Scite: Smart Citations and Evidence-Aware Research

A modest product update can reveal where a company believes the future is going. Scite's July 2026 release did exactly that.

The platform combined its Feed and Alerts into a single Research Feed, bringing followed papers, researchers, and topics into one stream. It also introduced an MCP User Dashboard that shows signed-in users which Scite tools their connected AI applications are calling and how often. Neither feature sounds as dramatic as a new language model. Together, they point to something more useful: Scite is trying to become part of the researcher's normal environment rather than another tab visited only when a citation looks suspicious.

That direction has been building throughout 2026. Scite MCP can bring Smart Citations and full-text literature search into ChatGPT, Claude, Gemini, Copilot, Cursor, Claude Code, and other compatible tools. Collections can now be created and managed from those AI clients. A Claude connector arrived in April. Collections replaced the older Dashboard workflow in May. By July, Scite was giving users a unified way to follow research and see how their connected assistants were using the service.

The result is a more interesting product than the phrase citation checker suggests. Scite now sits somewhere between a scholarly database, an evidence monitor, an AI research assistant, and an infrastructure layer for other AI tools.

The feed is the real clue

Traditional literature searching has a clear beginning and an unreliable end. A researcher writes a query, screens the results, saves useful papers, and promises to return later. The field keeps moving. New studies appear, older findings are challenged, papers are corrected or retracted, and a source that once looked central can become less convincing.

Scite's unified Research Feed is designed around that reality. A followed topic, researcher, or paper can continue producing relevant activity after the first search is over. Collections can alert users to new supporting or contrasting citations and flag retractions. The research object is no longer a static folder. It becomes something that can change as the literature changes.

This is especially valuable for long projects. A doctoral candidate may work on one claim for years. A medical-affairs team may need to watch evidence around a therapy between formal reviews. A product-safety group may need to notice when an assumption receives credible challenge. A literature review prepared six months ago can look polished while already missing the paper that changes its conclusion.

The danger is alert fatigue. Following everything produces a feed that no one can meaningfully assess. A useful setup begins with decisions, not topics. Follow the papers, authors, and claims that could alter a protocol, conclusion, risk assessment, research direction, or investment. Give each Collection an owner and a review cadence. If an alert cannot lead to an action, it may not deserve attention.

Scite is leaving its own tab

Scite MCP is the other half of the shift. Instead of asking a researcher to move from ChatGPT or Claude into Scite, the connection brings Scite's search and citation intelligence into the AI tool already being used.

This matters because general AI assistants are convenient but uneven research partners. They can produce fluent summaries, yet scholarly coverage may be incomplete and references can be wrong, invented, or detached from the claim they are meant to support. Scite's value is not simply that it returns more citations. It can expose citation statements, show how later papers referred to earlier work, search scholarly full text where available, and connect a response to verifiable records.

The MCP page describes support for ChatGPT, Claude, Gemini, Copilot, and other compatible clients. Research Solutions also names Cursor and Claude Code, which makes the connection relevant beyond essay writing. A developer building a scientific pipeline can query literature without leaving the coding environment. An analyst can investigate a technical assumption while working through data. An agent can create or update a Collection as part of a broader review workflow.

The new MCP User Dashboard is a sensible governance step. If AI tools are calling search, citation, Collection, or paper-access functions on a user's behalf, people need to know what is happening. Usage visibility does not solve every oversight problem, but it gives researchers and organisations a place to start asking whether the connection is being used as intended.

The human still needs to inspect the result. A connector can improve retrieval and citation grounding without making an AI answer correct. The question may be poorly framed. The search may miss a synonym. The evidence may be heterogeneous. A cited paper may not support the sentence in the way the assistant implies. Better sources make verification easier; they do not make verification optional.

Why Smart Citations still matter

Scite's defining feature remains Smart Citations. A conventional citation count tells you how often later work referred to a paper. It does not tell you why. Scite analyses citation statements and classifies them as supporting, contrasting, or mentioning the cited work.

That extra context can transform a literature trail. A highly cited paper may be used mainly as background. A modestly cited experiment may have several direct replications. A famous claim may have attracted a small number of technically important challenges that disappear inside a total count. Reading the actual citation statement helps a researcher reach the relevant part of the later paper without guessing from the title.

Scite says it searches more than 300 million scholarly articles and has indexed more than 1.6 billion citations. The size matters because citation intelligence depends on seeing both the cited paper and the surrounding text in later work. Publisher relationships and open-access sources give the system more context than metadata alone.

The feature is particularly useful when a researcher already has a paper or claim worth examining. Search can find related literature, but Smart Citations can show what happened after publication. That makes Scite a strong second step after discovery and a useful challenge to the habit of treating citation volume as approval.

