How to Check What Research Says With Consensus

November 8, 2026

Consensus answers yes/no questions with a share of papers: “Does intermittent fasting improve insulin sensitivity?” returns “Yes, 78% of 42 papers”. Then it lists those papers with one-line findings.

That format is the reason it’s useful for a quick reality check and dangerous if you stop there. The percentage is a summary of abstracts, not a meta-analysis.

What is Consensus and how is it different from Elicit?

Consensus is a search engine for research questions, built by Consensus NLP, launched in 2022. It indexes over 200 million papers, pulling from PubMed, Semantic Scholar and other sources.

The difference from a literature-review tool: Consensus is built around the direction of findings — positive, negative, mixed — and aggregates that into the Consensus Meter. Elicit is built around extraction into a table.

Use Consensus when the question is “does X do Y”. Use Elicit when the question is “what do these 60 studies measure and in whom”.

How do I ask a question that gets a clean meter?

Phrase it as a claim you could test.

Works:

  • “Does caffeine improve athletic endurance performance?”
  • “Is remote work associated with lower employee turnover?”
  • “Does vitamin D supplementation reduce depression scores?”

Fails:

  • “What is good for sleep?” — no claim, no direction
  • “Why do people procrastinate?” — too open for a yes/no meter
  • “Is AI bad?” — undefined terms

The meter only appears for yes/no framing. For open questions you get a plain list, which is still useful but loses the summary.

What does the Consensus Meter actually measure?

It shows the proportion of retrieved papers whose abstract supports, contradicts, or is neutral on the claim. Three things to know:

  • The denominator is the retrieved set, not the literature. Change the phrasing and the set changes. 78% of 42 is not 78% of the field.
  • Study quality is not weighted. A randomized trial and a cross-sectional survey count the same.
  • Sample size is not weighted. A 20-person pilot and a 50,000-person cohort contribute equally to the bar.

That’s why you should read the meter as “how much disagreement is there in this slice”, not “what is true”.

For a quick sanity check before a meeting, it’s fine. For a literature claim in a paper, cite the underlying studies.

How do I use the filters well?

The filter panel is the difference between a usable answer and a useless one:

  • Study type — RCT, meta-analysis, observational. Filtering to RCTs often flips the meter’s story, because observational studies dominate nutrition and behaviour literature.
  • Year range — set to the last 5 years for fast-moving fields, 10+ for basic science.
  • Sample size — the underused filter. Most public users never touch it.
  • Journal quality — where available, narrow to peer-reviewed venues.

Always click into the individual papers. Each result shows a one-sentence takeaway pulled from the abstract; reading five of those takes two minutes and catches the two papers that actually decided the question.

What does it cost?

As of late 2026: Free gives a limited number of questions per month plus unlimited search within a cap. Premium is about $8.99/month billed annually with unlimited Consensus Meter questions and GPT-4-based features; Enterprise is custom for institutions.

Lower than most research tools. For students and clinicians doing quick checks, Premium is a fair $9.

Where does Consensus mislead?

  • Consensus is not correctness. A well-run trial contradicting 40 weak ones is still the better evidence, and the meter can’t show that.
  • Health and nutrition fields have publication bias. Positive findings get published more; the meter inherits that.
  • Abstract-level reading misses subgroup effects that are often the real story.

If a decision is high-stakes — clinical, medical, safety — the meter is a starting point and nothing more.

Quick Summary

  • Consensus answers yes/no research questions and shows the share of papers supporting each direction.
  • Frame questions as testable claims; open questions return lists without a meter.
  • The meter reflects a retrieved slice, not the whole field, and doesn’t weight study quality or sample size.
  • Filter to RCTs and recent years to avoid observational-study dominance.
  • Free tier covers light use; Premium is about $8.99/month — but never treat the percentage as ground truth.