Short answer / top recommendation
Use a mixed workflow: Elicit or Perplexity (for evidence-linked extraction and quick scoping) + Zotero (collect PDFs/DOIs) + scite.ai (verify citation context). If you need repeated, reproducible reviews at scale, build a small RAG pipeline (GPT-4/GPT-4o or equivalent model + vector DB) that preserves metadata/DOIs for every retrieval.

Why this combination
- Elicit: designed for literature review tasks, extracts methods/results and returns paper metadata and links — excellent for systematic scoping.
- Perplexity / Consensus: fast Q&A with inline links to sources, good for initial synthesis and finding relevant papers.
- Zotero: reliable bibliographic collection, snapshots, and export to CSV/BibTeX for reproducibility.
- scite.ai: shows citation context (supporting/contradicting/mention), useful to spot unreliable or mis-cited claims.
- RAG + model: when you need integrated writing with traceable provenance and repeated queries, a custom retrieval pipeline keeps the source id/DOI attached to every generated claim.

Decision criteria (use these to pick a tool)
- Provenance: Does the tool return direct links and DOIs for each claim? (High priority.)
- Verifiability: Can you get the citation context (quotes, figures, methods)?
- Coverage: Are paywalled/full-text contents accessible for your field? (PubMed/Europe PMC vs paywalled journals.)
- Reproducibility: Can you export query logs, snapshots, or bibliographic data? (Needed for audits.)
- Cost & skill: Low-cost tools (Elicit/Perplexity/Zotero) vs. high-cost dev effort for custom RAG.

Best-for / Avoid-if
- Elicit: Best for systematic literature extraction and screening; avoid if you need full-text behind paywalls or custom metadata pipelines.
- Perplexity / Consensus: Best for quick evidence-backed answers; avoid using their answers as the final source without verifying DOIs and full text.
- scite.ai: Best for checking whether the literature supports a claim; avoid if you only need raw lists of papers.
- Custom RAG (LangChain + vector DB + GPT-4): Best for high-volume, reproducible reviews integrated into your workflow; avoid if you don’t have engineering resources or budget.

Practical checklist (step-by-step)
1) Create a precise search strategy (databases, keywords, time range). Save it. 2) Run broad scoping in Elicit or Perplexity to assemble candidate papers. 3) Pull PDFs and metadata into Zotero (capture DOI, URL, snapshot). 4) For each key claim, use scite.ai to pull citation context and check whether citations support the claim. 5) Extract method/results into a spreadsheet (or Elicit export). 6) Run targeted full-text checks (manual) for high-impact claims. 7) If repeating work, ingest PDFs into a vector DB and enable RAG that returns source id + quote + DOI. 8) Keep an evidence matrix (claim / source / quote / DOI / verification status). 9) Export references and append a reproducible appendix listing queries and snapshots. 10) Always spot-check model-generated text against the primary paper before citing.

When the right answer depends
- Budget: free/low cost (Elicit, Perplexity, Zotero); mid (Perplexity Pro, scite basics); high (custom RAG + paid APIs + scite subscriptions).
- Skill/team size: single researcher → Elicit + Zotero + manual scite checks; teams or labs → custom RAG + ingest pipelines and QA.

Concrete next step
If you want a ready-to-run setup: start with Elicit for extraction, Zotero for capture, and scite for verification; then evaluate whether building a small RAG pipeline is worth the engineering cost for recurring reviews.

Compare ChatGPT and Gemini