Consensus Review: Academic Search, Deep Review and Research Agent

Consensus is expanding the material its research agent can actually read, not only the abstracts it can index. On August 5, 2026, the company announced a full-text partnership with De Gruyter Brill, adding licensed journal content with particular relevance to humanities and social-science research. The agreement is the latest in a series of publisher relationships that also includes Wiley, Taylor & Francis, Sage, the American Chemical Society, the American Psychological Association, and the American Association for the Advancement of Science.

That change matters because a research assistant that can inspect methods, results, discussion, and supporting context has more evidence to work with than one limited to titles and abstracts. It also gives Consensus a better opportunity to serve fields where the argument may unfold across a long article and where a short abstract cannot represent the author's evidence, framing, or disagreement with prior scholarship.

The partnership should not be read as universal access or automatic accuracy. Consensus can use licensed text for search and analysis while a reader may still need a university login or personal subscription to open the original article. Full-text coverage varies by publisher and publication, and the announcement focuses on journals rather than promising comprehensive access to every book, monograph, archive, or historical source associated with De Gruyter Brill.

### Full text changes retrieval before it changes writing

The most important benefit of full-text access happens before a polished answer appears. Search systems have more material for matching a question to relevant work. A concept that is absent from a title or abstract may be discussed in detail inside a methods section, footnote, literature review, or conclusion. An intervention may look effective in the abstract while the paper's limitations narrow that conclusion to one population, measurement choice, or short follow-up period.

Consensus says its search combines semantic retrieval with keyword methods, then reranks candidate papers using relevance and quality signals. Pro answers can synthesize up to 20 papers. Deep Review uses multiple targeted searches, examines more than 1,000 candidates, and can select up to about 50 papers for a structured report. When full text is available, those stages can draw on more of each paper.

More available text does not guarantee better interpretation. A model can still connect the wrong passage to a claim, confuse an author's summary of another study with the author's own finding, miss a qualification, or mistake correlation for causation. Researchers should use the full-text indicator and source snippets as verification tools, then open the original paper and inspect the surrounding section. The right question is not simply whether an answer has citations. It is whether each citation supports the exact claim made, in the relevant population and context.

### Humanities research is not a tidy consensus problem

De Gruyter Brill is particularly strong in areas such as history, philosophy, theology, classics, linguistics, law, and social sciences. These disciplines can benefit from better discovery, but they also expose the limits of a product name built around consensus. Many humanities questions are not properly answered by counting papers as positive or negative. Scholars may disagree because they use different archives, interpret the same source through different frameworks, define a period differently, or ask different normative questions.

A useful AI workflow in these fields should map arguments rather than force them into a vote. It can identify schools of thought, trace how a concept changes over time, compare evidence bases, surface counterarguments, and build a reading list. It should distinguish a historian's interpretation from an empirical claim and preserve who is arguing what. A confident paragraph that blends incompatible frameworks is less useful than a table that identifies authors, periods, sources, claims, and contested assumptions.

The Consensus Meter is especially easy to misuse here. The feature analyzes 5 to 20 highly relevant papers for eligible directional questions and classifies their findings as Yes, No, Possibly, or Mixed. Consensus itself warns that the sample is not the complete literature, that the model can misclassify a paper, and that question nuance can be lost. Those limitations apply in any field and become more pronounced when the scholarship is interpretive rather than experimental.

### Publisher access is not reader access

The practical distinction between analysis rights and reader rights deserves attention. Consensus says publisher full text can be used to improve retrieval and answer generation, and its partnership announcement says results direct users back to the original work. It also says the licensed content is not used to train AI models. Those are useful commitments, but they do not remove a publisher paywall.

A user may receive a summary influenced by an article that they cannot independently download. For casual orientation, that may still be valuable. For academic writing, peer review, clinical decisions, policy work, or any claim likely to be challenged, the inability to inspect the source is a material limitation. Researchers should mark inaccessible sources in the evidence table, use institutional access or interlibrary services where available, and avoid quoting or making decisive claims from material they cannot verify.

