Short answer: if your main problem is finding relevant papers to build a review, Elicit is the most coherent starting point of the three; if you want to start from a question and quickly map what the literature says, Consensus is more direct; if the task is to check how a paper or claim was cited later, Scite has the most specialized proposition. For rigorous research, however, the best strategy is not to choose a universal winner: it is to separate literature retrieval from citation auditing and combine tools with a transparent search in databases such as PubMed or the reference database for your field.
The short answer: there is no single winner
| Task | Best fit | Why | Main limitation |
|---|---|---|---|
| Discover relevant papers without already knowing every keyword | Elicit | Semantic search, relevance reranking, and screening/extraction workflows | It does not replace an exhaustive, documented search strategy |
| Ask a question and quickly understand the direction of the evidence | Consensus | Combines academic search, synthesis, and features such as the Consensus Meter | The synthesis depends on the retrieved set and can miss nuance |
| Check how a study was cited, challenged, or supported | Scite | Smart Citations show citation context and classify statements as supporting, contrasting, or mentioning | That strength does not mean it is the best paper retriever for every question |
The distinction matters because “finding the right paper” and “knowing whether a citation really supports a claim” are different problems. Putting everything into one ranking favors whichever tool happens to match the chosen metric.
What the July 2026 study actually measured
The strongest recent comparison is the study Comparative evaluation of artificial intelligence–assisted literature search tools for identifying clinically meaningful evidence in cardiology, published on July 28, 2026 in the European Heart Journal – Digital Health. It compared ChatGPT-5, Elicit, Consensus, Scite, and PubMed across four cardiology topics, using standardized queries and blinded assessment by three experts.
The design is more useful than a simple “who won” because it separated two gold standards. The first measured relevance: experts judged whether each retrieved reference was pertinent to the question. The second measured retrieval of key references: for each topic, the authors defined a set of central studies based on expert assessment. In other words, a tool could return many plausible papers and still miss studies that experts considered essential.
In that experiment, ChatGPT-5 had the highest proportion of papers rated relevant, 90%, and the highest overlap with key references, 60%. Scite was at the lower end, with 20% relevance and 10% overlap with key references; Elicit, Consensus, and PubMed occupied intermediate positions that varied by topic. Prompt formulation did not have a significant effect on relevance, while the tool itself explained a substantial share of the performance variation.
This does not turn the study into a universal ranking. The tests were conducted in cardiology, and the free web interfaces were accessed between August 2025 and January 2026. Since then, Elicit, Consensus, and Scite have received meaningful changes in search, full-text access, agents, filters, and evidence organization. The result should be read as evidence that tool architecture and coverage matter—not as a definitive table for every discipline in August 2026.
Did Scite “lose”? Only if the question is literature retrieval
Scite’s low result in the benchmark matters, but it answers a specific task: retrieving the most relevant papers and key references. Scite’s main current differentiator sits at another layer: showing how later papers cite a publication and letting users navigate citation passages classified as support, contrast, or simple mention. Judging that function only by literature recall would be like rating a reference checker by the quality of its automatic summary.
Elicit: strongest when the bottleneck is search, screening, and extraction
According to official documentation consulted in August 2026, Elicit Search searches more than 138 million academic papers and conference proceedings, while also connecting to other evidence workflows. Its semantic search is designed to find studies even when the query and paper title do not share the exact same words. The system first ranks the corpus by similarity, then performs a deeper review of top results before presenting the final set.
In practice, that design favors users who start with a conceptual question and do not yet know the field’s vocabulary. Elicit also connects search to screening, data extraction, reports, and systematic reviews. For a review that needs to move from dozens or hundreds of references to a comparable table of population, intervention, outcome, and results, that workflow continuity is a genuine advantage.
The caution is not to confuse “convenient semantic search” with “exhaustive search.” Formal systematic reviews still need to record databases, strategies, dates, criteria, and inclusion decisions. Elicit itself positions search as part of a broader process; when reproducibility is the goal, an explicit strategy in PubMed, Embase, Scopus, Web of Science, or another field-specific database remains the backbone.
Consensus: strongest when the question comes before the bibliography
Consensus accepts natural-language questions, keywords, and Boolean queries. In more advanced modes, Research Agent can break complex questions into multiple searches, apply filters, read across papers, and return a citation-backed answer. In 2026 the product also expanded filters for study design, publishers, and human/animal settings, as well as full-text access from publishing partners when available.
