Autonomous Research Tools: Market Map Digest — Cycle 4
TL;DR
This cycle sharpened the two edges of the market map that matter most for positioning: the academic deep-search layer now has a complete price ladder from free incumbent to enterprise vertical, and Google just showed what a trust moat looks like when a host platform owns the content supply chain — licensed, entitlement-checked publisher books woven directly into the notebook. The squeeze on paid deep-search startups is now structural, not competitive.
The academic deep-search price ladder is complete — and both ends are free
The academic slice of the research-tool market now has every rung filled: free incumbent at the top, commercial SaaS in the middle, enterprise vertical below it, and a non-profit giving away both agents and the underlying index at the bottom (academic-deep-search-undermind-asta). Undermind, the YC S24 startup founded by two MIT-trained physicists, sells an adaptive multi-step Deep Search agent where results come back "in a table with a match score for each paper and a stated reason the paper was included" — but its Pro tier is $16/mo, verified against undermind.ai by CASRAI's guide (CASRAI guide).
The non-profit pole makes that price hard to defend. Ai2's Asta, announced August 26, 2025, offers free research agents plus a developer layer whose Scientific Corpus Tool is "an MCP extension of our Semantic Scholar API that provides agents free access to a 200M+ normalized index of papers, already serving over 1.5 billion queries per year" (Ai2 blog). Here's the trap in Undermind's position:
"Currently searches Semantic Scholar's index where classification runs on titles and abstracts." — CASRAI guide
Undermind's data layer is the index Ai2 gives away. That means its $16/mo can only be defended on the agent loop and enterprise workflow — and the enterprise workflow is where the money actually is: "GSK deployed Undermind as part of its internal 'AI Scientist' stack, with over 1,000 scientists using the tool" (CASRAI guide). The consumer tier is a funnel; the institution is the customer. Both poles are also converging on backend infrastructure — Undermind "markets an enterprise offering positioned as infrastructure for other AI research agents," while Asta is MCP-native from day one (academic-deep-search-undermind-asta) — which reinforces the destination-vs-backend bifurcation from previous cycles.
What to watch: Whether independent verification ever lands for Undermind's recall claims — the only evidence today is one librarian review (JCHLA) and vendor case studies, and a new thread is open on exactly that.
Google's trust moat is now a licensed-content supply chain, not a citation format
The host layer's answer to provenance anxiety arrived on August 27, 2026: Expert Intelligence, "a cross Google initiative that helps you engage with your trusted sources through Google AI products," launching in Gemini Notebook with "more than 100,000 books from major publishers, including Bloomsbury, De Gruyter Brill, Johns Hopkins University Press, Macmillan Publishers, O'Reilly Media, and Penguin Random House" (Google blog). The mechanism is ownership-gated grounding — and Google built the entitlement into sharing itself:
"for people to interact with a book in Gemini Notebook, they must own the book through Google Play Books. If a notebook is shared, other collaborators will be prompted to purchase their own copy to engage with the content further." — Google blog
This is a third publisher-payment model, distinct from the other two on the map. Parallel pays publishers per agent-use; Sacra sells premium proprietary data as agent infrastructure; Google converts reader purchases into AI-grounding entitlements — the user pays the publisher, Google takes the engagement, and publishers get a discovery channel instead of a payout (google-expert-intelligence-licensed-sources). The roadmap is the real signal for the market map: "Soon, people will also be able to add more sources like third-party subscriptions, business research reports, and textbooks" (Google blog). That line points directly at the supply premium research-data vendors sell — a watch thread is now registered on it.
For Hey, Lefty, the lesson cuts both ways: model-agnostic orchestration must also be source-agnostic. The durable workspace should ingest licensed, user-owned, and open sources alike rather than compete on content supply — because competing on supply against Google's entitlement machinery is a fight no startup can fund (google-expert-intelligence-licensed-sources).
What to watch: The first non-book source type going live in Expert Intelligence — the moment subscriptions or business research reports enter the grounding layer, Google is in direct contact with Sacra-style premium data vendors.
What surprised us
- Undermind's moat and its constraint are the same thing. Its adaptive agent loop is the product, but its dependence on Semantic Scholar's titles-and-abstracts index means its data layer is literally the free commodity its non-profit competitor distributes (academic-deep-search-undermind-asta
). Very few $16/mo products have their supply chain given away by the entity across the table.
- Enterprise "AI Scientist" stacks keep showing up as the actual revenue. GSK's 1,000-scientist deployment is the second consecutive cycle where institutional workflow, not individual seats, is where deep-search tools monetize — the consumer tier is consistently priced as a funnel.
- Ownership-gating is a sharing mechanic, not just a licensing mechanic. Google turned collaboration itself into a purchase funnel — every notebook share generates a Play Books checkout prompt (Google blog). That's distribution no destination-layer startup can replicate.
- AstaBench changes the verification game. A free, open evaluation framework with 2,400+ problems across 11 benchmarks — reporting accuracy-vs-compute Pareto frontiers — makes vendor recall claims like Undermind's independently testable in principle (Ai2 blog). The open question thread should check whether anyone has actually run the comparison.
Open threads worth a vote
- [question] Independent verification of Undermind's recall claims vs. Google Scholar — should the next pass prioritize library-guide test queries or AstaBench-style replication? (threads)
- [watch] Google Expert Intelligence expands beyond Play Books — fires when subscriptions, business research reports, or textbooks go live in the grounding layer (threads)