Every AI research tool leads with a number: 138 million papers, 200 million, 310 million. Lined up like that, it looks like the one specification you can compare directly. It is the one you cannot.
The numbers are measured differently, none of them is audited by anyone outside the company quoting it, and several of these tools are counting overlapping copies of the same underlying corpus. What actually helps you choose is the kind of service you are looking at, because the kind determines what you can get out of it.
Each vendor counts its corpus its own way
Here is what each vendor said about itself on 22 August 2026.
| Service | Its own figure | What the figure covers |
|---|---|---|
| Semantic Scholar | 200 million on its About page, 214 million on its API page | Papers it indexes, from publisher partnerships, data providers and web crawls |
| Elicit | 138 million papers, plus 545,000 clinical trials | Its searchable corpus after de-duplication, drawn from Semantic Scholar, PubMed and OpenAlex |
| Consensus | 200 million | Scientific documents, mostly peer-reviewed articles with some conference papers and preprints |
| ResearchRabbit | 310 million | Not stated which database this comes from |
| Connected Papers | “hundreds of millions” | The Semantic Scholar corpus, which it licenses |
Two things fall out of that table immediately.
One vendor publishes the same figure at two precisions. Semantic Scholar says “over 200 million academic papers” on its About page and shows a counter reading 214 million on its API page [6]. The two agree — 214 million is over 200 million — and both were live on the same day. What that shows is that even inside one organisation the figure is rounded and refreshed independently on each surface. Assume the same width behind any single number a vendor quotes.
The corpora overlap. Elicit, Consensus and Connected Papers all build on Semantic Scholar, and Elicit and Consensus both also use OpenAlex. Elicit is the only one that explains the arithmetic [3]: its upstream sources advertise far more than 138 million between them — Semantic Scholar 200 million-plus, OpenAlex 243 million, PubMed 33 million-plus — and 138 million is what remains once overlapping and incomplete records are removed. So the smallest number on the list is the one produced by the most explicit method, and a bigger number elsewhere may simply mean nobody subtracted the duplicates.
Treat the counts as marketing copy that happens to be true, and choose on something else.
The three kinds of service
1. The index: it returns a list of papers
Semantic Scholar is the clearest example: a free, non-profit search and discovery service run at Ai2, which also publishes its corpus as open datasets and an API [5]. Several of the others are built on top of it — Connected Papers licenses it, and Elicit and Consensus both draw on it.
An index returns a list of papers. It does not answer your question, and when accuracy is what you need, that is an advantage: no stage between the paper and you rephrases anything.
2. The answer engine: it replies in prose
Elicit and Consensus retrieve first and generate second. Consensus describes its retrieval as a hybrid: semantic search over AI embeddings to catch the intent of a question, plus BM25 keyword search [2]. Elicit says it supports its generated claims with sentence-level citations, and can find up to 1,000 relevant papers and analyse up to 20,000 data points at once [3].
An answer engine gives you a paragraph with citations under it. What it gives you is bounded by what you asked.
3. The relationship map: it lays out related papers
Connected Papers and ResearchRabbit start from papers you already have rather than from a question. Connected Papers is explicit about its method: it is not a citation tree [1]. Similarity is computed from co-citation and bibliographic coupling — two papers whose reference lists and citing papers overlap heavily are presumed to be on related subjects — and the result is laid out as a force-directed graph. For each graph it analyses on the order of 50,000 papers and keeps the few dozen most strongly connected to your origin paper. ResearchRabbit works from starting papers, up to 50 of them on its free tier, which it says is “Free $0, Forever!” [4].
A map gives you the papers nearest your starting point. Without a starting paper there is nothing for it to work from.
What each kind structurally misses
| Kind | What you will not get from it |
|---|---|
| Index | Reading. You get the list and do the work, which is the right trade when you need to be sure nothing was rephrased |
| Answer engine | The thing you did not think to ask about. The answer’s boundary is your question’s boundary |
| Relationship map | A start from nothing. And because similarity is not citation, a paper that cites yours directly can sit outside the map, while one that never mentions you sits inside it |
That last row is the one people get wrong most often, and Connected Papers says it plainly on its own site [1]. If your question is “who has cited this since”, a similarity map is the wrong instrument; you want the citation edges themselves.
There is also a failure mode the answer engines share and cannot design away. Retrieving before generating removes invented references, but it does not remove misreadings of real papers — Consensus says so about its own product [2]. Because an answer engine excludes invented citations by construction, misreading is the error you will actually meet, so we covered how it shows up and how to check for it separately.
Picking one
- You need to be certain of coverage, for a systematic search. Use an index, and record the query. The wording you put into an answer engine is not a search strategy anyone can reproduce.
- You are new to a subject and need orientation. An answer engine is the fastest way in. Then verify the papers it names.
- You have one good paper and want the work near it. A relationship map.
- You want to know who has built on a specific result. Follow the citations forward, not a similarity map.
Most people end up using several services together, which is why the corpus numbers matter less than they look: you are not picking a single database, you are picking the service that fits the step you are on.
Doing the searching and the citation-following in one place
Litlas is built around the last two of those steps, with the reading and the writing attached.
One query runs across 16 scholarly sources at once — OpenAlex, Crossref, DataCite, OpenCitations, ROR/ORCID, arXiv, DBLP, OpenReview, ACL Anthology, PubMed, PMC Open Access, DOAJ, CORE, OpenAIRE Graph, Unpaywall and Common Crawl — matching on title, authors, DOI and venue, with filters for venue, author and year. Opening the citation graph on any paper gives you the edges themselves in both directions: what it cited, and who has cited it since. Papers you keep go into a board with your own notes and tags, and export as BibTeX, RIS or CSL-JSON.
For the reading step, the ✨ button hands the paper straight to ChatGPT, Claude, Gemini or Gemini Notebook in a new tab with a prompt already prepared, so the model you already trust is the one that reads it, and you can put the same paper in front of two of them.
The free plan needs no card and works right away.
Three things to remember
- Corpus counts do not compare. They are measured differently, several tools resell the same underlying corpus, and Semantic Scholar publishes its own figure at two precisions.
- Pick by kind of service: an index for coverage, an answer engine for orientation, a map for the work near a paper, citation edges for who built on what.
- Similarity is not citation. Connected Papers says it is not a citation tree, so do not use a similarity map to answer “who cited this”.
Bibliography
- Connected Papers. “About.” https://www.connectedpapers.com/about. Accessed 22 August 2026.
- Consensus. “How Consensus Works.” https://framer.consensus.app/home/blog/how-consensus-works/. Accessed 22 August 2026.
- Elicit. “Elicit’s Source for Papers.” Elicit Help Centre. https://support.elicit.com/en/articles/14758040-elicit-s-source-for-papers. Accessed 22 August 2026.
- ResearchRabbit. “Pricing.” https://www.researchrabbit.ai/pricing. Accessed 22 August 2026.
- Semantic Scholar. “About Semantic Scholar.” https://www.semanticscholar.org/about. Accessed 22 August 2026.
- Semantic Scholar. “Semantic Scholar Academic Graph API.” https://www.semanticscholar.org/product/api. Accessed 22 August 2026.