AI research tools

Choose an AI research tool by the sources your question requires and the work needed to check its answer. Web search finds current public material; a selected-document workflow examines an approved collection; paper search finds studies. Citations make checking possible, but do not establish that an answer is correct.

Start with the source you need

For each important claim, identify the source that could settle it. A current product capability calls for current provider documentation. An internal policy calls for the approved version held by its owner. A research finding calls for the relevant paper, including its methods and limits.

Decide what the answer must include: the finding, its date or version, relevant conditions, conflicting evidence and points the available sources cannot resolve. An incomplete source set cannot be repaired by a fluent summary.

Four research routes

Compare only candidates that can reach the sources the task requires. Each route below is documented by its provider; confirm current behaviour against those pages.

Perplexity Pro Search — a search route. Perplexity documents Pro Search and file uploads. Check which files an answer actually used, and how supporting material can be inspected.

ChatGPT Deep Research — a research-and-report route. OpenAI documents Deep Research. Check what sources it can use and how a report exposes supporting material.

Gemini Notebook — a selected-document route. Google documents adding or discovering sources and using chat in Gemini Notebook. Check what content is included from imported material and whether the workflow is confined to the intended collection.

Elicit — an academic-paper route. Elicit describes meaning-based paper search and screening criteria. Its documentation cautions that extraction can miss information in complex tables. Extracting data from figures requires Scale or Enterprise and particular settings. Check access to the underlying paper and read the relevant methods or results before relying on an extracted finding.

Check the paper-search pool and ranking

For a paper-led question, check that the search pool matches the evidence you need. Elicit says its academic-paper search covers more than 138 million papers and conference proceedings, with weekly updates, and includes more than 545,000 clinical trials registered on clinicaltrials.gov, updated in real time. These are distinct collections, so confirm which type of record the question calls for.

Elicit says it ranks papers primarily by similarity in content or meaning, not keyword frequency, also considering features such as citation count and recency. The service says it takes the top 1,000 papers for further prioritisation with language models and may show anywhere from the top 4 to the top 1,000, depending on the plan and workflow. Treat this as a discovery aid: check whether important papers are present and whether the evidence is current enough for the question.

Elicit says it helps specify broad research questions and points out ways to clarify your analysis, and users can run custom screening questions over the papers found to narrow them further. Before relying on a screened set, make sure the question and screening criteria reflect the population, intervention or subject, and outcome that matter to your task.

Where the available collection is not enough, Elicit describes an option to upload documents or integrate company subscriptions to access more full-text papers. Check whether the relevant papers are actually accessible.

Search results and extracted answers are not substitutes for reading the underlying methods and results.

Elicit Academic Paper Search: Coverage and Ranking

Check the answer before choosing

Give suitable candidates a question with a known answer, an important qualification and a point the available sources cannot establish. Keep the substantive source material comparable where the workflows permit.

Record the product, plan, settings, sources and date, including unsuccessful answers. A single good response cannot establish how a workflow performs across ordinary and difficult questions.

For an Australian organisation, check applicable privacy obligations and the actual account and data arrangements before considering restricted files. Begin with authorised, non-sensitive material.

Choose a workflow for a defined question and source set. Keep a person responsible for consequential claims, and revisit the choice when the sources, plan or product change.