Coverage should also be recorded, not assumed. Consensus describes a database of more than 220 million papers drawing from Semantic Scholar, OpenAlex, its own crawl, publisher agreements, and open-access documents. The broader corpus can include preprints and conference proceedings. A large number is not the same as comprehensive indexing, and it does not ensure equal representation across languages, regions, disciplines, or older collections. Preprint exclusion, publisher filters, study-design filters, and exact search logs are essential when publication status affects the conclusion.

### Deep Review still needs a documented method

Deep Review and the Research Agent can reduce the mechanical burden of an early literature review. They can expand terminology, run several searches, screen a larger candidate pool, compare papers, produce a structured report, and export citations or evidence tables. The June 2026 update also brought Deep Search into My Library and collections, with custom outputs such as PICO or SPIDER tables, reading lists, and scoping maps.

Those capabilities are best treated as research infrastructure, not a completed method. A systematic review needs a protocol, appropriate database selection, reproducible searches, explicit inclusion and exclusion criteria, duplicate screening where required, critical appraisal, and a transparent account of missing and excluded evidence. Consensus can assist several of those steps, but a final report should not obscure which steps the user performed, which steps the agent performed, and where judgment changed the result.

A defensible workflow starts with Paper Search and a written question. The researcher records the exact query, date, filters, and result count, examines whether relevant terms and known papers appear, and revises the strategy. Only then should Pro or Deep generate a synthesis. The resulting table should include design, population, setting, intervention or exposure, comparator, outcome definition, follow-up, uncertainty, funding, and limitations. Blank cells are better than invented details.

The search should then be repeated in discipline-appropriate databases. Citation chasing, grey-literature searches, reference-list review, and expert consultation may still be necessary. Consensus is especially useful for finding vocabulary and candidate papers, but no academic search product should silently become the only source of recall for a high-stakes or publication-grade review.

### What researchers should verify before citing

First, check publication status. Do not label a preprint, conference item, or manuscript as peer reviewed without confirmation. Second, check whether Consensus analyzed full text or only an abstract and metadata. Third, verify that the cited passage supports the claim and that the paper studied the relevant population, exposure, comparator, and outcome. Fourth, inspect design quality and uncertainty rather than using citation count or journal reputation as a substitute for appraisal. Fifth, record inaccessible full-text sources and do not rely on quotations you cannot inspect.

The same discipline applies to uploaded and imported material. Consensus can accept up to 20 PDFs at a time, but scanned or image-only documents may fail. Zotero import is one way rather than continuous synchronization, and not every imported item type is currently included in My Library search. Researchers should confirm that the intended documents are present, searchable, current, and in the correct collection before asking cross-document questions.

Privacy also belongs in the method. Consensus says uploads are private unless shared through a collection and that user queries, content, and personal data are not used to train its own or third-party large language models. Its policy describes encryption and says it does not sell personal information. However, it uses cloud service providers and retains data for as long as reasonably necessary for stated purposes rather than promising one fixed period for every category. Unpublished manuscripts, confidential reports, participant data, patient information, and contract-restricted content still need an institutional data-handling decision before upload.

The De Gruyter Brill agreement is a meaningful expansion because it improves the chance that Consensus can retrieve and inspect the substance of scholarship outside the heavily represented biomedical and technical literature. It makes the platform more useful for discovery across humanities and social sciences. The responsible conclusion is narrower than the marketing headline: more licensed full text can improve search and evidence extraction, but it does not make coverage universal, remove paywalls, turn interpretation into consensus, or eliminate the obligation to read and evaluate the source.

WhatAI Decision Box

Consensus searches more than 220 million scholarly papers and returns cited answers, paper lists, evidence tables, study snapshots, and structured research reports. Its strongest workflows combine question refinement, targeted filters, full-text analysis where available, and direct inspection of the cited sources.

Consensus combines semantic and keyword retrieval, quality-aware reranking, and AI analysis. Paper Search offers a browse-first result list, Pro summarizes as many as 20 papers, and Deep Review examines more than 1,000 candidates across multiple searches before selecting up to about 50 papers for a report. The Research Agent can plan steps, apply filters, inspect papers, and generate a cited response.

