Scholar Labs: Google Scholar's Conservative "Deep Search" Entry — and the Academic AI Quadrant Map
On November 18, 2025 — the same day Google launched Gemini 3 — the Google Scholar team quietly launched Scholar Labs, the free incumbent's first generative-AI layer. Aaron Tay's early review delivers the key classification: Scholar Labs is a "Deep Search" tool, not "Deep Research" — it finds and ranks papers with LLM-judged relevance, but does not synthesize an answer across them.
Mechanics (from Tay's testing, confirmed and extended by his Katina comparative review): the system analyzes the natural-language question, decomposes it into multiple sub-queries (10+ observed; one test ran 12), executes them against the full Google Scholar index, merges/dedupes/reranks, then has an LLM evaluate candidates paper-by-paper, emitting a generated rationale for each relevant hit. It pauses at milestones of 10/20/30/40 relevant papers and hard-caps at 50 relevant papers found or 300 evaluated. Tay's post-review annotation: Scholar Labs has since gotten faster and can exceed 300 evaluated (600+ observed), with additional stopping rules. It is free, with low interpretability ("we can see the number of queries made but not what they were").
Why it matters — index scale as the moat:
"If Scholar Labs inherits this exact index (and isn't walled off from paywalled full-text), it should dominate in non-STEM disciplines where other databases are weaker." — Aaron Tay, Scholar Labs Early Review
"Academics don't need sub-second latency; they need better ranking of top-k." — Aaron Tay
But it is deliberately not a research agent:
"It is just Deep Search, or 'paper finder.' With most AI powered search tools pivoting towards Deep Research, which provide long reports and visualizations, this is something that Scholar Labs does not touch and likely may not for a while." — Aaron Tay
The 2×2 quadrant map of academic AI search
Tay's framework (popularized in his Katina Magazine comparative review) classifies tools on two axes — how much searching (quick vs. iterative/deep) and what is returned (ranked list vs. generated synthesis):
- Q1 Quick Search — classic Google Scholar, Web of Science Smart Search (hybrid Boolean + semantic, bundled free, no generation)
- Q2 Deep Search — Scholar Labs, Ai2 Asta Paper Finder, Undermind
- Q3 Quick RAG — Consensus Pro, Primo Research Assistant, Scopus Discovery AI
- Q4 Deep Research — OpenAI Deep Research, Gemini Deep Research, Perplexity, SciSpace Deep Review, Consensus Deep Search
He also draws the key architectural line: "Deep search refers to behavior: the tool runs multiple searches before returning results. Agentic search refers to architecture: an LLM decides which tools to call, in what order, and when to stop." Most 2025 academic deep-research tools (Undermind, Scholar Labs, SciSpace Deep Review) are workflow-bound; Consensus's Research Agent, Undermind Projects, and Elicit's Research Agent are the first genuinely agentic ones.
Consensus: the vertical's benchmark trajectory
From the same Katina review: Consensus raised an $11.5M Series A in July 2024 (Union Square Ventures; 400k MAU, $1.5M ARR), then "in May 2026, Consensus announced a $30 million round led by GreatPoint Ventures, claiming over 2.5 million monthly active users and a place among the top five most-used research tools." Its corpus spans 200M+ records (Semantic Scholar + OpenAlex + own crawl) with partnerships with "6 of the 12 largest publishers" (Wiley, Taylor & Francis, APA, AAAS, Sage, ACS). Pro Search = three-stage retrieval (embeddings+BM25 → rerank top 1,500 with quality signals → re-score top 20); Deep Search = up to 20 sub-searches, 1,000+ papers reviewed, ~50 selected, PRISMA-style flow diagram and a "Consensus Meter." Its agentic mode chains four tools (Paper Search, DOI Lookup, Author Search, Citation Crawling), and it is the only one of the three reviewed tools exposing an MCP server (joining Scite, SciSpace, Wiley, Elicit). Pricing: institutional FTE tiers, with Deep Search capped at 50 searches/month even on paid accounts; Scholar Labs is free; WoS Smart Search is bundled.
Reproducibility as a purchase criterion — Katina ran identical queries 5× per tool: WoS Smart Search top-10 Jaccard similarity 0.30; Consensus Deep top-50 Jaccard 0.94 but rank-unstable (18.4/50 same-rank); Scholar Labs initially appeared perfectly reproducible, later corrected for caching (Jaccard 0.66). Conclusion: "A search step that cannot be reproduced cannot be reported as a reproducible component of a systematic or scoping review without explicit acknowledgement."
Tay's market call:
"I suspect the market will split along workflow lines: casual or open-ended research through LLM platforms, and structured evidence work through platforms, like Consensus, with specialized features and affordances." — Katina Magazine
What this means for Hey, Lefty
- The free incumbent now occupies "better retrieval, no synthesis" — the synthesis/report quadrant is where startups and generalist agents compete, and Google is signaling it won't chase that soon.
- MCP is becoming the distribution hedge: academic search providers are becoming backend tools callable from ChatGPT/Claude — validating the orchestrator thesis in Integrating Custom MCP Servers into Hey, Lefty's Research Cycles and pressuring destination-only products.
- Reproducibility and interpretability are becoming measurable institutional procurement criteria — an opening for tools that expose their search steps (see AI Is Breaking the Proxies for Expertise: Goodhart's Law Comes for Research Output).