There is no single best paper search site, and the services that market themselves as one are usually the ones least willing to say what they index. What exists instead is half a dozen tools, each clearly best at one specific thing. Here is what each does well, what it quietly misses, and how to combine two or three of them.
The short version
| Site | Best at | Where it falls short |
|---|---|---|
| Google Scholar | The broadest first pass in any field. “Cited by” makes it a usable citation index, not just a search box. | Google does not publish what it indexes. Filtering is thin — a year range, a review-articles toggle, patents and citations on or off, and an advanced-search box for author and venue — with no field-level control like MeSH and no official API. There is no way to export a result set directly; you have to save items to My Library first. Peer-reviewed and unreviewed items sit side by side. |
| PubMed | Biomedicine and life sciences. MeSH terms and the query builder let you write a search precise enough to report. | Nothing outside its subject scope. Records are metadata and abstracts; full text lives with the publisher or in PMC. |
| arXiv | Physics, maths, computer science, statistics and neighbours. Full text of everything, the day it is posted. | Moderated, not peer reviewed. Basic search. The posted version may differ from the published one. |
| Semantic Scholar | Working outwards from one paper: citation contexts, author pages, a free API, and one-line TLDR summaries on computer science and biomedical papers. | Non-English literature is thin. TLDRs are AI-generated and are not offered outside those two areas; merged author records need checking. |
| OpenAlex | Computing over the literature rather than browsing it: CC0 data and a free download of the whole database. It now also caches full text — PDF and TEI XML — for a large subset of works. Built to replace Microsoft Academic Graph by the non-profit that renamed itself OpenAlex in 2025, formerly OurResearch. | An aggregated catalogue, not a curated index — expect duplicates and gaps. Since early 2026 the hosted API is metered: a small free daily allowance, larger with a free API key, and billed by usage beyond that. The data itself stays free. |
| CiNii Research | Japanese-language scholarship: articles, university bulletins, dissertations, books and research data in one search. | Many records are metadata only and hand you off elsewhere for the text. |
| J-STAGE | Full text of journals published by Japanese scholarly societies, much of it free to read. | A hosting platform, not an index. A journal that is not on it is not in it. |
Start from where you actually are
The right site depends less on your field than on what you already have.
- A topic and no papers. Google Scholar first, then a subject index to see what it missed. This is the only case where a general engine is the right opening move.
- One good paper and a suspicion there are more. Stop searching by keyword. Open its references and the papers citing it — via Google Scholar’s “Cited by”, Semantic Scholar, or a citation-graph tool. Two or three hops usually surfaces the cluster faster than another hour of query variations.
- A field with a canonical index. Use it and skip the general engine: PubMed for biomedicine, arXiv and DBLP for computer science, ACL Anthology for NLP, ERIC for education. A curated index tells you what it covers; a general engine does not.
- A search you will have to defend in writing. You need a documented scope and a repeatable query string — see the subscription databases below.
The literature a general search under-covers
Every general engine is weak in the same three places, and none warns you.
Scholarship published outside English. A large share of the world’s research is indexed nationally and effectively nowhere else. For Japan that index is CiNii Research, run by the National Institute of Informatics, which absorbed the older CiNii Articles service in April 2022 and now searches articles, university bulletins, dissertations, books, research data and funded projects together. The full text of Japanese society journals mostly sits on J-STAGE, run by the Japan Science and Technology Agency, much of it openly readable. The equivalents elsewhere are SciELO for Latin America, Spain and Portugal, KCI for Korea, and CNKI for China.
Older and out-of-print material. National libraries have digitised far more than search engines index. Japan’s National Diet Library consolidated its catalogues into a single NDL Search in January 2024 — its own holdings, a union catalogue of libraries nationwide, and its periodical article index — and since May 2022 its digital collections have let registered individuals read digitised out-of-print books and periodicals from home. For anything published before the web, the library catalogue beats the search engine.
Repository copies and grey literature. Theses, technical reports and working papers live in institutional repositories; CORE, BASE and OpenAIRE aggregate them across institutions. Unpaywall finds a legal open copy of a paywalled DOI when one exists.
When Scopus and Web of Science are worth it
Both are subscription products, normally reached through an institution. What you pay for is not more results but a defined and documented scope: a curated journal list, consistent metadata, and a query you can quote in a methods section so somebody else can rerun it. That matters for a systematic review, a bibliometric analysis, or any claim of the form “we searched the literature and found nothing”.
For everyday work the free services above cover it; paying out of pocket for either is rarely the right call.
A routine that holds up
- Write the query before you run it. Break the question into two or three concepts and list synonyms for each. Deciding it in the search box means you never quite know what you searched.
- Run it in at least two places: one broad engine, and one index that documents its scope.
- Record where and when you searched. Indexes change underneath you: a search without a date is not repeatable, including by you in three months.
- Walk citations both ways from the two or three best hits. Backwards tells you where the idea came from; forwards tells you whether it survived.
- Save as you go, into one place. Not at the end. The tab you meant to come back to is the one you lose.
Where a cross-search tool fits
Steps 2 and 5 are the tedious ones: the same query in four tabs, then hand-copying the results into whatever holds your library.
Litlas is one of the tools that collapse those two steps. A single search runs against 16 scholarly sources — OpenAlex, Crossref, DataCite, OpenCitations, ROR/ORCID, arXiv, DBLP, OpenReview, ACL Anthology, PubMed, the PMC Open Access subset, DOAJ, CORE, OpenAIRE Graph, Unpaywall and Common Crawl — filterable by venue, author and year, sortable by citation count. From any result you can open a citation graph and expand backwards into references or forwards into citing work, save papers to a library with your own notes and tags, import BibTeX, export BibTeX, RIS or CSL-JSON, and format references in APA, Chicago (author-date), IEEE, MLA or Vancouver. The free plan needs no card: five searches a day, three boards, 100 saved items, five graph views a day.
What it is not, because that decides when to reach for it:
- It has no index of its own for non-English scholarship. All 16 sources are international or EU-wide; none is a national index of Japanese-, Korean-, Chinese- or Spanish-language work. A Japanese university bulletin that CiNii Research finds may simply not be there. Cross-search replaces the four English-language tabs, not the national index.
- It does not summarise papers. The ✨ button hands a PDF or page off to Gemini, ChatGPT, Claude or Gemini Notebook and opens it there. The reading is done by whichever of those you chose; Litlas runs no model of its own.
- It is not a systematic review platform. No screening workflow, no conflict resolution, no PRISMA diagram. If the search has to appear in a methods section, run it in a database whose scope is documented.
- It is not a bibliometrics tool. The citation graph expands one paper at a time — no corpus-wide co-citation clustering, no network export of the kind VOSviewer expects.
- Five searches a day is five searches a day. Enough to explore a question, not enough to sweep a field in an afternoon. The daily caps come off on a paid plan.
If you only remember three things
- Start broad, then confirm in an index that publishes its scope. Google Scholar is the best first pass and the worst last word.
- If any of the literature you need was not published in English, go to that language’s national index directly. No general engine covers it well, and none will tell you what it missed.
- Follow citations in both directions from the two or three papers that matter most. That finds more relevant work than another hour of keyword variations.