It can also help with language. Academic disagreement is rarely announced with the words this study is wrong. A later paper may narrow a result, identify a boundary condition, fail to reproduce an effect, criticise a method, or reach a different conclusion in another population. Citation context gives the reviewer a place to investigate those differences.

Citation context is not a verdict

The three labels are useful precisely because they are simple. They are also easy to misuse.

A supporting citation does not prove that the original paper is methodologically strong. A contrasting citation does not prove it is false. A mentioning citation may contain important criticism that does not fit neatly into a classification. The model is interpreting a local citation statement, not conducting a full risk-of-bias assessment across both studies.

Research quality depends on questions the label cannot answer alone. Was the study designed well? Was the sample appropriate? Were outcomes selected after the fact? Is the effect clinically meaningful or merely statistically significant? Has the analysis been corrected? Do conflicts of interest matter? Does the later paper truly test the same claim under comparable conditions?

Coverage also shapes the picture. A new preprint may have no later citations because time has not passed. A niche field may have less full-text coverage. A paper can be influential in practice without being widely cited. An absent contrasting citation may mean consensus, or it may mean that the relevant criticism is not indexed, not expressed through a formal citation, or too recent to appear.

The right interpretation is modest: Smart Citations are prioritisation signals. They help a researcher decide what to read and where to look. The citation statement provides evidence about how one paper referred to another. The scientific judgement still belongs to the person who reads the methods, results, limitations, corrections, and wider literature.

Collections become living reviews

Collections, introduced as the successor to Dashboards, give Scite's evidence signals a place to accumulate. A user can group papers into a curated set, monitor changes, run focused work, and keep a topic organised over time. The current pricing page says Collections can hold up to 10,000 papers.

The more important change is that Collections now work through MCP. A researcher can ask an AI tool to find papers, add selected records to a Collection, update that set, and return to it without constantly moving between interfaces. That is a natural fit for literature review, research surveillance, competitive intelligence, and evidence monitoring.

A Collection should not be mistaken for a completed review. It does not automatically record why papers were included, who screened them, which version was assessed, how duplicates were handled, or whether the search was comprehensive. Those details matter in any workflow that needs to be reproducible.

The best use is to give each Collection a purpose. One might hold all studies screened for a defined question. Another might monitor sentinel papers after a review closes. A third might collect methods that could change an experiment. Add a written protocol, inclusion rules, dates, owners, and an export schedule outside the AI conversation. The Collection becomes more valuable when its boundaries are explainable.

Reference Check catches problems before submission

Reference Check lets a user upload a PDF or DOCX and evaluate the cited sources against Scite's data. It can surface retractions, contrasting citations, and other flags linked to the references.

This is a practical final-stage tool. A manuscript may have been written over months while the status of a cited source changed. A reference may be technically correct but weak for the sentence it supports. An influential paper may have attracted serious criticism since it was added. Reference Check gives the author, supervisor, reviewer, or editor a structured list of places that deserve another look.

It is not an automatic integrity certificate. A clean result does not show that every sentence accurately represents its source. It cannot guarantee that the most appropriate literature was selected. It may not recognise every editorial notice or issue. Authors still need to read the cited work, verify quotations and numbers, confirm current status, and make sure the reference supports the precise claim.

The API documentation adds an important commercial detail: API-based Reference Check requires a paid licence and is not included in a standard self-service Pro key. Organisations planning automated manuscript screening should confirm access, limits, retention, privacy, and result handling with Scite rather than assuming the consumer plan covers a production pipeline.

The search and access gap

Scite's Assistant and Search can answer natural-language questions, search full text where available, and reveal the papers behind an answer. The browser extension and Zotero plugin bring citation signals into familiar discovery and reference workflows. The API exposes papers, tallies, citations, search, Collections, Assistant, and other functions for teams building their own systems.

Finding a paper is not the same as having the right to read it. Some content is open access, some is available through an institution, and some remains behind a publisher paywall. Scite MCP can connect to holdings and Research Solutions describes access resolution through services such as LibKey and GetFTR, with Article Galaxy available when a document needs to be acquired.

That integration is useful because literature work often breaks at the handoff between discovery and access. An AI answer cites a source, the user clicks, and the workflow ends at a login page. A resolver can route the researcher to an entitled copy or a legitimate purchase path.

The boundaries remain important. A citation statement is not a substitute for the full paper. Publisher access rights can differ from AI processing rights. Institutions need to review which content is being sent into external AI tools, what the connector can retrieve, how prompts and results are retained, and whether licences permit the intended analysis.

For sensitive R&D, clinical, legal, or unpublished work, the question is not only can Scite find the literature. It is also what does the connected AI service learn from the question?

Pricing makes the decision clearer

Scite's current individual Basic plan is listed at $20 per month when billed yearly, with a seven-day free trial. The page advertises unlimited Assistant usage, full-text literature search, Smart Citations, integrations, and Collections containing up to 10,000 papers. Annual billing therefore makes the visible individual cost $240, with the stated saving measured against monthly billing.

