention and deletion options.

Prevention research usually requires only a general question and public literature. Entering health histories, names or identifiable incident details adds information processing unnecessary for finding studies. If authorized internal documents are used, the decision should consider what is shared, who can access the collection and which connected service receives the content.

Reviewing attachments before sharing a project helps prevent a collection of articles accidentally including an internal report. Workspace configuration and specific contractual conditions matter alongside the no-training commitment.

Reference documentation: Privacy Policy — November 1st, 2026.

Relevance to occupational safety and health

The following uses are proposals from Sabentis. Consensus can help explore research on preventive measures, work organization and risk communication. Its contribution concerns preparing evidence; the tool does not independently know workplace conditions or demonstrate the effectiveness of a local intervention.

• Organization and fatigue: distinguishing sleep duration, sleepiness, perceived tiredness and errors prevents different outcomes being presented as equivalent.
• Safety training: searching separately for knowledge, observed behavior and safety outcomes helps establish what was measured and over what period.
• Participation: finding studies on worker consultation can suggest questions for project design, distinguishing satisfaction with participation from verified changes in exposure.

A useful report explains which evidence directly concerns the population of interest and which provides context only. It also retains conflicting findings and conditions of application. If the literature concerns a different sector, schedule or task, that distance should remain visible before proposing an organizational evaluation.

Reference documentation: Research Agent · Advanced Search Filters.

Hypothetical case: researching fatigue in shift work

In this hypothetical case proposed by Sabentis, the question is whether particular organizational interventions are associated with less fatigue among shift workers. The practitioner uses the results to locate articles, separates observational studies from interventions and prepares a synthesis that explains the limitations of the evidence.

1. Define population, intervention and outcome: rotating shifts, organizational changes and fatigue measured using identifiable instruments.
2. Request searches with synonyms and retain the query, filters and date.
3. Ask for a table covering design, sample, duration, comparator and outcomes, including studies with different findings.
4. Open the papers and check that the table reflects the specific outcome; better sleep alone does not demonstrate fewer accidents.
5. Document disagreements and gaps before proposing a local evaluation.

The useful deliverable is a traceable evidence matrix with checked references and explicit uncertainties, rather than an isolated support percentage.

To make the comparison reproducible, record exclusion criteria too. A study of tiredness in students may explain a mechanism without directly answering a question about night workers. Separating direct from contextual evidence avoids treating them equally. If the assistant proposes a broad conclusion, ask which particular papers support it and which outcome each measured.

Before completing the report, check whether the studies use compatible scales and periods. A statistical difference does not automatically mean a meaningful workplace improvement; retain magnitudes and uncertainty where available in the original.

Reference documentation: Research Agent · Advanced Search Filters · Using the Study Snapshot.

What the tool cannot establish

Neither the Meter nor an extensive report replaces a systematic review with a protocol, reproducible searches and explicit risk-of-bias assessment. Nor does a paper’s absence from results establish that it does not exist. Scope depends on the query, filters, indexing and available access.

Three errors deserve attention in an OSH synthesis: counting several publications from the same sample as independent confirmations; turning associations into causal effects; and transferring findings between populations without examining their differences. These interpretation problems can persist even when every DOI is correct.

The conclusion should distinguish what was found, what remains unchecked and what needs further investigation. A nuanced explanation may be less forceful than an automated answer, but it helps establish which information is missing before action.

Reference documentation: The Consensus Meter · Oklahoma State University: Consensus AI-powered Academic Search Engine.

When to choose it and what can complement it

Choose Consensus when the main requirement is exploring a scientific question and following connections between papers. Elicit can complement planned extraction into columns, while Scite helps examine how a study is cited. Comparing outputs means examining the question and underlying corpus, rather than taking a vote between assistant responses.

Reference documentation: Research Agent.

Frequently asked questions

Does the Meter establish scientific consensus?

No. It summarizes classifications within a set of results; study quality, independence and relevance still require assessment.

Can it find a known paper?

Research Agent documents DOI and author searches. Check the title, year and publication to avoid confusing different versions.

Can it directly justify a preventive measure?

It can help locate evidence. The decision requires considering the workplace context, alternatives and the quality of that evidence.

Reference documentation: Research Agent · The Consensus Meter.