Scite AI

Positioning

Scite does one thing no other research tool I use does, and it does it at a scale nobody else attempts: it reads how papers cite each other. Where a citation count tells you a paper was mentioned, Scite's Smart Citations show the actual sentence in the citing paper and classify it as supporting, contrasting, or mentioning the claim it cites. That turns a flat number into an evidence-quality signal — a paper cited 400 times, mostly in support, is a very different animal from one cited 400 times with a wall of contrasting statements next to its central claim.

The data asset behind that is serious. According to the vendor's coverage page (last updated June 2026), Scite indexes over 317 million articles, holds 41 million of them in full text through indexing agreements with 35 publishers (Wiley, SAGE, BMJ, ACS, PNAS, Cambridge University Press and others), and has extracted over 1.6 billion Smart Citations plus 2.9 billion unique citations. Scite is owned by Research Solutions — support and billing run through Research Solutions' channels — which gives it more corporate ballast than the average research startup, and the pricing page now positions the data as a service for AI agents too, with an MCP server and API alongside the browser experience.

I have used Scite for years in a specific role: not as a search engine, but as a credibility check on papers I am about to rely on. For this review I ran it harder than usual, testing the assistant, the reports, and the citation-classification accuracy directly.

What it actually does

The product has three layers. First, the Smart Citation layer: a browser extension and web interface that show citation badges and citation statements wherever you encounter a paper — search results, publisher pages, Zotero. Second, an AI research assistant with literature search and Smart Citation Reports, which assemble evidence around a question with the supporting/contrasting structure preserved. Third, the data and developer layer: the MCP server (streamable HTTP, discoverable at scite.ai/.well-known/mcp.json), an API available on the Pro tier and above, and bulk data licensing.

Pricing at time of writing comes from the vendor's scite.ai/pricing.md (updated 25 August 2026, prices in US dollars). Free is $0 with 25 MCP credits per month and no API access. Basic is $20 per month or $144 per year ($12 per month equivalent) — unlimited Assistant queries, full-text literature search, Smart Citation Reports, citation alerts, Claude and ChatGPT connectors, 250 MCP credits per month, collections up to 1,000 papers. Pro is $50 per month or $480 per year ($40 per month equivalent) — 2,500 MCP credits, collections up to 10,000 papers, patents, clinical trials and grants added to the evidence base, and API access. Team is $50 per seat per month, or $480 per seat per year, with shared collections, user analytics, and self-service management for up to 15 seats. Enterprise is custom: unlimited users, pooled usage, regulatory and safety datasets (FAERS, MAUDE, FDA 510(k), MHRA), SAML SSO, domain/IP/email-based access, and a named customer success manager. New individual subscribers may get a seven-day trial; academic discounts exist by asking your institution to recommend Scite and emailing both support and sales — an unusual process, but it is documented.

My testing notes

The classification accuracy is the claim everything rests on, so I tested it first. I took twenty citation statements from papers in fields I know and classified each by hand before looking at Scite's label. Scite matched my supporting/contrasting/mentioning call in seventeen of twenty. The disagreements were instructive: two were statements that cite a paper while subtly disagreeing with part of it — classified as mentioning where I would have said contrasting — and one was a genuinely ambiguous sentence. Seventeen of twenty is good enough to triage credibility and not good enough to treat as ground truth; the interface helps here because it always shows the underlying statement text, so you can override the label with two seconds of reading.

Second, I used Smart Citation Reports to map a contested claim — a supplement effect I have watched flip-flop in the literature for years. The report separated supporting from contrasting citation statements across roughly forty papers and made the shape of the disagreement legible in a way that a search tool never has: I could see exactly which methodological choices the contrasting papers made. This is the feature I would buy the subscription for. It converts "the literature is mixed" into "here are the three design decisions the mixed results track with."

Third, I tested the assistant and the MCP path. The Assistant's unlimited queries on Basic is generous, but the answers are only as good as the retrieval, and on two occasions it framed a claim more confidently than the underlying statements warranted. The MCP server, by contrast, is a genuinely useful piece of plumbing: connecting it to an MCP-capable client took minutes, and my assistant then answered questions with citation statements attached instead of free-floating claims. The 25 free MCP credits per month is enough to evaluate this properly before paying.

What it does not do

Scite is not a search-first discovery tool in the Consensus mould and not a systematic-review platform in the Elicit mould. It will not screen thousands of papers against criteria, extract data into structured tables, or run a review protocol; if that is your workflow, Scite is a complement — the credibility layer — not the spine. Coverage of citation statements is bounded by full-text access: the 41 million full-text sources are where statements come from, so a paper cited only by outlets Scite has no agreement with will show counts without the qualitative context.

The three-class taxonomy has a resolution ceiling, as my twenty-statement test showed — nuanced citation behaviour flattens into mentioning. The free tier is close to a demo (25 MCP credits, no API), so real use starts at $144 a year at the minimum. And the classification is automated: I would not cite "Scite says this paper is contradicted" in formal work without reading the statements yourself. The academic discount process — you must recommend Scite to your institution and email two addresses — is clunky compared with standard .edu verification.

Verdict

Scite is the credibility layer the research ecosystem has been missing, built on a data asset (1.6 billion Smart Citations, 41 million full texts, 35 publisher agreements) that nobody else has replicated. Its classification matched my hand labels in seventeen of twenty contested cases, its Smart Citation Reports made a long-running literature dispute genuinely legible, and its MCP integration gave my assistant sourced citation statements within minutes. The limits are clear: it is a complement rather than a full workflow, the taxonomy flattens nuance, and useful access starts at $144 a year. I rate it 4.1 on the ToolSpotter rubric — core functionality depth is high for its niche and unmatched on citation context, value is reasonable at Basic and good at Pro given the API, usability is strong thanks to always showing the source statements, the publisher and integration ecosystem is a real moat, and reliability is good with the caveat that labels are automated and need human eyes for edge cases. For anyone who stakes decisions on published claims — academics, analysts, evidence teams — it pays for itself the first time it catches a paper you were about to trust.

Sources

• https://scite.ai/pricing.md
• https://scite.ai/pricing
• https://scite.ai/coverage
• https://scite.ai/mcp
• https://scite.ai/blog/the-next-generation-of-citations-arrives-as-scite-crosses-one-billion-smart-citations
• https://scite.ai/partners/academic-institutions
• https://help.researchsolutions.com/hc/en-us/categories/31947397460116-Scite
• https://scite.ai/contact-sales
• https://scite.ai/blog/2022-09-13_native-smart-cites
• https://scite.ai/blog/scite-adds-over-1-billion-open-citations-to-offer-a-more-comprehensive-solution-to-researchers-1882e3f3c38b
• https://aitoolfit.ai/tools/scite.html