Team and enterprise access is quoted separately. Institutions should ask about authentication, administration, usage reporting, API rights, Reference Check automation, publisher coverage, MCP access, data handling, training, support, and how off-campus or affiliated researchers are licensed.

The price is easier to justify for someone who regularly evaluates evidence than for someone who only needs occasional paper discovery. Semantic Scholar and PubMed cover broad search without an individual subscription. ResearchRabbit and Connected Papers are strong for exploration. Elicit focuses more heavily on structured review and extraction. Consensus offers direct evidence synthesis around questions.

Scite earns its fee when citation context changes decisions. If the user frequently needs to see whether later work supported, narrowed, or challenged a finding, the product has a distinctive role. If the main requirement is drafting, reference management, PDF chat, or a one-off class assignment, another tool may cover more of the workflow for less.

The workflow WhatAI would actually trust

Begin with a question precise enough to search. Write the population, intervention or exposure, comparison, outcome, setting, date range, and study types where they apply. Save the original wording before asking Assistant to transform it.

Inspect the search strategy. Add missing terminology, spellings, abbreviations, and controlled vocabulary. Run complementary searches in the databases expected by the field. Scite's breadth is valuable, but a serious review should not quietly become dependent on one index.

Use Smart Citations to prioritise reading, not to vote on truth. Open the citation statements, then open the papers. Separate direct replications from background mentions, methodological criticism, population differences, and theoretical disagreement.

Build a Collection with explicit inclusion rules. Record who added each paper and why. Export records into the reference manager or review system that holds the durable project history. Turn on alerts only for questions someone is responsible for monitoring.

If MCP is enabled, use the Scite dashboard to review which tools are being called. Keep sensitive questions out of unapproved AI clients. Require the assistant to provide traceable sources and mark uncertainty. Recheck every citation attached to a high-stakes claim.

Before submission, run Reference Check and investigate every retraction, correction, contrasting citation, and unusual source. Then conduct the slower human check: does each reference actually support the words written?

The WhatAI verdict

Scite is becoming more valuable because it is resisting the temptation to become a generic writing assistant. Its strongest asset is still the citation statement: the small piece of text that shows how later research treated an earlier paper. The 2026 product changes put that evidence into more useful places.

The unified Research Feed turns literature searching into monitoring. Collections create a living set of papers rather than a forgotten folder. MCP brings Scite into ChatGPT, Claude, Gemini, Copilot, and coding tools. The new usage dashboard gives users some visibility into what those connections are doing. Reference Check remains a practical safeguard near publication.

This makes Scite one of the more credible additions to an AI research stack, but credibility comes from inspectability, not automation. Its labels are model outputs. Its database has boundaries. Its Assistant can still misunderstand a question. A paper with many supporting citations can still be weak, and a new paper with no citations can still be important.

Choose Scite when the question is not only what has been published, but what happened to a claim after publication. Use it to find the challenging paper, the confirming study, the retraction, the overlooked citation statement, and the new evidence that arrives after a review is supposedly finished.

Then read the work. Scite can make the trail clearer. It cannot walk the final mile of judgement for you.

WhatAI Decision Box

Scite combines scholarly search with citation context, showing whether later papers support, contrast, or mention earlier work. Its 2026 workflow includes Assistant, Reference Check, Collections, alerts, a unified Research Feed, and MCP access from major AI tools.

The individual Basic plan is listed at $20 per month when billed yearly and includes a seven-day trial. Team and enterprise access is custom. Scite is strongest when citation context matters, but its automated labels do not replace methodological appraisal or reading the source.

Share how you use Scite for literature reviews, evidence monitoring, Reference Check, Collections, Smart Citations, Assistant, MCP, or institutional research. Include the field and workflow because coverage and citation patterns vary.

Use Cases

Pricing varies by plan and region — see current pricing.

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

Scite Community Discussions

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

Scite Pros & Cons

How to Get Results with Scite: Step-by-Step Workflow

1. Define the research decision

   Write the claim, population, intervention or exposure, comparison, outcome, setting, date range, study types, and what evidence would change the conclusion.

2. Build a transparent search

   Ask Assistant or Search, then inspect and save the actual strategy. Add synonyms, abbreviations, spellings, subject headings, exclusions, and complementary databases required by the field.

3. Screen for relevance

   Review titles, abstracts, dates, study designs, populations, outcomes, and access status. Record inclusion and exclusion reasons outside the chat interface.

4. Inspect Smart Citations

   Use supporting, contrasting, and mentioning labels to prioritise reading. Open the citation statements and separate replication, criticism, background use, and population differences.

