Why research tools need a different test
A writing assistant can be judged on style. A research tool must be judged on truth. The worst failure is not a vague answer but a confident one supported by a citation that does not say what the tool claims. So our test focused on verification: for every answer, we opened the cited sources and checked whether they supported the specific claim.
We used three question sets. Ten general-knowledge questions with recent developments, ten academic questions from health and climate science, and five deliberately ambiguous questions where the honest answer is "the evidence is mixed". The last group proved the most revealing.
Answer engines
Perplexity — best for fast, sourced answers
Perplexity combines a language model with live web search and places numbered citations next to each sentence. In our checks, the large majority of citations supported their claims, and it was quickest at summarising recent news. It sometimes leans on secondary sources such as blogs when primary sources exist, so for important claims, click through. Its focus modes for academic papers and specific sites help narrow the search.
ChatGPT and Gemini with search
Both general assistants now browse the web and cite pages. They are convenient if you already use them for writing, but their citations were less consistently linked to specific sentences, which makes verification slower. Use them for orientation, not for final references.
Academic research assistants
Elicit — best for literature reviews
Elicit searches a large corpus of academic papers and extracts key details such as sample size, methods and outcomes into a table. For anyone starting a literature review, this turns days of skimming into an afternoon. It handled our ambiguous questions well, surfacing papers on both sides rather than forcing a single conclusion.
Consensus — best for yes-or-no evidence questions
Consensus is built around questions like "does creatine improve memory?" It shows a meter summarising how many studies lean each way, with links to each paper. It is excellent for a quick evidence check, though the meter can oversimplify; always look at study quality, not just count.
Semantic Scholar and Google Scholar — the free foundations
Semantic Scholar, from the Allen Institute for AI, offers free paper search with AI-generated one-line summaries and citation graphs. Google Scholar remains the broadest index and the best place to check who has cited a paper since. Neither writes answers for you, which is precisely why they are reliable.
Managing what you find
Zotero — the reference manager we recommend
Whatever search tool you choose, you need somewhere to keep sources. Zotero is free, open source and captures citations from your browser in one click. It formats bibliographies in thousands of styles, including APA and Chicago, and syncs across devices. Pair it with Elicit or Perplexity and you have a complete, mostly free research workflow.
| Tool | Source type | Citation reliability | Free use |
|---|---|---|---|
| Perplexity | Web + papers | High | Yes, generous |
| Elicit | Academic papers | High | Limited credits |
| Consensus | Academic papers | High | Limited searches |
| Semantic Scholar | Academic papers | Direct links only | Fully free |
| Google Scholar | Academic papers | Direct links only | Fully free |
| Zotero | Reference manager | Not applicable | Free, paid storage |
A simple verification habit
Whichever tool you use, adopt one rule: never cite a source you have not opened. Read at least the abstract and the sentence that supports your claim. If a tool cannot produce a link to the original, treat its statement as a lead to investigate, not a fact.