This format works well for questions such as “is there evidence that X improves Y?” or “which factors are associated with Z?”. The Consensus Meter analyzes leading results for compatible questions and shows the distribution across positions such as “yes,” “no,” “possibly,” or “mixed,” with additional indicators for methods, recency, journal, and citations.
The advantage is orientation speed. The limitation sits in the same place: a visual synthesis can look more conclusive than the literature really is. Consensus’s own documentation warns that classifications can be wrong, that question-specific nuance can be lost, and that its database does not contain all scientific research. The Meter is therefore a gateway to reading, not a substitute for evaluating study design, population, outcomes, and limitations.
Scite: strongest when the question is “how has this work been cited?”
Scite says it covers more than 300 million scholarly items and has indexed more than 1.6 billion citation statements. Its distinctive feature is Smart Citations: instead of only showing how many times a paper has been cited, the system shows the passage where the citation appears and classifies that citation statement as supporting, contrasting, or mentioning.
For auditing an existing bibliography, this is especially useful. The Reference Check feature can analyze the references in a manuscript and look for retractions, editorial notices, or contrasting citations. Full-text search and Assistant broaden the tool toward discovery, but its most distinctive editorial value remains auditing the citation trail.
There is an important caveat: the label applies to a citation statement, not to the paper as a whole. One article may cite another positively in one paragraph and disagree with an interpretation elsewhere. “Supporting” also does not mean “true”; it means that the cited passage was classified as supportive. Reading the context and the original study remains necessary.
How to run a fair benchmark with a small Portuguese-language gold standard
If the choice matters for a lab, newsroom, course, or research group, it is worth testing all three tools on your own topic. The cardiology study’s method can be adapted without turning the exercise into a months-long project.
- Define a narrow question. Use a real question in Portuguese, including population, intervention/exposure, and outcome when appropriate.
- Build a small gold standard. Set aside 12 to 20 references that subject-matter experts consider central. A recent systematic review, a guideline, and a manual database search can help compose the set. Record DOI, year, and the reason for inclusion.
- Standardize the test. Run the same question in Portuguese and, if the field publishes mainly in English, an equivalent English version. Use the same date, the same plan level, and clean history.
- Measure retrieval and relevance separately. Calculate recall@10 or recall@20 as the share of gold-standard references appearing in the first results. Calculate relevance@10 as the share of the top ten results judged genuinely pertinent by an evaluator.
- Measure citation auditing on a separate axis. For key papers, check whether the tool exposes citation context, helps locate contrasting evidence, and makes the passage traceable to the citing work. Do not mix that result with search recall in a single score.
This separation avoids a common mistake: penalizing Scite for not being the best discovery engine when the intended use is auditing, or rewarding Elicit/Consensus for synthesis quality when the actual problem is whether a specific reference was challenged later.
Which one should you choose in each scenario?
- “I’m entering a field and need to find the right studies.” Start with Elicit, then confirm coverage and terminology in a traditional bibliographic database.
- “I have a question and want to understand quickly where the evidence points.” Consensus tends to offer the most direct experience, as long as you open the studies behind the synthesis.
- “I already have the papers and want to know whether my citations are well supported.” Scite is the clearest fit, especially for tracing support, contrast, retractions, and context.
- “I am doing a review that must be defensible and reproducible.” Use a field-standard database as the documented backbone, add Elicit or Consensus for complementary discovery, and use Scite as a citation-audit layer.
What not to do with any of the three
- Do not cite an AI synthesis without opening the original paper.
- Do not assume a proprietary database contains “all the literature.”
- Do not treat citation counts or a “supporting” label as an automatic measure of quality.
- Do not change prompts, filters, and time windows between tools and then call the result a benchmark.
- Do not use an AI-generated reference list as a silent replacement for a documented search strategy when the work requires reproducibility.
Verdict
For readers who need to find papers and organize screening/extraction, Elicit is the most complete fit of the three by workflow scope—not because the cardiology study proved universal superiority over Consensus. For users who want to start with a question and get a quick evidence map, Consensus is more natural. For anyone who needs to verify the context and trajectory of a citation, Scite still has the most specialized function.
The 2026 study reinforces the main caution: performance changes substantially by tool and problem, and neither a polished interface nor a good summary guarantees coverage. The more robust choice is to treat search and verification as separate stages—and keep a transparent bibliographic database at the center whenever the research must be auditable.