5. Read and appraise the papers

   Check methods, sample, outcomes, effect size, uncertainty, bias, funding, conflicts, corrections, retractions, and applicability. Do not treat a citation label as a quality score.

6. Build a governed Collection

   Group included and sentinel papers with a defined purpose, owner, dates, inclusion rules, version history, and export plan. Enable only alerts that can lead to a decision.

7. Use MCP with controls

   Connect only approved AI clients, review the MCP User Dashboard, keep sensitive questions out of unapproved tools, and require traceable sources and explicit uncertainty.

8. Verify every synthesis

   Compare each generated claim with the cited passage and full paper. Correct overstatement, distinguish association from causation, and preserve disagreement and uncertainty.

9. Run Reference Check

   Before submission or publication, upload the manuscript, investigate every retraction, correction, contrasting citation, and unusual source, then manually verify that each citation supports the written sentence.

10. Monitor and update

    Review the Research Feed on a defined cadence, triage new evidence, document changes to conclusions, export durable records, and schedule a formal review when alerts reveal material change.

Scite Gotchas and Limits to Know Before You Start

• Smart Citation labels are automated interpretations of citation statements and can misclassify nuanced language.
• Supporting, contrasting, and mentioning labels do not assess full study quality, bias, effect size, or applicability.
• A citation count measures attention and use, not truth, replication, clinical importance, or methodological strength.
• Recent papers and preprints may have few citations simply because later literature has not had time to appear.
• Coverage varies by publisher, field, document type, language, full-text availability, and indexing delay.
• The Assistant can misunderstand a question, omit terminology, overstate evidence, or attach a source too broadly.
• Scite should complement subject databases and established systematic-review methods rather than silently replace them.
• Full-text access depends on open access, institutional entitlements, publisher arrangements, or purchase options.
• MCP use can send research questions and context through external AI services with separate privacy and retention rules.
• Collections require protocols, owners, inclusion logic, exports, and version history to become reproducible review artifacts.
• Alerts can create noise if papers, researchers, and topics are followed without a decision or responsible reviewer.
• API Reference Check requires paid enablement and is not included in a standard self-service Pro key.
• The Basic price is presented per month but billed yearly, so users should confirm the full annual commitment.
• No verified Scite affiliate program was found, and the old ref query was not retained.
• Reference Check flags concerns but cannot prove that every manuscript sentence accurately represents its source.

Which Scite Feature Fits Your Use Case

Starter Prompts for Scite

What is Scite?

Scite is a scholarly search and citation-intelligence platform. It shows citation statements in context and classifies them as supporting, contrasting, or mentioning the cited work, alongside Assistant, Reference Check, Collections, alerts, MCP, and API tools.

How do Smart Citations work?

Scite analyses the text around a citation and uses a model to classify the citation statement. Users can read the statement and surrounding context rather than relying only on a total citation count.

Do supporting citations prove that a paper is correct?

No. A supporting label describes how one citation statement relates to the cited work. It does not assess the complete methods, bias, sample, effect size, conflicts, corrections, or wider body of evidence.

How much does Scite cost?

The official pricing page lists Basic at $20 per month when billed yearly, with a seven-day free trial. Team and enterprise pricing is custom. API and automation rights can require separate arrangements.

Does Scite have a permanent free plan?

The current pricing page promotes a seven-day trial rather than a permanent free individual plan. The MCP page also allows users to try the connection with an account, but sustained feature access depends on the applicable subscription.

What does Scite MCP do?

Scite MCP connects Smart Citations, scholarly search, Collections, and literature context to compatible tools including ChatGPT, Claude, Gemini, Copilot, Cursor, and Claude Code.

What is the Scite Research Feed?

The unified Research Feed combines the earlier Feed and Alerts experience. Updates from followed papers, researchers, and topics flow into one place so users can monitor changing evidence.

What are Scite Collections?

Collections are curated groups of papers that users can return to, monitor, and manage. They can alert users to new supporting or contrasting citations and retractions and can be managed through MCP-enabled AI tools.

What does Reference Check find?

Reference Check extracts references from a PDF or DOCX and evaluates them against Scite data, surfacing retractions, contrasting citations, and other source flags for manual investigation.

Can Scite replace a systematic-review database?

Not by itself. A rigorous review may require multiple subject databases, registered protocols, reproducible strategies, duplicate screening, extraction, risk-of-bias assessment, and full-text appraisal beyond Scite.

Does Scite provide full-text access to every paper?

No. Access depends on open-access status, publisher coverage, institutional entitlements, and document-delivery options. Search results and citation statements are not a universal licence to read or process every full paper.

Is Scite owned by Research Solutions?

Yes. Research Solutions acquired Scite in November 2023, and Scite now operates as part of the Research Solutions product family.

Related Research & Knowledge Work Tools

Explore the Network