What is Consensus?
Consensus is an AI search engine for scientific and academic research. It searches over 200 million academic papers and uses language models to help you find, understand, and synthesize the literature faster.
For every query, Consensus first retrieves the most relevant academic papers. Then, our AI analyzes the top results and generates a clear, cohesive synthesis of the findings. Every response includes citations, so you can trace each insight back to the original source.
What Does Consensus Search Over?
Our corpus includes over 200 million scientific documents across all domains of science —primarily peer-reviewed journal articles, along with some conference papers and preprints.
We’ve built this corpus by aggregating data from three major sources:
- Semantic Scholar
- OpenAlex
- Our own crawl of the scholarly web to fill in important coverage gaps
By combining these sources, Consensus covers nearly all of the highest-impact journals and the entirety of PubMed. Think of Consensus as an AI-native alternative to Google Scholar with a more-refined corpus.
How Does a Consensus Search Work?
When you run a search on Consensus, we try to find the most relevant scientific papers to deliver to you and ground our AI-generated summaries.
Behind the scenes, Consensus uses a classic multi-step approach for information retrieval to narrow in on the best sources.
Step 1: Cast a wide-net
First, we scan our entire 200-million-paper corpus to find the most relevant results. This is done by matching your query against the titles and abstracts of each paper using a hybrid search approach that combines:
- Semantic search, powered by AI embeddings, to capture the intent behind your question and support natural language queries.
- Keyword search, powered by BM25, a proven traditional method, this anchors results to the exact terms in your query for precise keyword matching.
It’s like having a librarian who not only finds books that match your exact words, but also understands what you’re really looking for, and brings you both.
These two methods work together to assign a relevance score to each paper. Think of this step as a fast, intelligent filter that screens the entire corpus to find what’s worth reading next.
Step 2: Refine by Quality
After the initial relevance filter, we take the top 1,500 papers and re-rank them, not just by how well they match your query, but also by how strong the research is.
This second pass blends the textual relevance score from step 1 with research quality signals like:
-
- Recency of publication
- Citation count
- Journal reputation and impact
The result is a prioritized list of papers that are not only relevant to your question, but also more likely to reflect high-quality, credible science.
Step 3: Produce final ranking of papers
In the final step, we focus on precision, ranking the top 20 papers as accurately as possible.
This step works like Step 2, but with two key upgrades:
- We recalculate textual relevance using a larger, more powerful AI model that is optimized for precision, not just broad recall.
- Because this model is more compute-intensive, we run it only on the top 20 papers.
We still factor in research quality signals like recency, citation count, and journal impact, so the final list reflects both relevance and rigor.
How Does Consensus Use AI Responsibility?
At Consensus, we’re committed to using AI ethically, transparently, and with purpose. Always in service of helping researchers, students, and professionals save time without compromising trust and transparent attribution.
The simplest way to understand our approach is this: We only use AI after we search the scientific literature. This ensures that every response is grounded in real, citable research, not speculative content generated by a model.
Once relevant papers are retrieved, we use AI in two key ways:
- To analyze individual papers in depth (e.g. Ask Paper, Study Snapshot)
- To synthesize findings across multiple papers (e.g. Pro Analysis, Consensus Meter)
We use two types of AI models in these AI features:
- Commercial models (like OpenAI) for general-purpose summarization
- Fine-tuned open-source models for domain-specific features, like the Consensus Meter
Regardless of the model or use-case, we never use user data to train systems, and we never share your data with third parties. Your data stays private and stays yours.
How Consensus Deals With Hallucinations:
Hallucinations are a common issue in AI systems where models generate something that is not true. These usually fall into three categories:
- Fake sources – the AI cites a paper or article that doesn’t exist.
- Wrong facts – the AI generates a confident answer from internal memory that’s simply incorrect with no source.
- Misread sources – the AI summarizes a real paper or source, cites it, but gets it wrong.
Thanks to how we’ve built Consensus, only the third type of hallucination is possible, and we work hard to minimize it.
Consensus isn’t a chatbot. It’s a search engine that uses AI to summarize real scientific papers. Every time you ask a question, we search a database of peer-reviewed research.
That means:
- Every paper we cite is guaranteed to be real
- Every summary is based on actual research, not a model’s guess or internal memory
Still, no AI system is perfect. Sometimes a model can misinterpret a paper and summarize it incorrectly and this can happen in Consensus. To reduce this risk, we’ve added safeguards like “checker models” that verify a paper’s relevance before summarizing it. If we cannot find any relevant information to your query, we will tell you, not let a model loose to fill in gaps that don’t exist. This safe guard is our attempt at trying to “set the models up for success”.
Most importantly, we’ve designed the product to make it easy for you to dive into the source material yourself. The best use of Consensus isn’t just getting a quick summary, it’s using our tools to explore the research in depth. That’s when real understanding happens.