Semantic Scholar's Citation Graph vs Agentic AI: When the Old Stack Still Wins
Semantic Scholar's Academic Graph indexes paper, author, and citation records and exposes them through a public API — a paper endpoint, an author endpoint, and a recommendations endpoint that accepts sets of positive and negative paper IDs and returns a ranked list of related work (Semantic Scholar API). None of that output is generated text. A graph query returns edges that were already indexed from real reference lists; it structurally cannot invent a citation the way a language model completing a sentence can. What it can do is miss one — coverage gaps are a real limitation, just a different one from fabrication, and the graph spans a large but not exhaustive slice of the published literature.
What the Graph Actually Returns
The Recommendations endpoint does not write a report — it ranks. Feed it a small seed set of relevant papers marked positive and a set of off-topic papers marked negative, and it returns recently published, topically adjacent work drawn from existing citation and co-citation structure (Recommendations API). There is no synthesis step in that process where a fabricated title, year, or author list could enter, because there is no generation step at all — only ranking over what the graph already contains.
Where Agentic Tools Actually Add Value
Agentic research tools solve a different problem: turning a scattered set of sources into readable, synthesized prose. A 2026 benchmark across 14 LLM-based research agents found citation link validity above 90% but factual accuracy — whether a cited source actually said what the report claimed — at only 39 to 77% (arXiv:2605.06635), the same gap covered in the comparison of current deep-research agents. That is the tradeoff: synthesis is exactly the step a citation graph does not perform, and exactly the step where an agentic tool's factual accuracy drops below its link validity.
When the Graph Wins
Seed expansion — finding what a known paper connects to, forward and backward — is a structural query the graph answers directly. A chat-style agent answers the same question by re-searching and re-synthesizing each time, reintroducing the same fabrication risk with every pass. For snowball searching from a small set of anchor papers, the graph's ranked output has nothing to fact-check beyond "these papers are connected," because it never asserts more than that. A screening pass built on citation and co-citation edges also stays reproducible in a way a re-run agentic query does not: the same seed set returns the same ranked neighbors, which matters when a search strategy has to be reported and defended later.
When the Agent Wins
A graph query cannot summarize what a paper argues, resolve a natural-language question across a dozen documents, or draft prose. For orienting quickly in an unfamiliar area or producing a first-pass background section, an agentic tool does something the graph structurally cannot — paid for with a factual-accuracy gap that has to be checked before any of it lands in a manuscript, the same discipline citation hallucination in AI-assisted writing already requires for any AI-drafted section.
A Practical Split
Use the graph for discovery and snowballing: seed papers in, ranked related papers out, nothing generated to verify. Use an agentic tool for synthesis of a source set already trusted, then verify what it produced before it reaches a draft. The literature search tool inside AI for Academic queries PubMed and OpenAlex directly rather than routing every lookup through a generative synthesis step, which keeps discovery closer to the graph model than the open-ended agent model. Free to start.