Consensus fits students, researchers, clinicians, analysts, journalists, and evidence-focused teams that need faster discovery and synthesis. It should support, not replace, formal database searches, systematic-review protocols, critical appraisal, statistical analysis, professional judgment, or reading the primary paper. Verify full-text status, publication type, methods, population, and quoted evidence before relying on an answer.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-16.

Consensus.app Community Discussions

Explore community discussions. Ask and answer questions on Consensus.app to grow and learn together.

Consensus.app Pros & Cons

How to Get Results with Consensus.app: Step-by-Step Workflow

1. Define the decision

   State the question, intended use, population, setting, outcomes, timeframe, acceptable evidence, and decisions the research may inform. Separate exploratory work from a formal review.

2. Run Paper Search

   Start with natural language, inspect terminology in relevant papers, then build keyword and Boolean variants. Record every query, date, result count, and filter used.

3. Apply evidence filters

   Choose dates, study designs, populations, publishers, citation thresholds, open-access status, human or animal settings, and preprint rules that match the question.

4. Review retrieval quality

   Inspect the first results for relevance, missing synonyms, unsuitable populations, duplicate records, and overrepresented sources. Revise the search before generating a synthesis.

5. Choose Pro or Deep

   Use Pro for focused orientation and Deep Review for a wider, structured investigation. Review the Research Agent plan and adjust the intended outputs before it runs.

6. Build an evidence table

   Extract design, sample, setting, intervention or exposure, comparator, outcome, effect, uncertainty, follow-up, funding, and limitations. Leave unsupported fields blank.

7. Verify every key claim

   Open the cited paper, confirm whether full text was analyzed, locate the supporting passage, and check that the summary matches the population, method, outcome, and direction of effect.

8. Search beyond Consensus

   Repeat the documented strategy in appropriate discipline databases and publisher indexes. Add citation chasing, reference-list review, grey literature, and expert consultation where required.

9. Appraise and synthesize

   Assess risk of bias, directness, consistency, precision, and publication status. Separate findings from interpretation and explain conflicts instead of averaging them away.

10. Export an audit trail

    Save the report, evidence table, citations, exact queries, filters, search date, inclusion decisions, exclusions, corrections, and unresolved uncertainties with the final research output.

Consensus.app Gotchas and Limits to Know Before You Start

• The database is broad, but not every indexed item is peer reviewed; preprints and conference proceedings can appear.
• More than 220 million indexed papers does not mean every publisher, discipline, language, or historical archive is covered equally.
• Search ranking favors relevance and quality signals, which can still miss niche, older, negative, or differently worded evidence.
• Pro answers synthesize up to 20 papers rather than the entire literature on a question.
• Deep Review examines many candidates but selects up to about 50 papers for synthesis, so its inclusion logic still needs review.
• The Consensus Meter requires at least 5 relevant results and analyzes no more than 20.
• Yes, No, Possibly, and Mixed labels can flatten dose, population, comparator, timeframe, and outcome differences.
• The model can misclassify a study's findings or overstate what its design supports.
• A paper marked as full text is easier to verify, but the generated interpretation can still be wrong.
• Licensed publisher text may be searchable by Consensus even when the user cannot open the full article without a subscription.
• Full-text partnerships do not imply that every journal, issue, article type, book, or monograph from a publisher is covered.
• When full text is unavailable, analysis may rely on an abstract and metadata that omit methods, caveats, and subgroup results.
• Medical Mode is a focused research interface, not medical advice, diagnosis, prescribing guidance, or a patient-specific decision system.
• Clinical guidelines can become outdated and recommendations can differ by country, specialty, and patient population.
• PDF upload accepts up to 20 documents at a time and individual files can be limited to 50 MB.
• Scanned or image-only PDFs may fail because searchable text is required.
• Zotero import is one way, not an automatic two-way sync, and updates require another manual import.
• Books, book sections, conference papers, manuscripts, reports, theses, and some other Zotero items may be stored but not included in My Library searches.
• An uploaded document is private by default but can become accessible to others if placed in a shared collection.
• Incognito mode prevents a search from appearing in history, but it is not a guarantee that no service data is processed or retained.
• The privacy policy describes purpose-based retention rather than one fixed deletion period for every data category.
• Vendor statements about retrieval accuracy, privacy, and coverage should be evaluated as vendor claims unless independently audited.
• Citation exports can contain incomplete or incorrect metadata and should be checked before publication.
• AI-generated prose can conceal conflicting evidence, low certainty, indirect populations, or weak study designs behind fluent language.
• Research Agent and Deep Review outputs are research starting points, not completed systematic reviews or meta-analyses.
• Students and authors should follow institutional disclosure, authorship, plagiarism, and AI-use policies.
• Sensitive unpublished manuscripts, patient information, participant data, and confidential reports need an approved data-handling decision before upload.
• Discounts for students, faculty, and clinicians require eligibility verification and can change.

