EvidenceSeek

EvidenceSeek is a search tool for scientific literature that finds papers based on natural-language queries. Unlike a keyword search, EvidenceSeek lets you input a detailed empirical question or hypothesis, and it will find and rank papers based on a precise understanding of your query. It performs many database searches and then identifies the most relevant papers using detailed criteria that you can customize. Read more on the About page.

Enter a query in natural language. Adding more specific information will help narrow down the search results, but EvidenceSeek also works well with broadly formulated queries.

EvidenceSeek offers three input modes: Question takes a natural-language research question. Hypothesis lets you test a specific claim, such as “Probiotics reduce inflammation in IBD.” PICO lets you set Population, Intervention, Comparator, and Outcome separately, which gives you the most control over clinical searches.

Use these controls when you want more visibility or control over a search before it runs.

In Hypothesis and PICO modes, Add Additional Search Directions lets you enter free-text preferences that shape which papers are included and how they are ranked. For example, in a clinical query, adding “prefer studies with ≥6-month follow-up” will push those studies higher in the results. You can also set explicit inclusion or exclusion rules, such as “only include double-blind randomized controlled trials.”

For further visibility and control, Review Search Criteria First shows you the exact criteria EvidenceSeek will use to screen papers and lets you edit them before the search runs. You can add, drop, or reword a criterion, and set how important each one is.

When a search finishes, each paper in the results carries a relevance score, shown in blue, from 0 to 1. The score is EvidenceSeek's estimate of how much evidential weight a study merits for your query. Treat it as a measure specific to EvidenceSeek's ranking, not as a probability or an absolute grade. As a rough guide, a score above 0.7 is high, 0.3 to 0.7 moderate, and below 0.3 low. Because the score reflects relevance to your specific query, the same papers will receive different scores under a different search.

Traditional search engines like Google Scholar and PubMed rank results based on keyword match. While the search process itself is fast, keywords are a crude proxy for finding what you are actually looking for, so you may end up running many searches and sifting through hundreds of results.

EvidenceSeek differs in a few ways. First, it accepts your full research question or hypothesis in natural language, so you don't have to translate it into keywords. Second, it ranks results by their evidential relevance — how strongly each paper's findings bear on your question — rather than by keyword frequency or citation counts. In practice, one well-framed query usually surfaces the most relevant studies, instead of leaving you to filter through dozens of loosely related pages.

Elicit's paper search tool and the Allen Institute's (Ai2) Asta search use large language models to improve the search process, as does EvidenceSeek. But it diverges from them in several ways. First, it uses a different ranking approach: it explicitly assesses evidential relevance and incorporates it into an overall relevance score, rather than relying on topical match alone. For each candidate paper, EvidenceSeek judges how strongly its findings bear on your question or hypothesis, based on the study's methodological features, such as its design, sample size, and study population. This means it will often find highly relevant papers even when they have less surface-level overlap with your query.

Second, EvidenceSeek gives you more transparency and control over the search itself. Before a search runs, you can review and modify the criteria it will use to rank papers. For example, for a materials science query, you can require it to prioritize experimental synthesis studies over computational predictions and prefer those that report measurements taken at room temperature and ambient pressure. Once a search has completed, you can also add custom filters to the search results.

Finally, in our test searches, EvidenceSeek retrieves more relevant papers and ranks them more accurately than other AI-assisted search tools. In practice, this means that you're less likely to miss an important paper, but the easiest way to see the difference is to run the same search side by side and compare.

Yes. Use the Export button above the results to download a formatted report (PDF or Markdown) or a citation file (BibTeX or RIS) for your reference manager.

We don't keep a searchable record of your queries, build a profile tied to you, or store your searches and results in any database. They live briefly in our server's memory and in your browser session, and we temporarily cache some derived information to speed up repeated searches. To interpret your query and score papers, we send it to third-party AI providers and scholarly databases, which handle it under their own policies. Because of that, searches aren't fully private end-to-end, so don't enter confidential or personally identifiable information. See our Privacy Policy for details.

Currently, EvidenceSeek assesses relevance using paper abstracts and metadata. This means that for papers with sparse abstracts, EvidenceSeek's relevance score carries more uncertainty than it does for papers with detailed abstracts. In the future, EvidenceSeek may incorporate the full text of papers it can legally access.

Because scoring is performed in part by large language models (LLMs), EvidenceSeek can make mistakes in assessing studies. EvidenceSeek uses several strategies to limit their impact. Most importantly, the relevance score aggregates many diverse signals, including metadata the LLMs never see, so that no single mistake has a large effect. For judgments that matter most, EvidenceSeek also queries several LLMs independently and then combines their answers, which mitigates the risk posed by any single model's biases or errors. EvidenceSeek's search pipeline is continually tested to reduce errors, though no system of this kind can remove them entirely.