Which Consensus.app Feature Fits Your Use Case

Starter Prompts for Consensus.app

What is Consensus.app?

Consensus is an AI academic search and research synthesis platform. It searches more than 220 million papers, ranks relevant evidence, and can generate cited answers, study summaries, evidence tables, and structured reviews.

Is every paper in Consensus peer reviewed?

No. The broader database includes scholarly journal content but can also include preprints and conference proceedings. Use publication-type and preprint filters, then verify the status of each source before presenting it as peer reviewed.

How much does Consensus cost?

As verified on August 16, 2026, Free costs $0, Pro costs $20 monthly or $144 yearly, and Deep costs $65 monthly or $540 yearly. Teams and Enterprise use custom pricing.

What is included in the Free plan?

The Free plan includes unlimited Paper Searches, a limited monthly allocation of Pro messages, 3 Deep Reviews per month, and 10 Study Snapshots per month. Limits and feature names can change, so check the plan page before purchase.

What is the difference between Pro and Deep Review?

A Pro answer synthesizes a focused set of up to 20 relevant papers. Deep Review runs multiple targeted searches, examines more than 1,000 candidates, selects up to about 50 papers, and creates a more structured report with methods, results, discussion, and export options.

What does the Consensus Research Agent do?

The Research Agent plans and executes multi-step research tasks. It can refine a question, chain searches, apply filters, read across results, and produce a cited answer. Its plan and evidence still require user review.

How does the Consensus Meter work?

For eligible directional questions, the Meter classifies findings from 5 to 20 highly relevant papers as Yes, No, Possibly, or Mixed. It is a limited snapshot, not a vote across all research, and the model can misclassify a paper.

Does Consensus read full papers?

Sometimes. It can analyze full text from open-access sources, supported publisher agreements, and user-uploaded PDFs. When full text is unavailable, it may rely on an abstract and metadata. Look for the full-text indicator and inspect the cited passage.

Does licensed full text remove a publisher paywall?

No. Consensus may be permitted to search and analyze licensed text while the user still needs a personal or institutional subscription to open or download the article from the publisher.

What is Medical Mode?

Medical Mode narrows search toward roughly 8 million papers and guidelines from selected clinical sources, including about 50,000 guidelines. It improves focus but does not provide medical advice or replace current guidelines and clinician judgment.

Can Consensus replace a systematic review?

No. It can accelerate discovery, screening, evidence mapping, and drafting, but a systematic review needs a registered or documented protocol, reproducible searches across appropriate databases, independent screening, critical appraisal, and transparent synthesis.

Can I upload private PDFs?

Yes. Users can upload as many as 20 PDFs at a time, with a 50 MB file limit, and ask cited questions across them. Scanned or image-only documents may not work. Uploaded documents are private unless placed in a shared collection.

Does Consensus sync automatically with Zotero?

No. Zotero import is one way and updates require another manual import. It can preserve folders, but current search and analysis support within My Library is centered on journal articles rather than every Zotero item type.

Does Consensus use my searches to train AI models?

The current privacy policy says queries, content, and personal data are not used to train Consensus or third-party large language models. This is a vendor policy commitment, and the policy does not promise a single fixed retention period for all data.

Are Consensus answers always accurate?

No. Citations make verification easier, but retrieval can miss evidence and AI can misread methods, flatten nuance, choose an unsuitable population, or classify a paper incorrectly. Check the source, passage, study design, and applicability.

Related Education Tools

Explore the Network